Merge pull request #6 from sisl/dev-idm-videos

Optionally save videos of evaluated policies

README.md

@@ -1,10 +1,22 @@
 # InteractionImitation
-Imitation Learning with the INTERACTION Dataset
+Imitation Learning with the [Interaction Dataset](https://interaction-dataset.com/) via the [InteractionSimulator](https://github.com/sisl/InteractionSimulator) gym environments.
+
+Code for "[SHAIL: Safety-Aware Hierarchical Adversarial Imitation Learning for Autonomous Driving in Urban Environments](https://arxiv.org/abs/2204.01922)".
+If you find this repository useful, please cite the paper:
+
+```
+@article{jamgochian2022shail,
+  author = {Arec Jamgochian and Etienne Buehrle and Johannes Fischer and Mykel J. Kochenderfer},
+  title = {{SHAIL}: Safety-Aware Hierarchical Adversarial Imitation Learning for Autonomous Driving in Urban Environments},
+  journal = {arXiv:2204.01922 [cs]},
+  year = {2022}
+}
+```
 
 ## Getting started
-Clone InteractionSimulator and pip install the module.
+Clone the `InteractionSimulator` with the `shail` tag and pip install the module.
 ```
-git clone https://github.com/sisl/InteractionSimulator.git
+git clone --branch shail https://github.com/sisl/InteractionSimulator.git
 cd InteractionSimulator
 pip install -e .
 cd ..
@@ -19,53 +31,28 @@ The INTERACTION dataset contains a two folders which should be copied into a fol
   - the contents of `recorded_trackfiles` should be copied to `./InteractionSimulator/datasets/trackfiles`
   - the contents of `maps` should be copied to `./InteractionSimulator/datasets/maps`
 
-## Processing, saving, and loading expert demos
+## Processing and saving expert demos
-Once the repository has been set up, you can process and save expert track demonstrations with:
+Once the repository has been set up, you need to generate two separate sets of expert demos for tracks 0-4. The first command generates true joint and individual states and actions necessary for evaluating, saving them in `expert_data/`. The second command generates trajectory rollouts according to individual agent observations, which is later used as expert data for the learning models.
 ```
-python src/expert_data.py --loc [LOCNUM] --track [TRACKNUM]
+python -m src.expert --locs='[DR_USA_Roundabout_FT]' --tracks='[0,1,2,3,4]'
-```
+python -m intersimple-expert-rollout-setobs2 --tracks='[0,1,2,3,4]'
-You can (and should) process all tracks at once at location 0 with:
-```
-python src/expert_data.py --all-tracks
-```
-
-You can then train a default behavior cloning policy with the following. Be sure to check help for main.py for running options.
-```
-python src/main.py --train
-```
-You can run tensorboard by running the following and opening `localhost:6006` (or alternatively port-forwarding 6006 from the remote server)
-```
-tensorboard --logdir output/
-```
-You can then test the learned policy with the following, and see the animation file in `output/`:
-```
-python src/main.py --test
 ```
 
 
-You can load the experts actions manually
+## Tuning hyperparameters and training finalized models
-```
+To tune models, we use `ray[tune]` grid searches. You can run see the commands we used to train in the top half of `train_models.sh`, as well as the hyperparameters we search over in `bc-experiment.py`, `gail-experiment.py`, and `shail-experiment.py`. After training the models, configurations get saved in `best_configs/` (the best SHAIL confg gets copied to a HAIL config, with the appropriate environment parameters changed for ablation). However, upon manual inspection of the training runs, we note some better performance than the automatically-set configs at earlier epochs, so we adjust the `best_configs` manually.
-from src import expert_data
+
-observations, actions = expert_data.load_expert_data(loc = [LOCNUM], track = [TRACKNUM])
+After the `best_configs/` are set, we rerun each configuration with multiple seeds. The commands to do so are in the bottom half of `train_models.sh`. This saves different learned policy files to `test_policies/`.
-for (s, a) in zip (observations, actions):
+
-    # do some imitation learning
+
-```
+## Evaluating models
+To evaluate the learned policies, we rerun each model in particular setting, evaluate all our metrics, and average over different trained model seeds. The commands to do so are in `evaluate_models.sh`. 
 
 
 ## Package Structure
 ```
 InteractionImitation
-|- demos
+|- TODO
-|- algorithms
-   |- BC
-   |- AdVIL
-|- nets
-   |- Encoder
-   |- DeepSet
-   |- Decoder
-|- policies
-|- discriminators
-|- demo_generators
 ```
 
 ## Type Definitions

bc-experiment.py

@@ -55,6 +55,7 @@ def training_function(config):
                 ),
                 check_collisions=True,
                 stop_on_collision=config['trainenv']['stop_on_collision'],
+                use_idm=config['trainenv']['use_idm'],
             ), collision_distance=6, collision_penalty=100),
             lambda obs: (obs - obs_min) / (obs_max - obs_min + 1e-10)
         )) for _ in range(60)]
@@ -68,6 +69,8 @@ def training_function(config):
                 ),
                 check_collisions=True,
                 stop_on_collision=config['trainenv']['stop_on_collision'],
+                use_idm=config['trainenv']['use_idm'],
+                track=track,
             ), collision_distance=6, collision_penalty=100),
             lambda obs: (obs - obs_min) / (obs_max - obs_min + 1e-10)
         )) for _ in range(15)] for track in range(4)],[])
@@ -159,6 +162,7 @@ if __name__ == '__main__':
                 'experiment': args.train,
                 'trainenv': {
                     'stop_on_collision': False, 
+                    'use_idm':True,
                 },
                 'policy': {
                     'learning_rate': 3e-4, 

best_configs/bc_expA.json

@@ -1,7 +1,8 @@
 {
     "experiment": "A",
     "trainenv": {
-        "stop_on_collision": false
+        "stop_on_collision": false,
+        "use_idm": true
     },
     "policy": {
         "learning_rate": 0.0003,

best_configs/bc_expB.json

@@ -1,7 +1,8 @@
 {
     "experiment": "B",
     "trainenv": {
-        "stop_on_collision": false
+        "stop_on_collision": false,
+        "use_idm": true
     },
     "policy": {
         "learning_rate": 0.0003,

best_configs/gail_expA.json

@@ -1,7 +1,8 @@
 {
     "experiment": "A",
     "trainenv": {
-        "stop_on_collision": false
+        "stop_on_collision": false,
+        "use_idm": true
     },
     "policy": {
         "learning_rate": 0.0003,

best_configs/gail_expB.json

@@ -1,7 +1,8 @@
 {
     "experiment": "B",
     "trainenv": {
-        "stop_on_collision": false
+        "stop_on_collision": false,
+        "use_idm": true
     },
     "policy": {
         "learning_rate": 0.0003,

best_configs/hail_expA.json

@@ -3,7 +3,8 @@
     "trainenv": {
         "stop_on_collision": false,
         "safe_actions_collision_method": null,
-        "abort_unsafe_collision_method": null
+        "abort_unsafe_collision_method": null,
+        "use_idm": true
     },
     "policy": {
         "learning_rate": 0.0003,

best_configs/hail_expB.json

@@ -3,7 +3,8 @@
     "trainenv": {
         "stop_on_collision": false,
         "safe_actions_collision_method": null,
-        "abort_unsafe_collision_method": null
+        "abort_unsafe_collision_method": null,
+        "use_idm": true
     },
     "policy": {
         "learning_rate": 0.0003,

best_configs/shail_expA.json

@@ -3,7 +3,8 @@
     "trainenv": {
         "stop_on_collision": false,
         "safe_actions_collision_method": "circle",
-        "abort_unsafe_collision_method": "circle"
+        "abort_unsafe_collision_method": "circle",
+        "use_idm": true
     },
     "policy": {
         "learning_rate": 0.0003,

best_configs/shail_expB.json

@@ -3,7 +3,8 @@
     "trainenv": {
         "stop_on_collision": false,
         "safe_actions_collision_method": "circle",
-        "abort_unsafe_collision_method": "circle"
+        "abort_unsafe_collision_method": "circle",
+        "use_idm": true
     },
     "policy": {
         "learning_rate": 0.0003,

config/networks.json5

@@ -1,42 +0,0 @@
-{
-    ego_encoder: {
-        input_dim: 5, // number of state vars
-        hidden_n: 0,
-        hidden_dim: 5,
-        output_dim: 5
-    },
-    deepsets: {
-        input_dim: 6, // number of relative state vars for others
-        phi: {
-            hidden_n: 2,
-            hidden_dim: 20,
-        },
-        latent_dim: 20,
-        rho: {
-            hidden_n: 2,
-            hidden_dim: 10,
-        },
-        output_dim: 10
-    },
-    path_encoder: {
-        input_dim: 40, // 2 * path length for (x,y) coordinates
-        hidden_n: 0,
-        hidden_dim: 20,
-        output_dim: 10,
-    },
-    head: {
-        input_dim: 0, // computed in policy constructor
-        hidden_n: 3,
-        hidden_dim: 50,
-        output_dim: 1, // number of outputs e.g. number of actions, or just one
-        final_activation: 'sigmoid',
-    },
-    optim: {
-            optimizer: 'adam',
-            lr: 1e-3,
-            weight_decay: 0.1,
-    },
-    train_epochs: 200,
-    train_batch_size: 32,
-    loss: 'huber',
-}

config/value_dice.json5

@@ -1,85 +0,0 @@
-{
-    policy_net: {
-        ego_encoder: {
-            input_dim: 5, // number of state vars
-            hidden_n: 0,
-            hidden_dim: 5,
-            output_dim: 5
-        },
-        deepsets: {
-            input_dim: 6, // number of relative state vars for others
-            phi: {
-                hidden_n: 2,
-                hidden_dim: 20,
-            },
-            latent_dim: 20,
-            rho: {
-                hidden_n: 2,
-                hidden_dim: 10,
-            },
-            output_dim: 10
-        },
-        path_encoder: {
-            input_dim: 40, // 2 * path length for (x,y) coordinates
-            hidden_n: 0,
-            hidden_dim: 20,
-            output_dim: 10,
-        },
-        head: {
-            input_dim: 0, // computed in policy constructor
-            hidden_n: 3,
-            hidden_dim: 50,
-            output_dim: 1, // number of outputs e.g. number of actions, or just one
-            final_activation: 'sigmoid',
-        },
-    },
-    value_net: {
-        ego_encoder: {
-            input_dim: 5, // number of state vars
-            hidden_n: 0,
-            hidden_dim: 5,
-            output_dim: 5
-        },
-        deepsets: {
-            input_dim: 6, // number of relative state vars for others
-            phi: {
-                hidden_n: 2,
-                hidden_dim: 20,
-            },
-            latent_dim: 20,
-            rho: {
-                hidden_n: 2,
-                hidden_dim: 10,
-            },
-            output_dim: 10
-        },
-        path_encoder: {
-            input_dim: 40, // 2 * path length for (x,y) coordinates
-            hidden_n: 0,
-            hidden_dim: 20,
-            output_dim: 10,
-        },
-        action_dim: 1, // number of actions
-        head: {
-            input_dim: 0, // computed in policy constructor
-            hidden_n: 3,
-            hidden_dim: 50,
-            output_dim: 1, // number of outputs e.g. number of actions, or just one
-            final_activation: 'id',
-        },
-    },
-    policy_optim: {
-            optimizer: 'adam',
-            lr: 1e-3,
-            weight_decay: 0.1,
-    },
-    value_optim: {
-            optimizer: 'adam',
-            lr: 1e-3,
-            weight_decay: 0.1,
-    },
-    train_epochs: 200,
-    train_batch_size: 32,
-    discount: 0.95,
-    clip_grad_norm: 1.,
-}

cp-videos.sh

@@ -0,0 +1,10 @@
+# cp-videos videos/ videos/icra23/
+
+agents=( 5 27 39 43 47 53 63 81 83 87 93 96 105 113 124 127 130 134 )
+
+for a in "${agents[@]}"
+do
+    cp "$1/expert_agent/loc0/track0/agent${a}_ani.mp4" "$2/t${a}expert.mp4"
+    cp "$1/idm/loc0/track0/agent${a}_ani.mp4" "$2/t${a}idm.mp4"
+    cp "$1/shail/loc0/track0/agent${a}_ani.mp4" "$2/t${a}shail.mp4"
+done

eval_experiments.py

@@ -6,22 +6,24 @@ import json
 
 activations = [torch.nn.Tanh, torch.nn.LeakyReLU]
 
-def main(method:str='expert', folder:str=None, locations=[(0,0)], skip_running=False):
+def main(method:str='expert', folder:str=None, locations=[(0,0)], skip_running=False, save_videos:bool=False, videos_folder:str='videos', first_seed_only:bool=False):
     
     exclude_keys_from_policy_kwargs = {'learning_rate', 'learning_rate_decay', 'clip_ratio', 'iterations_per_epoch', 'option'}
     policy_kwargs = {}
 
-    if method in ['expert', 'idm']:
+    if method in ['expert', 'expert_agent']:
         env, env_kwargs ='NRasterizedRouteIncrementingAgent', {}
+    elif method in ['idm']:
+        env, env_kwargs ='NRasterizedRouteIncrementingAgent', {'use_idm':True}
     elif method in ['bc','gail']:
         env='NormalizedContinuousEvalEnv' 
-        env_kwargs={'stop_on_collision':True, 'max_episode_steps':1000}
+        env_kwargs={'stop_on_collision':True, 'max_episode_steps':1000, 'use_idm':True}
     elif method in ['hail']:
         env = 'NormalizedSafeOptionsEvalEnv'
-        env_kwargs={'stop_on_collision':True, 'max_episode_steps':1000, 'safe_actions_collision_method': None, 'abort_unsafe_collision_method': None}
+        env_kwargs={'stop_on_collision':True, 'max_episode_steps':1000, 'safe_actions_collision_method': None, 'abort_unsafe_collision_method': None, 'use_idm':True}
     elif method in ['shail']:
         env = 'NormalizedSafeOptionsEvalEnv'
-        env_kwargs={'stop_on_collision':True, 'max_episode_steps':1000}
+        env_kwargs={'stop_on_collision':True, 'max_episode_steps':1000, 'use_idm':True}
     else:
         raise NotImplementedError
 
@@ -30,8 +32,13 @@ def main(method:str='expert', folder:str=None, locations=[(0,0)], skip_running=F
     if folder is not None:
         files = [os.path.join(folder, f) for f in os.listdir(folder) if os.path.isfile(os.path.join(folder, f))]
         files = [f for f in files if f.endswith('.pt')]
+
+        if first_seed_only:
+            files = files[:1]
+
         with open(os.path.join(folder, 'config.json'), 'rb') as f:
             config = json.load(f)
+            
         print('%i policy files found in %s folder' %(len(files), folder))
         print('found policy config', config['policy'])
 
@@ -50,7 +57,8 @@ def main(method:str='expert', folder:str=None, locations=[(0,0)], skip_running=F
                                 policy_file=policy_file, 
                                 policy_kwargs=policy_kwargs,
                                 env=env, 
-                                env_kwargs=env_kwargs)
+                                env_kwargs=env_kwargs,
+                                videos_folder=None if not save_videos else videos_folder)
         outfolder = os.path.dirname(outbase)
     else:
         locstr = 'loc_'+'_'.join([f'r{ro}t{tr}' for (ro,tr) in locations])

evaluate_models.sh

@@ -1,19 +1,18 @@
-# can add --skip_running if you've run the runs before on the saved policies
+# can add --skip_running if you've already run the saved policies through the test environments and have appropriate
+# metrics in the out folder. Doing so will generate average metrics quickly.
 
+# Experiment A
 python -m eval_experiments 
-python -m eval_experiments --locations='[(0,4)]' 
 python -m eval_experiments --method idm 
-python -m eval_experiments --method idm --locations='[(0,4)]' 
 python -m eval_experiments --method bc --folder='test_policies/bc/expA' 
-python -m eval_experiments --method bc --folder='test_policies/bc/expB' --locations='[(0,4)]' 
 python -m eval_experiments --method gail --folder='test_policies/gail/expA' 
-python -m eval_experiments --method gail --folder='test_policies/gail/expB' --locations='[(0,4)]' 
 python -m eval_experiments --method hail --folder='test_policies/hail/expA'
-python -m eval_experiments --method hail --folder='test_policies/hail/expB' --locations='[(0,4)]'
+python -m eval_experiments --method shail --folder='test_policies/shail/expA' 
-python -m eval_experiments --method shail --folder='test_policies/shail/expA'
-python -m eval_experiments --method shail --folder='test_policies/shail/expB' --locations='[(0,4)]'
 
-python -m eval_experiments --method hail --folder='test_policies/hail-etienne/expA'
+# Experiment B
-python -m eval_experiments --method hail --folder='test_policies/hail-etienne/expB' --locations='[(0,4)]'
+python -m eval_experiments --locations='[(0,4)]' 
-python -m eval_experiments --method shail --folder='test_policies/shail-etienne/expA'
+python -m eval_experiments --method idm --locations='[(0,4)]'
-python -m eval_experiments --method shail --folder='test_policies/shail-etienne/expB' --locations='[(0,4)]'
+python -m eval_experiments --method bc --folder='test_policies/bc/expB' --locations='[(0,4)]'
+python -m eval_experiments --method gail --folder='test_policies/gail/expB' --locations='[(0,4)]'
+python -m eval_experiments --method hail --folder='test_policies/hail/expB' --locations='[(0,4)]'
+python -m eval_experiments --method shail --folder='test_policies/shail/expB' --locations='[(0,4)]'

experiments/experiment.py

@@ -1,203 +0,0 @@
-import json5
-from functools import partial
-import os
-opj = os.path.join
-
-# set up ray tune
-import ray
-from ray import tune
-from ray.tune import Analysis, ExperimentAnalysis
-from ray.tune.schedulers import ASHAScheduler
-from hyperopt import hp
-from ray.tune.suggest.hyperopt import HyperOptSearch
-
-# get graphs
-import intersim
-from intersim.graphs import ConeVisibilityGraph
-
-
-from src.main import basestr, main
-
-def parse_args():
-    """
-    Parse arguments to main
-    Returns:
-        kwargs: dictionary of arguments:
-            train (bool): whether to run train loop
-            test (bool): whether to run test loop
-            method (str): the method to try for imitation
-            loc (int): the location index of the roundabout
-            config (str): config path
-            seed (int): RNG seed
-    """
-    import argparse
-    parser = argparse.ArgumentParser(description='Save Expert Trajectories')
-    parser.add_argument('--loc', default=0, type=int,
-        help='location (default 0)')
-    parser.add_argument("--train", help="train model",
-        action="store_true")
-    parser.add_argument("--ray", help="use ray tune to run multiple experiments",
-        action="store_true")
-    parser.add_argument("--test", help="test model",
-        action="store_true") 
-    parser.add_argument("--method", help="modeling method",
-        choices=['bc', 'gail', 'advil', 'vd'], default='bc')
-    parser.add_argument("--config", help="config file path",
-        default=None, type=str)
-    parser.add_argument('--seed', default=0, type=int,
-        help='seed')    
-    parser.add_argument('--nframes', default=500, type=int,
-        help='frames for test animation') 
-    parser.add_argument('--nsamples', default=200, type=int,
-        help='number of ray samples') 
-    parser.add_argument('--graph', action='store_true',
-        help='whether to mask the relative states based on a ConeVisibilityGraph')
-    parser.add_argument('-d', default='./expert_data', type=str,
-                       help='data directory')
-    parser.add_argument('-o', default=None, type=str,
-                       help='output directory')                   
-    args = parser.parse_args()
-    kwargs = {
-        'train':args.train, 
-        'test':args.test, 
-        'method':args.method,
-        'loc':args.loc,
-        'config_path':args.config,
-        'seed':args.seed,
-        'ray':args.ray,
-        'nframes':args.nframes,
-        'nsamples':args.nsamples,
-        'datadir':os.path.abspath(args.d),
-        'graph':None,
-        'outdir': opj('output',args.method,'loc%02i'%(args.loc)),
-        'train_tracks':[0,1,2],
-        'cv_tracks':[3],
-        'test_tracks':[4],
-    }
-    if args.o:
-        kwargs['outdir'] = args.o
-    if args.graph:
-        kwargs['graph'] = ConeVisibilityGraph(r=20, half_angle=120)
-    return kwargs
-
-def get_full_config(ray_config:dict, method:str)->dict:
-    """
-    Get full model configuration from ray config and method string
-    Args:
-        ray_config (dict): ray config
-        method (str): method to get full configuration for
-    """
-    if method == 'bc':
-        from src.bc import bc_config
-        config = bc_config(ray_config)
-    elif method == 'vd':
-        from src.value_dice import vd_config
-        config = vd_config(ray_config)
-    else:
-        raise NotImplementedError
-    return config
-
-def get_ray_config(method:str)->dict:
-    """
-    Get configuration for ray based on method.
-    Args: 
-        method (str): method to get configuration for
-    Returns:
-        ray_config (dict): configuration for ray
-    """
-    if method == 'bc':
-        ray_config = {
-            "lr": tune.loguniform(1e-5, 1e-3),
-            "weight_decay": tune.choice([0, 0.1]),
-            "loss": tune.choice(['huber', 'mse']),
-            "train_batch_size": tune.choice([16,32,64]),
-            "deepsets_phi_hidden_n": tune.randint(1,5),
-            "deepsets_phi_hidden_dim": tune.lograndint(8,65),
-            "deepsets_latent_dim": tune.lograndint(8,129),
-            "deepsets_rho_hidden_n": tune.randint(0,3),
-            "deepsets_rho_hidden_dim": tune.lograndint(8,129),
-            "deepsets_output_dim": tune.lograndint(4,129),
-            "head_hidden_n": tune.randint(1,6),
-            "head_hidden_dim": tune.lograndint(16,257),
-            "head_final_activation": tune.choice(['sigmoid', None]),
-        }
-    elif method == 'vd':
-        ray_config = {
-            "policy_lr": tune.loguniform(1e-5, 1e-3),
-            "value_lr": tune.loguniform(1e-5, 1e-3),
-            "policy_weight_decay": tune.choice([0, 0.1]),
-            "value_weight_decay": tune.choice([0, 0.1]),
-            "train_batch_size": tune.choice([16,32,64]),
-            "deepsets_phi_hidden_n": tune.randint(1,5),
-            "deepsets_phi_hidden_dim": tune.lograndint(8,65),
-            "deepsets_latent_dim": tune.lograndint(8,129),
-            "deepsets_rho_hidden_n": tune.randint(0,3),
-            "deepsets_rho_hidden_dim": tune.lograndint(8,129),
-            "deepsets_output_dim": tune.lograndint(4,129),
-            "head_hidden_n": tune.randint(1,6),
-            "head_hidden_dim": tune.lograndint(16,257),
-            "head_final_activation": tune.choice(['sigmoid', None]),
-            "clip_grad_norm": tune.choice([.5, 1., 5., 10.]),
-            "discount": tune.choice([.95, .99])
-        }
-    else:
-        raise NotImplementedError
-    return ray_config
-
-if __name__ == '__main__':
-    kwargs = parse_args()
-    
-    # make prefix of output files
-
-    if kwargs['config_path']:
-        # load config
-        with open(kwargs['config_path'], 'r') as cfg:
-            config = json5.load(cfg)
-        if not os.path.isdir(kwargs['outdir']):
-            os.makedirs(kwargs['outdir'])   
-        filestr = opj(kwargs['outdir'], basestr(**kwargs)) 
-        if kwargs['ray']:
-            filestr = kwargs['config_path'].replace('_config.json','')
-        main(config, filestr=filestr, **kwargs)
-
-    elif kwargs['ray'] and kwargs['train']:
-
-        ray.shutdown() 
-        ray.init(log_to_driver=False)
-
-        def ray_train(config, datadir=None):
-            full_config = get_full_config(config, kwargs['method'])
-            main(full_config, filestr='exp', **kwargs)
-        
-        ray_config = get_ray_config(kwargs['method'])
-        search = HyperOptSearch(ray_config, max_concurrent=8, metric='cv_loss',mode="min",)
-        custom_scheduler = ASHAScheduler(metric='cv_loss', mode="min", grace_period=15)
-
-        analysis = tune.run(
-            ray_train, 
-            #config=ray_config,
-            search_alg=search,
-            scheduler=custom_scheduler,
-            local_dir=kwargs['outdir'],
-            #resources_per_trial={"cpu": 2},
-            time_budget_s=120*60,
-            num_samples=kwargs['nsamples'],
-        )
-    elif kwargs['ray'] and kwargs['test']:
-        analysis = Analysis(kwargs['outdir'], default_metric="cv_loss", default_mode="min")
-        config = analysis.get_best_config()
-        filepath = analysis.get_best_logdir()
-        filestr = opj(filepath, 'exp')
-        config_path = filestr+'_config.json'
-        with open(config_path, 'r') as cfg:
-            config = json5.load(cfg)
-        print("Best ray experiment:", filepath)
-        main(config, filestr=filestr, **kwargs)
-    else:
-        raise Exception('No valid config found')
-
-    
-    
-    
-
-    

experiments/experiments.sh

@@ -1,9 +0,0 @@
-#!/bin/sh
-
-python experiments/experiment.py --ray --train -d ./expert_data/base
-python experiments/experiment.py --ray --test -d ./expert_data/base --nframes 1000
-python experiments/experiment.py --ray --train -d ./expert_data/reg
-python experiments/experiment.py --ray --test -d ./expert_data/reg --nframes 1000
-python experiments/experiment.py --ray --train -d ./expert_data/reg_graph --graph
-python experiments/experiment.py --ray --test -d ./expert_data/reg_graph --graph --nframes 1000
-

experiments/train_vd.sh

@@ -1,5 +0,0 @@
-#!/bin/sh
-
-# python experiments/experiment.py --method vd --train --ray -d expert_data/reg -o output/vd/loc00/reg --nsamples 400
-# python experiments/experiment.py --test --ray --method vd -d expert_data/normal -o output/vd/loc00/normal --nframes 1000
-python experiments/experiment.py --train --method vd --config config/value_dice.json5

gail-experiment.py

@@ -53,6 +53,7 @@ def training_function(config):
                 ),
                 check_collisions=True,
                 stop_on_collision=config['trainenv']['stop_on_collision'],
+                use_idm=config['trainenv']['use_idm'],
             ), collision_distance=6, collision_penalty=100),
             lambda obs: (obs - obs_min) / (obs_max - obs_min + 1e-10)
         )) for _ in range(60)]
@@ -67,6 +68,7 @@ def training_function(config):
                 ),
                 check_collisions=True,
                 stop_on_collision=config['trainenv']['stop_on_collision'],
+                use_idm=config['trainenv']['use_idm'],
                 track=track,
             ), collision_distance=6, collision_penalty=100),
             lambda obs: (obs - obs_min) / (obs_max - obs_min + 1e-10)
@@ -169,7 +171,8 @@ if __name__ == '__main__':
             config={
                 'experiment': args.train,
                 'trainenv': {
-                    'stop_on_collision': False, 
+                    'stop_on_collision': False,
+                    'use_idm': True, 
                 },
                 'policy': {
                     'learning_rate': 3e-4, 

generate_demos.sh

@@ -1,10 +0,0 @@
-#DEFAULT PARAMETERS:
-# locs:list=None, (default to all locations)
-# tracks:list=None, (default to all tracks)
-# env_class:str='NRasterizedIncrementingAgent',
-# env_args:dict={width:36,height:36,m_per_px:2},
-# expert_class:str='NRasterizedRouteIncrementingAgent',  
-# expert_args:dict={mu:0.001}):
-
-# python -m src.data.expert --locs='[DR_USA_Roundabout_FT]' --tracks='[0]'
-python -m src.data.expert --locs='[DR_USA_Roundabout_FT]' --tracks='[0]'

generate_videos.sh

@@ -0,0 +1,20 @@
+# can add --skip_running if you've already run the saved policies through the test environments and have appropriate
+# metrics in the out folder. Doing so will generate average metrics quickly.
+
+# Experiment A
+python -m eval_experiments 
+python -m eval_experiments --method expert_agent --save_videos --first_seed_only
+python -m eval_experiments --method idm --save_videos --first_seed_only
+python -m eval_experiments --method bc --folder='test_policies/bc/expA' --save_videos --first_seed_only
+python -m eval_experiments --method gail --folder='test_policies/gail/expA' --save_videos --first_seed_only
+python -m eval_experiments --method hail --folder='test_policies/hail/expA' --save_videos --first_seed_only
+python -m eval_experiments --method shail --folder='test_policies/shail/expA' --save_videos --first_seed_only
+
+# Experiment B
+python -m eval_experiments --locations='[(0,4)]' 
+python -m eval_experiments --method expert_agent --locations='[(0,4)]' --save_videos --first_seed_only
+python -m eval_experiments --method idm --locations='[(0,4)]' --save_videos --first_seed_only
+python -m eval_experiments --method bc --folder='test_policies/bc/expB' --locations='[(0,4)]' --save_videos --first_seed_only
+python -m eval_experiments --method gail --folder='test_policies/gail/expB' --locations='[(0,4)]' --save_videos --first_seed_only
+python -m eval_experiments --method hail --folder='test_policies/hail/expB' --locations='[(0,4)]' --save_videos --first_seed_only
+python -m eval_experiments --method shail --folder='test_policies/shail/expB' --locations='[(0,4)]' --save_videos --first_seed_only

intersimple-expert-rollout-setobs2.py

@@ -0,0 +1,65 @@
+import torch
+import functools
+from src.core.sampling import rollout_sb3
+from intersim.envs import IntersimpleLidarFlatIncrementingAgent
+from intersim.envs.intersimple import speed_reward
+from intersim.expert import NormalizedIntersimpleExpert
+from src.util.wrappers import CollisionPenaltyWrapper, Setobs
+import numpy as np
+from gym.wrappers import TransformObservation
+
+obs_min = np.array([
+    [-1000, -1000, 0, -np.pi, -1e-1, 0.],
+    [0, -np.pi, -20, -20, -np.pi, -1e-1],
+    [0, -np.pi, -20, -20, -np.pi, -1e-1],
+    [0, -np.pi, -20, -20, -np.pi, -1e-1],
+    [0, -np.pi, -20, -20, -np.pi, -1e-1],
+    [0, -np.pi, -20, -20, -np.pi, -1e-1],
+]).reshape(-1)
+
+obs_max = np.array([
+    [1000, 1000, 20, np.pi, 1e-1, 0.],
+    [50, np.pi, 20, 20, np.pi, 1e-1],
+    [50, np.pi, 20, 20, np.pi, 1e-1],
+    [50, np.pi, 20, 20, np.pi, 1e-1],
+    [50, np.pi, 20, 20, np.pi, 1e-1],
+    [50, np.pi, 20, 20, np.pi, 1e-1],
+]).reshape(-1)
+
+def main(track:int, loc:int=0):
+    env = IntersimpleLidarFlatIncrementingAgent(
+        loc=loc,
+        track=track,
+        n_rays=5,
+        reward=functools.partial(
+            speed_reward,
+            collision_penalty=0
+        ),
+    )
+
+    policy = NormalizedIntersimpleExpert(env, mu=0.001)
+
+    env = Setobs(TransformObservation(
+        CollisionPenaltyWrapper(
+            env,
+            collision_distance=6, collision_penalty=100
+        ), lambda obs: (obs - obs_min) / (obs_max - obs_min + 1e-10)
+    ))
+    print(env.nv, 'vehicles')
+    expert_data = rollout_sb3(env, policy, n_episodes=150, max_steps_per_episode=200)
+
+    states, actions, rewards, dones = expert_data
+    print(f'Expert mean episode length {(~dones).sum() / states.shape[0]}')
+    print(f'Expert mean reward per episode {rewards[~dones].sum() / states.shape[0]}')
+    print(f'Observation mean', states[~dones].mean(0))
+    print(f'Observation std', states[~dones].std(0))
+
+    torch.save(expert_data, f'intersimple-expert-data-setobs2-loc{loc}-track{track}.pt')
+
+def loop(tracks:list=[0]):
+    for track in tracks:
+        main(track)
+
+if __name__=='__main__':
+    import fire
+    fire.Fire(loop)

scratch/arec/intersimple/commands.txt

@@ -1,22 +0,0 @@
-python -m render_options --model_name='gail_options_image_mid_wcollision' --env='NRasterizedRoute' --options=True --width=36 --height=36 --m_per_px=2 --agent=50 --stop_on_collision=False
-
-import torch, os
-from src.data import load_experts
-folder = 'expert_data/DR_USA_Roundabout_FT/track0000'
-single_agent = os.path.join(folder, 'expert.pkl')
-multi_agent = os.path.join(folder,'joint_expert_states.pt')
-multi_agent_actions = os.path.join(folder,'joint_expert_actions.pt')
-demonstrations = load_experts([single_agent], flatten=False)
-demonstrations[0].__dict__.keys()
-len(demonstrations[0].obs)
-single_agent_lengths = [len(demonstration.obs) for demonstration in demonstrations]
-states = torch.load(multi_agent)
-actions = torch.load(multi_agent_actions)
-multi_agent_lengths = [sum(~torch.isnan(states[:,i,0])).item() for i in range(states.shape[1])]
-
-single_agent_actions = [demonstration.acts for demonstration in demonstrations]
-multi_agent_actions = [actions[~torch.isnan(actions[:,i,0])] for i in range(actions.shape[1])]
-
-import pickle
-with open(single_agent, "rb") as f:
-    new_trajectories = pickle.load(f)

scratch/arec/intersimple/data/expert.py

@@ -1,145 +0,0 @@
-from intersim.envs.intersimple import Intersimple
-from stable_baselines3.common.policies import BasePolicy
-import gym
-import intersim.envs.intersimple
-import imitation.data.rollout as rollout
-from stable_baselines3.common.vec_env.dummy_vec_env import DummyVecEnv
-from imitation.data.wrappers import RolloutInfoWrapper
-
-class IntersimExpert(BasePolicy):
-    
-    def __init__(self, intersim_env, mu=0, *args, **kwargs):
-        super().__init__(
-            observation_space=gym.spaces.Space(),
-            action_space=gym.spaces.Space(),
-            *args, **kwargs
-        )
-        self._intersim = intersim_env
-        self._mu = mu
-
-    def forward(self, *args, **kwargs):
-        raise NotImplementedError()
-
-    def _predict(self, *args, **kwargs):
-        raise NotImplementedError()
-
-    def _action(self):
-        target_t = min(self._intersim._ind + 1, len(self._intersim._svt.simstate) - 1)
-        target_state = self._intersim._svt.simstate[target_t]
-        return self._intersim.target_state(target_state, mu=self._mu)
-
-    def predict(self, *args, **kwargs):
-        return self._action(), None
-
-class IntersimpleExpert(BasePolicy):
-
-    def __init__(self, intersimple_env, mu=0, *args, **kwargs):
-        super().__init__(
-            observation_space=intersimple_env.observation_space,
-            action_space=intersimple_env.action_space,
-            *args, **kwargs
-        )
-        self._intersimple = intersimple_env
-        self._intersim_expert = IntersimExpert(intersimple_env._env, mu=mu)
-    
-    def forward(self, *args, **kwargs):
-        raise NotImplementedError()
-
-    def _predict(self, *args, **kwargs):
-        raise NotImplementedError()
-
-    def _action(self):
-        return self._intersim_expert._action()[self._intersimple._agent]
-
-    def predict(self, *args, **kwargs):
-        return self._action(), None
-
-class NormalizedIntersimpleExpert(IntersimpleExpert):
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-    def predict(self, *args, **kwargs):
-        action, _ = super().predict(*args, **kwargs)
-        return self._intersimple._normalize(action), None
-
-class DummyVecEnvPolicy(BasePolicy):
-
-    def __init__(self, experts):
-        self._experts = [e() for e in experts]
-
-    def forward(self, *args, **kwargs):
-        raise NotImplementedError()
-
-    def _predict(self, *args, **kwargs):
-        raise NotImplementedError()
-    
-    def predict(self, *args, **kwargs):
-        predictions = [e.predict() for e in self._experts]
-        actions = [p[0] for p in predictions]
-        states = [p[1] for p in predictions]
-        return actions, states
-    
-    def forward(self, *args, **kwargs):
-        raise NotImplementedError()
-
-    def _predict(self, *args, **kwargs):
-        raise NotImplementedError()
-
-def save_video(env, expert):
-    env.reset()
-    env.render()
-    done = False
-    while not done:
-        actions, _ = expert.predict()
-        _, _, done, _ = env.step(actions)
-        env.render()
-    env.close()
-
-def demonstrations(expert='NormalizedIntersimpleExpert', env='NRasterizedIncrementingAgent', path=None, min_timesteps=None, min_episodes=None, video=False, env_args={}, policy_args={}):
-    """Rollout and save expert demos.
-    
-    Usage:
-        python -m intersimple.expert <flags>
-    Args: 
-        expert (class): class of expert
-        env (class): class of env intersim.envs.intersimple
-        path (str): path to store output
-        min_timesteps (int): min number of timesteps for call to rollout.rollout_and_save
-        min_episodes (int): min number of episodes for call to rollout.rollout_and_save
-        video (bool): whether to save a video of the expert until a single environment instantiation stops
-        env_args (dict): dictionary of kwargs when instantiating environment class
-        policy_args (dict): dictionary of kwargs when instantiating Expert policy
-    """
-    
-    Env = intersim.envs.intersimple.__dict__[env]
-    Expert = globals()[expert]
-
-    env = Env(**env_args)
-    info_env = RolloutInfoWrapper(env) # getting rollout info (dictionary) from environment
-    venv = DummyVecEnv([lambda: info_env]) # making a DummyVecEnv with a list of a function that when called returns the rollout info
-
-    policy = Expert(env, **policy_args) # instantiate an expert policy from specified class with instantiated environment and policy kwargs
-    venv_policy = DummyVecEnvPolicy([lambda: policy]) # make a DummyVecEnvPolicy with a list of a function that when called returns the Expert policy
-
-    if min_timesteps is None and min_episodes is None:
-        min_episodes = env.nv # one episode per vehicle being controlled in environment (hopefully an incrementing agent environment)
-
-    if video:
-        save_video(env, policy)
-    
-    path = path or (policy.__class__.__name__ + '_' + env.__class__.__name__ + '.pkl')
-    suntil = rollout.make_sample_until(
-            min_timesteps=min_timesteps,
-            min_episodes=min_episodes,
-        )
-    rollout.rollout_and_save(
-        path=path,
-        policy=venv_policy,
-        venv=venv,
-        sample_until=suntil
-    )
-
-if __name__ == '__main__':
-    import fire
-    fire.Fire(demonstrations)

scratch/arec/intersimple/data/generate.sh

@@ -1,9 +0,0 @@
-#python -m expert --env=IntersimpleReward --min_timesteps=200 --env_args='{agent:51}' --path='NormalizedIntersimpleExpert_IntersimpleRewardAgent51.pkl'
-#python -m expert --env=IntersimpleReward --min_timesteps=200 --env_args='{agent:51}' --policy_args='{mu:0.005}' --path='NormalizedIntersimpleExpert_IntersimpleRewardAgent51Mu.005.pkl'
-#python -m expert --env=IntersimpleReward --min_timesteps=200 --env_args='{agent:51}' --policy_args='{mu:0.001}' --path='NormalizedIntersimpleExpert_IntersimpleRewardAgent51Mu.001.pkl'
-#python -m expert --env=NRasterized --min_timesteps=200 --env_args='{agent:51,width:36,height:36,m_per_px:2}' --policy_args='{mu:0.001}' --path='NormalizedIntersimpleExpertMu.001_NRasterizedAgent51w36h36mppx2.pkl'
-# python -m expert --env=NRasterizedRandomAgent --min_timesteps=10000 --env_args='{width:36,height:36,m_per_px:2}' --policy_args='{mu:0.001}' --path='NormalizedIntersimpleExpertMu.001N10000_NRasterizedRandomAgentw36h36mppx2.pkl'
-#python -m expert --env=NRasterized --min_timesteps=200 --env_args='{agent:51,width:36,height:36,m_per_px:2}' --policy_args='{mu:0.001}' --path='NormalizedIntersimpleExpertMu.001_NRasterizedAgent51w36h36mppx2.pkl'
-#python -m expert --env=NRasterized --min_timesteps=3000 --video --env_args='{width:36,height:36,m_per_px:2}' --policy_args='{mu:0.001}' --path='NormalizedIntersimpleExpertMu.001_NRasterizedRandomAgentw36h36mppx2.pkl'
-#python -m expert --env=NRasterizedIncrementingAgent --env_args='{width:36,height:36,m_per_px:2}' --policy_args='{mu:0.001}' --path='NormalizedIntersimpleExpertMu.001_NRasterizedIncrementingAgentw36h36mppx2.pkl'
-python -m process_all_experts --env_args='{width:36,height:36,m_per_px:2}' --policy_args='{mu:0.001}'

scratch/arec/intersimple/data/load_experts.py

@@ -1,23 +0,0 @@
-import pickle
-import imitation.data.rollout as rollout
-from tqdm import tqdm
-
-def load_experts(expert_files=[]):
-    """
-    Load expert trajectories from files and combine their transitions into a single RB
-
-    Args:
-        expert_files (list): list of expert file strings
-    Returns:
-        transitions (list): list of combined expert episode transitions 
-    """
-    transitions = []
-    for file in tqdm(expert_files):
-        with open(file, "rb") as f:
-            trajectories = pickle.load(f)
-        transitions = transitions + rollout.flatten_trajectories(trajectories)
-    return transitions
-
-if __name__=='__main__':
-    import fire
-    fire.Fire(load_experts)

scratch/arec/intersimple/data/process_all_experts.py

@@ -1,48 +0,0 @@
-import tqdm
-import expert
-import copy
-import os
-import intersim
-from tqdm import tqdm
-
-def process_all_experts(filename='expert.pkl',env_args={}, policy_args={}):
-    """
-    Process all experts in the Interaction Dataset
-    For now, using NormalizedIntersimpleExpert with NRasterizedIncrementingAgent environment
-
-    Args:
-        filename (str): name for track file
-        env_args (dict): default environment kwargs
-        policy_args (dict): default policy kwargs
-    """
-    I, J = len(intersim.LOCATIONS), intersim.MAX_TRACKS
-    pbar = tqdm(total=I*J)
-    for loc in range(I):
-        for track in range(J):
-            
-            it_env_args = copy.deepcopy(env_args)
-            it_env_args.update({
-                'loc':loc,
-                'track':track,
-            })
-            out_folder = os.path.join(intersim.LOCATIONS[loc], 'track%04i'%(track))
-            if not os.path.isdir(out_folder):
-                os.makedirs(out_folder) 
-            it_path = os.path.join(out_folder,filename)
-
-            expert.demonstrations(
-                expert='NormalizedIntersimpleExpert', 
-                env='NRasterizedIncrementingAgent', 
-                path=it_path, 
-                env_args=it_env_args, 
-                policy_args=policy_args,
-                )
-            pbar.update(1)
-    pbar.close()
-    
-
-if __name__=='__main__':
-    import fire
-    fire.Fire(process_all_experts)
-
-

scratch/arec/intersimple/gail/discriminator.py

@@ -1,101 +0,0 @@
-import torch
-
-# imitation.rewards.discrim_nets.DiscrimNetGAIL is composed of self.discriminator (nn.Module),
-# which gets called with inputs (state, action) when needed.
-
-class CnnDiscriminator(torch.nn.Module):
-    """ConvNet similar to stable_baselines3.common.policies.ActorCriticCnnPolicy."""
-
-    def __init__(self, env):
-        super().__init__()
-
-        obs_channels, _, _ = env.observation_space.shape
-        (action_size,) = env.action_space.shape
-        in_channels = obs_channels + action_size
-
-        self.cnn = torch.nn.Sequential(
-            torch.nn.Conv2d(in_channels, 32, kernel_size=(8, 8), stride=(4, 4)), # 5+1 -> 32
-            torch.nn.ReLU(),
-            torch.nn.Conv2d(32, 64, kernel_size=(4, 4), stride=(2, 2)), # 32 -> 64
-            torch.nn.ReLU(),
-            torch.nn.Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1)), # 64 -> 64
-            torch.nn.ReLU(),
-            torch.nn.Flatten(start_dim=1, end_dim=-1),
-            torch.nn.LazyLinear(512), # 28224 -> 512
-            torch.nn.ReLU(),
-            torch.nn.LazyLinear(1), # 512 -> 1
-        )
-    
-    @staticmethod
-    def _concatenate(state, action):
-        b, _, h, w = state.shape
-        _, a = action.shape
-        act = action.unsqueeze(-1).unsqueeze(-1).expand((b, a, h, w))
-        sa = torch.cat((state, act), -3)
-        return sa
-
-    def forward(self, state, action):
-        sa = self._concatenate(state, action)
-        assert sa.ndim == 4
-        return self.cnn(sa).squeeze(1)
-
-class CnnDiscriminatorFlatAction(torch.nn.Module):
-    """ConvNet similar to stable_baselines3.common.policies.ActorCriticCnnPolicy."""
-
-    def __init__(self, env):
-        super().__init__()
-
-        obs_channels, _, _ = env.observation_space.shape
-        (action_size,) = env.action_space.shape
-        in_channels = obs_channels
-
-        self.cnn = torch.nn.Sequential(
-            torch.nn.Conv2d(in_channels, 32, kernel_size=(8, 8), stride=(4, 4)), # in_channels -> 32
-            torch.nn.ReLU(),
-            torch.nn.Conv2d(32, 64, kernel_size=(4, 4), stride=(2, 2)), # 32 -> 64
-            torch.nn.ReLU(),
-            torch.nn.Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1)), # 64 -> 64
-            torch.nn.ReLU(),
-            torch.nn.Flatten(start_dim=1, end_dim=-1),
-            torch.nn.LazyLinear(128), # 28224 -> 128
-        )
-        self.decoder = torch.nn.Sequential(
-            torch.nn.LazyLinear(64), #128 + 2 -> 64
-            torch.nn.ReLU(),
-            torch.nn.LazyLinear(64), #64 -> 64
-            torch.nn.ReLU(),
-            torch.nn.LazyLinear(1) #64 -> 1
-        )
-    
-    @staticmethod
-    def _concatenate(state, action):
-        b, s= state.shape
-        b, a = action.shape
-        sa = torch.cat((state, action), -1)
-        return sa
-
-    def forward(self, state, action):
-        s = self.cnn(state.float())
-        sa = self._concatenate(s, action)
-        assert sa.ndim == 2
-        return self.decoder(sa).squeeze(1)
-
-class MlpDiscriminator(torch.nn.Module):
-    """MLP similar to stable_baselines3.common.policies.ActorCriticPolicy."""
-
-    def __init__(self, env=None):
-        super().__init__()
-        self.flatten = torch.nn.Flatten(start_dim=1, end_dim=-1)
-        self.mlp = torch.nn.Sequential(
-            torch.nn.LazyLinear(64), # 42 -> 64
-            torch.nn.Tanh(),
-            torch.nn.LazyLinear(64), # 64 -> 64
-            torch.nn.Tanh(),
-            torch.nn.LazyLinear(1),  # 64 -> 1
-        )
-    
-    def forward(self, state, action):
-        flat = self.flatten(state)
-        sa = torch.cat((action, flat), -1)
-        assert sa.ndim == 2
-        return self.mlp(sa).squeeze(1)

scratch/arec/intersimple/gail/test_discriminator.py

@@ -1,45 +0,0 @@
-from intersim.envs.intersimple import NRasterized
-from discriminator import CnnDiscriminator
-import torch
-
-def test_image_concatenation():
-    env = NRasterized()
-    disc = CnnDiscriminator(env)
-    s = torch.tensor(env.reset()).unsqueeze(0)
-    a = torch.tensor([[0.5]])
-    sa = disc._concatenate(s, a)
-
-    assert s.shape == (1, 5, 200, 200)
-    assert a.shape == (1, 1)
-    assert sa.shape == (1, 6, 200, 200)
-    assert torch.allclose(sa[:, :5], 1.0 * s)
-    assert (sa[:, 5] == a.unsqueeze(-1)).all()
-
-def test_image_concatenation3():
-    env = NRasterized()
-    disc = CnnDiscriminator(env)
-
-    s1 = env.reset()
-    a1 = 0.15
-    s2, _, _, _ = env.step(0.9)
-    a2 = 0.25
-    s3, _, _, _ = env.step(-0.9)
-    a3 = 0.35
-    
-    s = torch.stack([
-        torch.tensor(s1),
-        torch.tensor(s2),
-        torch.tensor(s3)
-    ], axis=0)
-    a = torch.tensor([
-        [a1],
-        [a2],
-        [a3],
-    ])
-    sa = disc._concatenate(s, a)
-
-    assert s.shape == (3, 5, 200, 200)
-    assert a.shape == (3, 1)
-    assert sa.shape == (3, 6, 200, 200)
-    assert torch.allclose(sa[:, :5], 1.0 * s)
-    assert (sa[:, 5] == a.unsqueeze(-1)).all()

scratch/arec/intersimple/gail_image_multiagent_nocollision.py

@@ -1,70 +0,0 @@
-# %%
-import pathlib
-import pickle
-import tempfile
-
-import stable_baselines3 as sb3
-from stable_baselines3.common.env_util import make_vec_env
-
-from imitation.algorithms import adversarial, bc
-from imitation.data import rollout
-from imitation.util import logger
-
-from intersim.envs.intersimple import NRasterized
-
-from gail.discriminator import CnnDiscriminatorFlatAction
-
-model_name = 'gail_image_multiagent_nocollision'
-
-# %%
-# Load pickled test demonstrations.
-with open("data/NormalizedIntersimpleExpertMu.001_NRasterizedRandomAgentw36h36mppx2.pkl", "rb") as f:
-    # This is a list of `imitation.data.types.Trajectory`, where
-    # every instance contains observations and actions for a single expert
-    # demonstration.
-    trajectories = pickle.load(f)
-
-# %%
-# Convert List[types.Trajectory] to an instance of `imitation.data.types.Transitions`.
-# This is a more general dataclass containing unordered
-# (observation, actions, next_observation) transitions.
-transitions = rollout.flatten_trajectories(trajectories)
-
-venv = make_vec_env(NRasterized, n_envs=2, env_kwargs={'stop_on_collision':False, 'width': 36, 'height': 36, 'm_per_px': 2})
-
-tempdir = tempfile.TemporaryDirectory(prefix="quickstart")
-tempdir_path = pathlib.Path(tempdir.name)
-print(f"All Tensorboards and logging are being written inside {tempdir_path}/.")
-
-# Train GAIL on expert data.
-# GAIL, and AIRL also accept as `expert_data` any Pytorch-style DataLoader that
-# iterates over dictionaries containing observations, actions, and next_observations.
-logger.configure(tempdir_path / "GAIL/")
-gail_trainer = adversarial.GAIL(
-    venv,
-    expert_data=transitions,
-    expert_batch_size=32,
-    #n_disc_updates_per_round=2048,
-    discrim_kwargs={'discrim_net': CnnDiscriminatorFlatAction(venv)},
-    gen_algo=sb3.PPO("CnnPolicy", venv, verbose=1, n_steps=1024),
-    allow_variable_horizon=True,
-)
-gail_trainer.train(total_timesteps=100000)
-gail_trainer.gen_algo.save(model_name)
-
-#del gail_trainer
-
-# %%
-model = sb3.PPO.load(model_name)
-
-env = NRasterized(stop_on_collision=False, width=36, height=36, m_per_px=2)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)

scratch/arec/intersimple/gail_image_singleagent_nocollision.py

@@ -1,70 +0,0 @@
-# %%
-import pathlib
-import pickle
-import tempfile
-
-import stable_baselines3 as sb3
-from stable_baselines3.common.env_util import make_vec_env
-
-from imitation.algorithms import adversarial, bc
-from imitation.data import rollout
-from imitation.util import logger
-
-from intersim.envs.intersimple import NRasterized
-
-from gail.discriminator import CnnDiscriminator
-
-model_name = 'gail_image_singleagent_nocollision'
-
-# %%
-# Load pickled test demonstrations.
-with open("data/NormalizedIntersimpleExpertMu.001_NRasterizedAgent51w36h36mppx2.pkl", "rb") as f:
-    # This is a list of `imitation.data.types.Trajectory`, where
-    # every instance contains observations and actions for a single expert
-    # demonstration.
-    trajectories = pickle.load(f)
-
-# %%
-# Convert List[types.Trajectory] to an instance of `imitation.data.types.Transitions`.
-# This is a more general dataclass containing unordered
-# (observation, actions, next_observation) transitions.
-transitions = rollout.flatten_trajectories(trajectories)
-
-venv = make_vec_env(NRasterized, n_envs=2, env_kwargs={'agent':51, 'stop_on_collision':False, 'width': 36, 'height': 36, 'm_per_px': 2})
-
-tempdir = tempfile.TemporaryDirectory(prefix="quickstart")
-tempdir_path = pathlib.Path(tempdir.name)
-print(f"All Tensorboards and logging are being written inside {tempdir_path}/.")
-
-# Train GAIL on expert data.
-# GAIL, and AIRL also accept as `expert_data` any Pytorch-style DataLoader that
-# iterates over dictionaries containing observations, actions, and next_observations.
-logger.configure(tempdir_path / "GAIL/")
-gail_trainer = adversarial.GAIL(
-    venv,
-    expert_data=transitions,
-    expert_batch_size=32,
-    #n_disc_updates_per_round=2048,
-    discrim_kwargs={'discrim_net': CnnDiscriminator(venv)},
-    gen_algo=sb3.PPO("CnnPolicy", venv, verbose=1, n_steps=1024),
-    allow_variable_horizon=True,
-)
-gail_trainer.train(total_timesteps=100000)
-gail_trainer.gen_algo.save(model_name)
-
-#del gail_trainer
-
-# %%
-model = sb3.PPO.load(model_name)
-
-env = NRasterized(agent=51, width=36, height=36, m_per_px=2, stop_on_collision=False)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)

scratch/arec/intersimple/gail_options_image.py

@@ -1,172 +0,0 @@
-# %%
-import sys
-sys.path.append('../../../')
-from src.discriminator import CnnDiscriminator, CnnDiscriminatorFlatAction
-from src.policies import OptionsCnnPolicy
-from src.util import render_env
-from src.data import load_experts
-from src.gail.options import OptionsEnv, LLOptions, HLOptions, RenderOptions
-from src.gail.train import train_discriminator, train_generator
-
-from imitation.algorithms import adversarial
-from imitation.util import logger
-import imitation.data.rollout as rollout
-
-import stable_baselines3
-from stable_baselines3.common.env_util import make_vec_env
-
-import torch
-import torch.utils.data
-import numpy as np
-import itertools
-import gym
-import pickle
-import tempfile
-import pathlib
-from tqdm import tqdm
-
-from intersim.envs.intersimple import NRasterized, NRasterizedRoute, NRasterizedRandomAgent, NRasterizedIncrementingAgent, NRasterizedRouteRandomAgent
-
-ALL_OPTIONS = [(v,t) for v in [0,2,4,6,8] for t in [5, 10]] # option 0 is safe fallback
-
-def flatten_transitions(transitions):
-    return {
-        'obs': np.stack(list(t['obs'] for t in transitions), axis=0),
-        'next_obs': np.stack(list(t['next_obs'] for t in transitions), axis=0),
-        'acts': np.stack(list(t['acts'] for t in transitions), axis=0),
-        'dones': np.stack(list(t['dones'] for t in transitions), axis=0),
-    }
-
-def train(expert_data, env_class=NRasterizedRouteRandomAgent, env_settings={}, 
-        epochs=10, discrim_batch_size=32, generator_steps=2048, discount=0.99):
-    """
-    Args: 
-        expert_data: list of transitions
-        env_class: environment class
-        env_settings: environment settings
-        epochs: number of epochs to train for
-        discrim_batch_size: discriminator batch size
-        generator_steps: number of steps taken in generator
-        discount: discount factor
-    Returns:
-        generator (stable_baselines3.PPO): options policy
-    """
-    env = env_class(**env_settings)
-    env.discount = discount
-
-    tempdir = tempfile.TemporaryDirectory(prefix="quickstart")
-    tempdir_path = pathlib.Path(tempdir.name)
-    logger.configure(tempdir_path / "GAIL/")
-    print(f"All Tensorboards and logging are being written inside {tempdir_path}/.")
-
-    venv = make_vec_env(env_class, n_envs=1, env_kwargs=env_settings)
-    discriminator = adversarial.GAIL(
-        expert_data=expert_data,
-        expert_batch_size=discrim_batch_size,
-        discrim_kwargs={'discrim_net': CnnDiscriminatorFlatAction(venv)},
-        #discrim_kwargs={'discrim_net': CnnDiscriminator(venv)},
-        venv=venv, # unused
-        gen_algo=stable_baselines3.PPO("CnnPolicy", venv), # unused
-    )
-
-    generator = stable_baselines3.PPO(
-        OptionsCnnPolicy,
-        OptionsEnv(env, options=ALL_OPTIONS),
-        verbose=1,
-        n_steps=generator_steps,
-    )
-
-    # PPO.train requires logger as set up in
-    # PPO._setup_learn (called by PPO.learn)
-    generator._logger = stable_baselines3.common.utils.configure_logger(
-        generator.verbose,
-        generator.tensorboard_log,
-    )
-
-    for _ in tqdm(range(epochs)):
-        train_discriminator(LLOptions(env, options=ALL_OPTIONS), generator, discriminator, num_samples=discrim_batch_size)
-        train_generator(HLOptions(env, options=ALL_OPTIONS), generator, discriminator, num_samples=generator_steps)
-    
-    return generator
-
-# %%
-if __name__ == '__main__':
-    # %%
-    model_name = 'gail_options_image_mid_wcollision'
-    env_class = NRasterizedRouteRandomAgent
-    env_settings = {'width': 36, 'height': 36, 'm_per_px': 2, 'stop_on_collision': False}
-    
-    #env_class = NRasterized
-    #env_settings = {'agent': 51, 'width': 36, 'height': 36, 'm_per_px': 2}
-    files = ['../../../expert_data/DR_USA_Roundabout_FT/track%04i/expert.pkl'%(i) for i in range(5)]
-    transitions=load_experts(files)
-
-    generator = train(
-        transitions,
-        env_class=env_class,
-        env_settings=env_settings,
-        epochs=2, 
-        discrim_batch_size=256, 
-        generator_steps=10,#256, 
-        discount=0.99
-        )
-
-    generator.save(model_name)
-
-    # Render
-    render_settings = {'width': 36, 'height': 36, 'm_per_px': 2, 'agent':51, 'stop_on_collision': False}
-    render_env(model_name=model_name, env='NRasterizedRoute', options=True, options_list=ALL_OPTIONS,
-        **render_settings)
-
-
-# %% Tests
-
-def test_ll_expert_data():
-    with open("data/NormalizedIntersimpleExpertMu.001_NRasterizedAgent51w36h36mppx2.pkl", "rb") as f:
-        expert_trajectories = pickle.load(f)
-    expert_transitions = rollout.flatten_trajectories(expert_trajectories)
-
-    env = LLOptions(NRasterized(agent=51, width=36, height=36, m_per_px=2))
-
-    gen_transitions = list(itertools.islice(env.sample_ll(
-        policy=stable_baselines3.PPO(
-            OptionsCnnPolicy,
-            OptionsEnv(env),
-            verbose=1,
-        )
-    ), 10))
-    gen_transitions = flatten_transitions(gen_transitions)
-
-    assert expert_transitions[:10].obs.shape == gen_transitions['obs'].shape
-    assert expert_transitions[:10].next_obs.shape == gen_transitions['next_obs'].shape
-    assert expert_transitions[:10].acts.shape == gen_transitions['acts'].shape
-    assert expert_transitions[:10].dones.shape == gen_transitions['dones'].shape
-
-def test_ll_states():
-    env = NRasterized()
-    policy = stable_baselines3.PPO(
-        OptionsCnnPolicy,
-        OptionsEnv(env),
-        verbose=1,
-    )
-    llenv = LLOptions(env)
-    transitions = list(itertools.islice(llenv.sample_ll(policy=policy), 100))
-
-    env2 = NRasterized()
-    s2 = env2.reset()
-    for i, t in enumerate(transitions):
-        assert i == 0 or np.array_equal(t['obs'], transitions[i-1]['next_obs'])
-        assert np.array_equal(t['obs'], s2)
-        assert t['acts'].shape == (1,)
-
-        nexts2, _, done2, _ = env2.step(t['acts'])
-        assert np.array_equal(t['next_obs'], nexts2)
-        assert np.array_equal(t['dones'], done2)
-
-        if done2:
-            break
-
-        s2 = nexts2
-
-def test_hl_transitions():
-    pass

scratch/arec/intersimple/gail_options_scratch.py

@@ -1,559 +0,0 @@
-# %%
-from gail.discriminator import CnnDiscriminator, CnnDiscriminatorFlatAction
-from imitation.algorithms import adversarial
-import stable_baselines3
-import torch.utils.data
-import numpy as np
-from intersim.envs.intersimple import NRasterized, NRasterizedRandomAgent
-import itertools
-from torch.distributions import Categorical
-import gym
-import torch
-import pickle
-import imitation.data.rollout as rollout
-import tempfile
-import pathlib
-from imitation.util import logger
-from stable_baselines3.common.env_util import make_vec_env
-from tqdm import tqdm
-
-import logging 
-logging.basicConfig(level=logging.DEBUG)
-
-ALL_OPTIONS = [(v,t) for v in [0,2,4,6,8] for t in [5, 10, 20]] # option 0 is safe fallback
-
-class OptionsCnnPolicy(stable_baselines3.common.policies.ActorCriticCnnPolicy):
-    """
-    Class for high-level options policy (generator)
-    """
-    def __init__(self, observation_space, *args, **kwargs):
-        super().__init__(observation_space['obs'], *args, **kwargs)
-
-    def _prior_distribution(self, s):
-        """
-        Return prior distribution over high-level options (before masking)
-        Args:
-            s (torch.tensor): observation
-        Returns:
-            values (torch.tensor): values from critic
-            dist (torch.distributions): prior distribution over actions
-        """
-        latent_pi, latent_vf, latent_sde = self._get_latent(s) 
-        distribution = self._get_action_dist_from_latent(latent_pi, latent_sde) 
-        values = self.value_net(latent_vf)
-        return values, distribution.distribution
-
-    def predict(self, obs):
-        """
-        Will mask invalid states before making action selections
-        Args:
-            obs: dict with keys:
-                obs (torch.tensor): (B,o)  true observations
-                mask (torch.tensor): (B,m) mask over valid actions
-        Returns:
-            ch (torch.tensor): (B,a) sampled actions
-            values (torch.tensor): (B,) predicted value at observation
-            log_probs (torch.tensor): (B,) log probabilities of selected actions
-        """
-        s, m = obs['obs'], obs['mask']
-        values, prior = self._prior_distribution(s)
-        posterior = Categorical(prior.probs * m)
-        ch = posterior.sample()
-        return ch, values, posterior.log_prob(ch)
-
-    def evaluate_actions(self, obs, ch):
-        """
-        Evaluate particular actions
-        Args:
-            obs: dict with keys:
-                obs (torch.tensor): (B,o) true observations
-                mask (torch.tensor): (B,m) masks over valid actions
-            ch (torch.tensor): (B,a) selected actions
-        Returns:
-            values (torch.tensor): (B,) predicted value at observation
-            log_probs (torch.tensor): (B,) log probabilities of selected actions
-            ent (torch.tensor): (B,) entropy of each distribution over actions
-        """
-        s, m = obs['obs'], obs['mask']
-        values, prior = self._prior_distribution(s)
-        posterior = Categorical(prior.probs * m)
-        return values, posterior.log_prob(ch), posterior.entropy() # additional values used by PPO.train
-
-class OptionsEnv(gym.Wrapper):
-    """
-    Wrap an intersimple environment with an options generator
-    """
-    def __init__(self, env, *args, **kwargs):
-        """
-        Initialize wrapped environment and set high-level action and observation spaces
-        """
-        super().__init__(env, *args, **kwargs)
-        num_hl_options = len(ALL_OPTIONS)
-        self.action_space = gym.spaces.Discrete(num_hl_options)
-        self.observation_space = gym.spaces.Dict({
-            'obs': env.observation_space,
-            'mask': gym.spaces.Box(low=0, high=1, shape=(num_hl_options,)),
-        })
-
-    def _after_choice(self):
-        pass
-    
-    def _after_step(self):
-        pass
-
-    def _transitions(self):
-        raise NotImplementedError('Use `LLOptions` or `HLOptions` for sampling.')
-    
-    def sample(self, generator):
-        """
-        yield transitions using a generator
-        Args:
-            generator (sb3.PPO)
-        Yields:
-
-        """
-        self.done = True
-        while True:
-            self.episode_start = False
-
-            if self.done:
-                # reset environment
-                self.s = self.env.reset()
-                self.m = available_actions(self.env)
-                self.done = False
-                self.episode_start = True
-
-            # set the action, the value of the start state, and the logprob of the action
-            # according to the current environment state and mask
-            self.ch, self.value, self.log_prob = generator.policy.predict({
-                'obs': torch.tensor(self.s).unsqueeze(0).to(generator.policy.device),
-                'mask': torch.tensor(self.m).unsqueeze(0).to(generator.policy.device),
-            })
-
-            # store a float list of actions to take given the option selected in the environment
-            self.plan = list(map(float, generate_plan(self.env, self.ch)))
-
-            # run whatever _after_choice might dictate in a child class
-            self._after_choice()
-
-            # some checks
-            assert not self.done
-            assert self.plan
-            assert feasible(self.env, self.plan, self.ch)
-
-            # execute the option so long as the episode isn't complete and the plan is still feasible
-            while not self.done and self.plan and feasible(self.env, self.plan, self.ch):
-                
-                # pop first action
-                self.a, self.plan = self.plan[0], self.plan[1:]
-                
-                # normalize action ??
-                self.a = self.env._normalize(self.a)
-                
-                # step through environment
-                self.nexts, _, self.done, _ = self.env.step(self.a)
-                self.nextm = available_actions(self.env)
-
-                # run whatever _after_step might dictate in child class
-                self._after_step()
-
-                # update state and mask to current
-                self.s = self.nexts
-                self.m = self.nextm
-
-            # transitions yielded from self._transitions() functions specied in child classes
-            yield from self._transitions()
-
-            ### NOTE: only yields after a full option has been executed / exited
-
-class LLOptions(OptionsEnv):
-    """Sample low-level (state, action) tuples for discriminator training."""
-    
-    def __init__(self, *args, **kwargs):
-        """
-        LLOption uses the true LL observations
-        """
-        super().__init__(*args, **kwargs)
-        # overwrite observation space to just output obs directly
-        self.observation_space = self.observation_space['obs'] 
-
-    def _after_choice(self):
-        """
-        After each option choice, initialize/reset the transition buffer
-        """
-        self._transition_buffer = []
-
-    def _after_step(self):
-        """
-        After each ll action, append s, s', a, done to transition buffer
-        """
-        self._transition_buffer.append({
-            'obs': self.s,
-            'next_obs': self.nexts,
-            'acts': np.array((self.a,)),
-            'dones': np.array(self.done),
-        })
-
-    def _transitions(self):
-        """
-        Yield from the transition buffer
-        """
-        yield from self._transition_buffer
-    
-    def sample_ll(self, policy):
-        """
-        Args:
-            policy
-        Returns:
-            gen: iterable which samples low-level transitions from the environment
-        """
-        return self.sample(policy)
-
-class HLOptions(OptionsEnv):
-    """Sample high-level (state, action, reward) tuples for generator training."""
-    
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-    def _after_choice(self):
-        """
-        After an option selection, initialize total reward and number of steps
-        """
-        self.r = 0
-        self.steps = 0
-
-    def _after_step(self):
-        """
-        After each low-level action, add the discounted discriminated reward score (given a discriminator)
-        """
-        self.r += self.discount**self.steps * self.discriminator.discrim_net.reward_train(
-            state=torch.tensor(self.s).unsqueeze(0).to(self.discriminator.discrim_net.device()),
-            action=torch.tensor([[self.a]]).to(self.discriminator.discrim_net.device()),
-            next_state=torch.tensor(self.s).unsqueeze(0).to(self.discriminator.discrim_net.device()), # unused
-            done=torch.tensor(self.done).unsqueeze(0).to(self.discriminator.discrim_net.device()), # unused
-        )
-        self.steps += 1
-    
-    def _transitions(self):
-        """
-        Yield a single dictionary per high-level selected action
-        Fields:
-            obs: high-level state and mask at selection
-            action: chosen high-level action
-            reward: accumulated option reward
-            episode_start: whether the action was chosen at the episode start
-            value: the value estimate from the starting state
-            log_prob: the log_prob of the selected action from the starting state
-            done: whether the episode has ended
-
-        """
-        yield {
-            'obs': {'obs': self.s, 'mask': self.m},
-            'action': self.ch,
-            'reward': self.r.detach(),
-            'episode_start': self.episode_start,
-            'value': self.value.detach(),
-            'log_prob': self.log_prob.detach(),
-            'done': self.done,
-        }
-    
-    def sample_hl(self, policy, discriminator):
-        """
-        Args:
-            policy
-            discriminator: function with which to score rewards
-        Returns:
-            gen: iterable which samples high-level transitions from the environment
-        """
-        self.discriminator = discriminator
-        return self.sample(policy)
-
-class RenderOptions(LLOptions):
-
-    def _after_step(self):
-        """
-        Render the environment after each low-level step
-        """
-        super()._after_step()
-        self.env.render()
-    
-    def close(self, *args, **kwargs):
-        """
-        On 'close', close the environment
-        """
-        self.env.close(*args, **kwargs)
-
-def available_actions(env):
-    """Return mask of available actions given current `env` state."""
-    valid = np.array([feasible(env, generate_plan(env, i), i) for i in range(len(ALL_OPTIONS))])
-    return valid
-
-def target_velocity_plan(current_v: float, target_v: float, t: int, dt: float):
-    """Smoothly target a velocity in a given number of steps"""
-    # for now, constant acceleration
-    a = (target_v - current_v)  / (t * dt)
-    return a*np.ones((t,))
-
-def generate_plan(env, i):
-    """Generate input profile for high-level action `i`."""
-    assert i < len(ALL_OPTIONS), "Invalid option index {i}"
-    target_v, t = ALL_OPTIONS[i]
-    current_v = env._env.state[env._agent, 1].item() # extract from env
-    plan = target_velocity_plan(current_v, target_v, t, env._env._dt)
-    assert len(plan) == t, "incorrect plan length"
-    return plan
-
-def check_future_collisions_fast(env, actions):
-    """Checks whether `env._agent` would collide with other agents assuming `actions` as input.
-    
-    Vehicles are (over-)approximated by single circles.
-
-    Args:
-        env (gym.Env): current environment state
-        actions (list of torch.Tensor): list of B (T, nv, adims) T-length action profiles
-    Returns:
-        feasible (torch.Tensor): tensor of shape (B,) indicating whether the respective action profiles are collision-free
-    """
-    B, (T, nv, _) = len(actions), actions[0].shape
-
-    states = torch.stack(env._env.propagate_action_profile(actions), axis=0)
-    assert states.shape == (B, T, nv, 5)
-
-    distance = ((states[:, :, :, :2] - states[:, :, env._agent:env._agent+1, :2])**2).sum(-1).sqrt()
-    distance = torch.where(distance.isnan(), np.inf*torch.ones_like(distance), distance) # only collide with spawned agents
-    distance[:, :, env._agent] = np.inf # cannot collide with itself
-    assert distance.shape == (B, T, nv)
-
-    radius = (env._env._lengths**2 + env._env._widths**2).sqrt() / 2
-    min_distance = radius[env._agent] + radius
-    min_distance = min_distance.unsqueeze(0).unsqueeze(0)
-    assert min_distance.shape == (1, 1, nv)
-
-    return (distance > min_distance).all(-1).all(-1)
-
-def check_future_collisions_circles(env, actions, n_circles:int=2):
-    """Checks whether `env._agent` would collide with other agents assuming `actions` as input.
-    
-    Vehicles are (over-)approximated by multiple circles.
-
-    Args:
-        env (gym.Env): current environment state
-        actions (list of torch.Tensor): list of B (T, nv, adims) T-length action profiles
-    Returns:
-        feasible (torch.Tensor): tensor of shape (B,) indicating whether the respective action profiles are collision-free
-    """
-    assert n_circles >= 2
-    B, (T, nv, _) = len(actions), actions[0].shape
-
-    states = torch.stack(env._env.propagate_action_profile(actions), axis=0)
-    assert states.shape == (B, T, nv, 5)
-    centers = states[:, :, :, :2]
-    psi = states[:, :, :, 3]
-    lon = torch.stack([psi.cos(), psi.sin()],dim=-1) # (B, T, nv, 2)
-
-    # offset between [-env._env.lengths+env._env.widths/2, env._env.lengths/2-env._env.widths/2]
-    back = (-env._env._lengths/2+env._env._widths/2).unsqueeze(-1) # (nv, 1)
-    length = (env._env._lengths-env._env._widths).unsqueeze(-1) # (nv, 1)
-    diff_d =  back + length*(torch.arange(n_circles)/(n_circles-1)).unsqueeze(0)  # (nv, n_circles) 
-    assert diff_d.shape == (nv, n_circles)
-    
-    offsets = diff_d[None, None, :, :, None] *  lon[:, :, :, None, :]
-    assert offsets.shape == (B, T, nv, n_circles, 2)
-
-    expanded_centers=centers.unsqueeze(-2) + offsets #(B, T, nv, n_circles, 2)
-    assert expanded_centers.shape == (B, T, nv, n_circles, 2)
-    agent_centers = expanded_centers[:,:,env._agent:env._agent+1,:,:] #(B, T, 1, n_circles, 2)
-    ds = expanded_centers.reshape((B, T, nv*n_circles, 1, 2)) - agent_centers #(B, T, nv*nc,1, 2) - (B, T, 1, nc, 2) = (B, T, nv*nc, nc, 2) 
-
-    distance = (ds**2).sum(-1).sqrt().reshape((B, T, nv, n_circles, n_circles)) # (B, T, nv, nc, nc)
-    distance = torch.where(distance.isnan(), np.inf*torch.ones_like(distance), distance) # only collide with spawned agents
-    distance[:, :, env._agent] = np.inf # cannot collide with itself
-    assert distance.shape == (B, T, nv, n_circles, n_circles)
-
-    radius = env._env._widths*np.sqrt(2) / 2
-    min_distance = radius[env._agent] + radius
-    min_distance = min_distance[None, None, :, None, None]
-    assert min_distance.shape == (1, 1, nv, 1, 1)
-
-    return (distance > min_distance).all(-1).all(-1).all(-1).all(-1)
-
-def feasible(env, plan, ch):
-    """Check if input profile is feasible given current `env` state. Action `ch=0` is safe fallback."""
-    
-    # zero pad plan - Take (T,) np plan and convert it to (T, nv, 1) torch.Tensor
-    full_plan = torch.zeros(len(plan), env._env._nv, 1)
-    full_plan[:, env._agent, 0] = torch.tensor(plan)
-    # valid = check_future_collisions_fast(env, [full_plan]) # check_future_collisions_fast takes in B-list and outputs (B,) bool tensor
-    valid = check_future_collisions_circles(env, [full_plan])
-    return ch == 0 or valid.item()
-
-def flatten_transitions(transitions):
-    return {
-        'obs': np.stack(list(t['obs'] for t in transitions), axis=0),
-        'next_obs': np.stack(list(t['next_obs'] for t in transitions), axis=0),
-        'acts': np.stack(list(t['acts'] for t in transitions), axis=0),
-        'dones': np.stack(list(t['dones'] for t in transitions), axis=0),
-    }
-
-def train_discriminator(env, generator, discriminator, num_samples):
-    transitions = list(itertools.islice(env.sample_ll(generator), num_samples))
-    generator_samples = flatten_transitions(transitions)
-    discriminator.train_disc(gen_samples=generator_samples)
-
-def train_generator(env, generator, discriminator, num_samples):
-    generator_samples = list(itertools.islice(env.sample_hl(generator, discriminator), num_samples+1))
-    
-    generator.rollout_buffer.reset()
-    for s in generator_samples[:-1]:
-        generator.rollout_buffer.add(
-            obs=s['obs'],
-            action=s['action'].cpu(),
-            reward=s['reward'].cpu(),
-            episode_start=s['episode_start'],
-            value=s['value'],
-            log_prob=s['log_prob'],
-        )
-    
-    generator.rollout_buffer.compute_returns_and_advantage(
-        last_values=generator_samples[-1]['value'],
-        dones=generator_samples[-1]['done'],
-    )
-    
-    generator.train()
-
-def train(expert_data, env_class=NRasterizedRandomAgent, env_settings={}, epochs=10, discrim_batch_size=32, generator_steps=2048, discount=0.99):
-    """
-    Args: 
-        expert_data: list of transitions
-        env_class: environment class
-        env_settings: environment settings
-        epochs: number of epochs to train for
-        discrim_batch_size: discriminator batch size
-        generator_steps: number of steps taken in generator
-        discount: discount factor
-    Returns:
-        generator (stable_baselines3.PPO): options policy
-    """
-    env = env_class(**env_settings)
-    env.discount = discount
-
-    tempdir = tempfile.TemporaryDirectory(prefix="quickstart")
-    tempdir_path = pathlib.Path(tempdir.name)
-    logger.configure(tempdir_path / "GAIL/")
-    print(f"All Tensorboards and logging are being written inside {tempdir_path}/.")
-
-    venv = make_vec_env(env_class, n_envs=1, env_kwargs=env_settings)
-    discriminator = adversarial.GAIL(
-        expert_data=expert_data,
-        expert_batch_size=discrim_batch_size,
-        discrim_kwargs={'discrim_net': CnnDiscriminatorFlatAction(venv)},
-        #discrim_kwargs={'discrim_net': CnnDiscriminator(venv)},
-        venv=venv, # unused
-        gen_algo=stable_baselines3.PPO("CnnPolicy", venv), # unused
-    )
-
-    generator = stable_baselines3.PPO(
-        OptionsCnnPolicy,
-        OptionsEnv(env),
-        verbose=1,
-        n_steps=generator_steps,
-    )
-
-    # PPO.train requires logger as set up in
-    # PPO._setup_learn (called by PPO.learn)
-    generator._logger = stable_baselines3.common.utils.configure_logger(
-        generator.verbose,
-        generator.tensorboard_log,
-    )
-
-    for _ in tqdm(range(epochs)):
-        train_discriminator(LLOptions(env), generator, discriminator, num_samples=discrim_batch_size)
-        train_generator(HLOptions(env), generator, discriminator, num_samples=generator_steps)
-    
-    return generator
-
-# %%
-if __name__ == '__main__':
-    # %%
-    model_name = 'gail_options_image'
-    env_class = NRasterizedRandomAgent
-    env_settings = {'width': 36, 'height': 36, 'm_per_px': 2}
-
-    with open("data/NormalizedIntersimpleExpertMu.001_NRasterizedIncrementingAgentw36h36mppx2.pkl", "rb") as f:
-        trajectories = pickle.load(f)
-    #import pdb 
-    #pdb.set_trace()
-    transitions = rollout.flatten_trajectories(trajectories)
-    generator = train(
-        transitions,
-        env_class=env_class,
-        env_settings=env_settings,
-        epochs=2, 
-        discrim_batch_size=32, 
-        generator_steps=2048, 
-        discount=0.99
-        )
-
-    generator.save(model_name) # save ppo sb3 generator class
-
-    # %%
-    model = stable_baselines3.PPO.load(model_name) # not actually used
-
-    env = RenderOptions(NRasterizedRandomAgent(**env_settings))
-    for s in env.sample_ll(generator):
-        if s['dones']:
-            break
-
-    env.close(filestr='render/'+model_name)
-
-# %% Tests
-
-def test_ll_expert_data():
-    with open("data/NormalizedIntersimpleExpertMu.001_NRasterizedAgent51w36h36mppx2.pkl", "rb") as f:
-        expert_trajectories = pickle.load(f)
-    expert_transitions = rollout.flatten_trajectories(expert_trajectories)
-
-    env = LLOptions(NRasterized(agent=51, width=36, height=36, m_per_px=2))
-
-    gen_transitions = list(itertools.islice(env.sample_ll(
-        policy=stable_baselines3.PPO(
-            OptionsCnnPolicy,
-            OptionsEnv(env),
-            verbose=1,
-        )
-    ), 10))
-    gen_transitions = flatten_transitions(gen_transitions)
-
-    assert expert_transitions[:10].obs.shape == gen_transitions['obs'].shape
-    assert expert_transitions[:10].next_obs.shape == gen_transitions['next_obs'].shape
-    assert expert_transitions[:10].acts.shape == gen_transitions['acts'].shape
-    assert expert_transitions[:10].dones.shape == gen_transitions['dones'].shape
-
-def test_ll_states():
-    env = NRasterized()
-    policy = stable_baselines3.PPO(
-        OptionsCnnPolicy,
-        OptionsEnv(env),
-        verbose=1,
-    )
-    llenv = LLOptions(env)
-    transitions = list(itertools.islice(llenv.sample_ll(policy=policy), 100))
-
-    env2 = NRasterized()
-    s2 = env2.reset()
-    for i, t in enumerate(transitions):
-        assert i == 0 or np.array_equal(t['obs'], transitions[i-1]['next_obs'])
-        assert np.array_equal(t['obs'], s2)
-        assert t['acts'].shape == (1,)
-
-        nexts2, _, done2, _ = env2.step(t['acts'])
-        assert np.array_equal(t['next_obs'], nexts2)
-        assert np.array_equal(t['dones'], done2)
-
-        if done2:
-            break
-
-        s2 = nexts2
-
-def test_hl_transitions():
-    pass

scratch/arec/intersimple/options_gail.py

@@ -1,510 +0,0 @@
-# %%
-from gail.discriminator import CnnDiscriminator, CnnDiscriminatorFlatAction
-from imitation.algorithms import adversarial
-import stable_baselines3
-import torch.utils.data
-import numpy as np
-from intersim.envs.intersimple import NRasterized, NRasterizedRandomAgent
-import itertools
-from torch.distributions import Categorical
-import gym
-import torch
-import pickle
-import imitation.data.rollout as rollout
-import tempfile
-import pathlib
-from imitation.util import logger
-from stable_baselines3.common.env_util import make_vec_env
-from tqdm import tqdm
-
-import logging 
-logging.basicConfig(level=logging.DEBUG)
-
-ALL_OPTIONS = [(v,t) for v in [0,2,4,6,8] for t in [5, 10, 20]] # option 0 is safe fallback
-
-class OptionsCnnPolicy(stable_baselines3.common.policies.ActorCriticCnnPolicy):
-    """
-    Class for high-level options policy (generator)
-    """
-    def __init__(self, observation_space, *args, **kwargs):
-        super().__init__(observation_space['obs'], *args, **kwargs)
-
-    def _prior_distribution(self, s):
-        """
-        Return prior distribution over high-level options (before masking)
-        Args:
-            s (torch.tensor): observation
-        Returns:
-            values (torch.tensor): values from critic
-            dist (torch.distributions): prior distribution over actions
-        """
-        latent_pi, latent_vf, latent_sde = self._get_latent(s) 
-        distribution = self._get_action_dist_from_latent(latent_pi, latent_sde) 
-        values = self.value_net(latent_vf)
-        return values, distribution.distribution
-
-    def predict(self, obs):
-        """
-        Will mask invalid states before making action selections
-        Args:
-            obs: dict with keys:
-                obs (torch.tensor): (B,o)  true observations
-                mask (torch.tensor): (B,m) mask over valid actions
-        Returns:
-            ch (torch.tensor): (B,a) sampled actions
-            values (torch.tensor): (B,) predicted value at observation
-            log_probs (torch.tensor): (B,) log probabilities of selected actions
-        """
-        s, m = obs['obs'], obs['mask']
-        values, prior = self._prior_distribution(s)
-        posterior = Categorical(prior.probs * m)
-        ch = posterior.sample()
-        return ch, values, posterior.log_prob(ch)
-
-    def evaluate_actions(self, obs, ch):
-        """
-        Evaluate particular actions
-        Args:
-            obs: dict with keys:
-                obs (torch.tensor): (B,o) true observations
-                mask (torch.tensor): (B,m) masks over valid actions
-            ch (torch.tensor): (B,a) selected actions
-        Returns:
-            values (torch.tensor): (B,) predicted value at observation
-            log_probs (torch.tensor): (B,) log probabilities of selected actions
-            ent (torch.tensor): (B,) entropy of each distribution over actions
-        """
-        s, m = obs['obs'], obs['mask']
-        values, prior = self._prior_distribution(s)
-        posterior = Categorical(prior.probs * m)
-        return values, posterior.log_prob(ch), posterior.entropy() # additional values used by PPO.train
-
-class OptionsEnv(gym.Wrapper):
-    """
-    Wrap an intersimple environment with an options generator
-    """
-    def __init__(self, env, render=False, *args, **kwargs):
-        """
-        Initialize wrapped environment and set high-level action and observation spaces
-        """
-        super().__init__(env, *args, **kwargs)
-        num_hl_options = len(ALL_OPTIONS)
-        self.action_space = gym.spaces.Discrete(num_hl_options)
-        self.observation_space = gym.spaces.Dict({
-            'obs': env.observation_space,
-            'mask': gym.spaces.Box(low=0, high=1, shape=(num_hl_options,)),
-        })
-        self._hl_transition_buffer = []
-        self._ll_transition_buffer = []
-        self.render=render
-
-    def _after_option_choice(self):
-        """
-        After initial option choice, 
-        """
-        self._hl_r = 0
-        self._hl_steps = 0
-    
-    def _after_step(self):
-        """
-        After each step, add the ll transition to the appropriate buffer, add to reward, add to steps, and possibly render
-        """
-
-        self._ll_transition_buffer.append({
-            'obs': self.s,
-            'next_obs': self.nexts,
-            'acts': np.array((self.a,)),
-            'dones': np.array(self.done),
-        })  
-        self.r += self.discount**self.steps * self.discriminator.discrim_net.reward_train(
-            state=torch.tensor(self.s).unsqueeze(0).to(self.discriminator.discrim_net.device()),
-            action=torch.tensor([[self.a]]).to(self.discriminator.discrim_net.device()),
-            next_state=torch.tensor(self.s).unsqueeze(0).to(self.discriminator.discrim_net.device()), # unused
-            done=torch.tensor(self.done).unsqueeze(0).to(self.discriminator.discrim_net.device()), # unused
-        )
-        self.steps += 1
-        if self.render:
-            self.env.render()
-    
-    def _after_option(self):
-        """
-        After each low-level action, add the discounted discriminated reward score (given a discriminator)
-        """
-        self._hl_transition_buffer.append({
-            'obs': {'obs': self.os, 'mask': self.m},
-            'action': self.ch,
-            'reward': self.r.detach(),
-            'episode_start': self.episode_start,
-            'value': self.value.detach(),
-            'log_prob': self.log_prob.detach(),
-            'done': self.done,
-            })
-
-    def close(self, *args, **kwargs):
-        """
-        On 'close', close the environment
-        """
-        self.env.close(*args, **kwargs)
-    
-    def sample(self, generator, controller):
-        """
-        yield transitions using a generator
-        Args:
-            generator (sb3.PPO)
-            controller (str): 'high' or 'low' to yield from proper buffer
-        Yields:
-
-        """
-        self.done = True
-        # DO I WANT TO EMPTY THE BUFFERS??? Probs naw
-        while True:
-            
-            # yield from buffers to empty what was stored previously
-            if controller = 'high':
-                yield from self._hl_transition_buffer
-            elif controller == 'low':
-                yield from self._ll_transition_buffer
-            else:
-                raise('Improper buffer')
-            
-            self.episode_start = False
-            if self.done:
-                # reset environment
-                self.s = self.env.reset()
-                self.done = False
-                self.episode_start = True
-
-            self.os = self.s.copy() # option start state
-            self.m = available_actions(self.env)
-
-            # set the action, the value of the start state, and the logprob of the action
-            # according to the current environment state and mask
-            self.ch, self.value, self.log_prob = generator.policy.predict({
-                'obs': torch.tensor(self.os).unsqueeze(0).to(generator.policy.device),
-                'mask': torch.tensor(self.m).unsqueeze(0).to(generator.policy.device),
-            })
-
-            # store a float list of actions to take given the option selected in the environment
-            self.plan = list(map(float, generate_plan(self.env, self.ch)))
-
-            # run whatever _after_choice might dictate in a child class
-            self._after_option_choice()
-
-            # some checks
-            assert not self.done
-            assert self.plan
-            assert feasible(self.env, self.plan, self.ch)
-
-            # execute the option so long as the episode isn't complete and the plan is still feasible
-            while not self.done and self.plan and feasible(self.env, self.plan, self.ch):
-                
-                # pop first action
-                self.a, self.plan = self.plan[0], self.plan[1:]
-                
-                # normalize action ??
-                self.a = self.env._normalize(self.a)
-                
-                # step through environment
-                self.nexts, _, self.done, _ = self.env.step(self.a)
-
-                # run whatever _after_step might dictate in child class
-                self._after_step()
-
-                # update state and mask to current
-                self.s = self.nexts
-            
-            # run whatever to do after option
-            self._after_option()
-
-    def sample_ll(self, policy):
-        """
-        Not quite sure how this works????
-        Why would you do this over LLOptions.sample(policy)
-        """
-        return self.sample(policy, 'low')
-
-    def sample_hl(self, policy, discriminator):
-        """
-        Args:
-            policy
-            discriminator: function with which to score rewards
-        Returns:
-            gen: an which samples high-level transitions from the environment
-        """
-        self.discriminator = discriminator
-        return self.sample(policy)
-
-def available_actions(env):
-    """Return mask of available actions given current `env` state."""
-    valid = np.array([feasible(env, generate_plan(env, i), i) for i in range(len(ALL_OPTIONS))])
-    return valid
-
-def target_velocity_plan(current_v: float, target_v: float, t: int, dt: float):
-    """Smoothly target a velocity in a given number of steps"""
-    # for now, constant acceleration
-    a = (target_v - current_v)  / (t * dt)
-    return a*np.ones((t,))
-
-def generate_plan(env, i):
-    """Generate input profile for high-level action `i`."""
-    assert i < len(ALL_OPTIONS), "Invalid option index {i}"
-    target_v, t = ALL_OPTIONS[i]
-    current_v = env._env.state[env._agent, 1].item() # extract from env
-    plan = target_velocity_plan(current_v, target_v, t, env._env._dt)
-    assert len(plan) == t, "incorrect plan length"
-    return plan
-
-def check_future_collisions_fast(env, actions):
-    """Checks whether `env._agent` would collide with other agents assuming `actions` as input.
-    
-    Vehicles are (over-)approximated by single circles.
-
-    Args:
-        env (gym.Env): current environment state
-        actions (list of torch.Tensor): list of B (T, nv, adims) T-length action profiles
-    Returns:
-        feasible (torch.Tensor): tensor of shape (B,) indicating whether the respective action profiles are collision-free
-    """
-    B, (T, nv, _) = len(actions), actions[0].shape
-
-    states = torch.stack(env._env.propagate_action_profile(actions), axis=0)
-    assert states.shape == (B, T, nv, 5)
-
-    distance = ((states[:, :, :, :2] - states[:, :, env._agent:env._agent+1, :2])**2).sum(-1).sqrt()
-    distance = torch.where(distance.isnan(), np.inf*torch.ones_like(distance), distance) # only collide with spawned agents
-    distance[:, :, env._agent] = np.inf # cannot collide with itself
-    assert distance.shape == (B, T, nv)
-
-    radius = (env._env._lengths**2 + env._env._widths**2).sqrt() / 2
-    min_distance = radius[env._agent] + radius
-    min_distance = min_distance.unsqueeze(0).unsqueeze(0)
-    assert min_distance.shape == (1, 1, nv)
-
-    return (distance > min_distance).all(-1).all(-1)
-
-def check_future_collisions_circles(env, actions, n_circles:int=2):
-    """Checks whether `env._agent` would collide with other agents assuming `actions` as input.
-    
-    Vehicles are (over-)approximated by multiple circles.
-
-    Args:
-        env (gym.Env): current environment state
-        actions (list of torch.Tensor): list of B (T, nv, adims) T-length action profiles
-    Returns:
-        feasible (torch.Tensor): tensor of shape (B,) indicating whether the respective action profiles are collision-free
-    """
-    assert n_circles >= 2
-    B, (T, nv, _) = len(actions), actions[0].shape
-
-    states = torch.stack(env._env.propagate_action_profile(actions), axis=0)
-    assert states.shape == (B, T, nv, 5)
-    centers = states[:, :, :, :2]
-    psi = states[:, :, :, 3]
-    lon = torch.stack([psi.cos(), psi.sin()],dim=-1) # (B, T, nv, 2)
-
-    # offset between [-env._env.lengths+env._env.widths/2, env._env.lengths/2-env._env.widths/2]
-    back = (-env._env._lengths/2+env._env._widths/2).unsqueeze(-1) # (nv, 1)
-    length = (env._env._lengths-env._env._widths).unsqueeze(-1) # (nv, 1)
-    diff_d =  back + length*(torch.arange(n_circles)/(n_circles-1)).unsqueeze(0)  # (nv, n_circles) 
-    assert diff_d.shape == (nv, n_circles)
-    
-    offsets = diff_d[None, None, :, :, None] *  lon[:, :, :, None, :]
-    assert offsets.shape == (B, T, nv, n_circles, 2)
-
-    expanded_centers=centers.unsqueeze(-2) + offsets #(B, T, nv, n_circles, 2)
-    assert expanded_centers.shape == (B, T, nv, n_circles, 2)
-    agent_centers = expanded_centers[:,:,env._agent:env._agent+1,:,:] #(B, T, 1, n_circles, 2)
-    ds = expanded_centers.reshape((B, T, nv*n_circles, 1, 2)) - agent_centers #(B, T, nv*nc,1, 2) - (B, T, 1, nc, 2) = (B, T, nv*nc, nc, 2) 
-
-    distance = (ds**2).sum(-1).sqrt().reshape((B, T, nv, n_circles, n_circles)) # (B, T, nv, nc, nc)
-    distance = torch.where(distance.isnan(), np.inf*torch.ones_like(distance), distance) # only collide with spawned agents
-    distance[:, :, env._agent] = np.inf # cannot collide with itself
-    assert distance.shape == (B, T, nv, n_circles, n_circles)
-
-    radius = env._env._widths*np.sqrt(2) / 2
-    min_distance = radius[env._agent] + radius
-    min_distance = min_distance[None, None, :, None, None]
-    assert min_distance.shape == (1, 1, nv, 1, 1)
-
-    return (distance > min_distance).all(-1).all(-1).all(-1).all(-1)
-
-def feasible(env, plan, ch):
-    """Check if input profile is feasible given current `env` state. Action `ch=0` is safe fallback."""
-    
-    # zero pad plan - Take (T,) np plan and convert it to (T, nv, 1) torch.Tensor
-    full_plan = torch.zeros(len(plan), env._env._nv, 1)
-    full_plan[:, env._agent, 0] = torch.tensor(plan)
-    # valid = check_future_collisions_fast(env, [full_plan]) # check_future_collisions_fast takes in B-list and outputs (B,) bool tensor
-    valid = check_future_collisions_circles(env, [full_plan])
-    return ch == 0 or valid.item()
-
-def flatten_transitions(transitions):
-    return {
-        'obs': np.stack(list(t['obs'] for t in transitions), axis=0),
-        'next_obs': np.stack(list(t['next_obs'] for t in transitions), axis=0),
-        'acts': np.stack(list(t['acts'] for t in transitions), axis=0),
-        'dones': np.stack(list(t['dones'] for t in transitions), axis=0),
-    }
-
-def train_discriminator(env, generator, discriminator, num_samples):
-    transitions = list(itertools.islice(env.sample_ll(generator), num_samples))
-    generator_samples = flatten_transitions(transitions)
-    discriminator.train_disc(gen_samples=generator_samples)
-
-def train_generator(env, generator, discriminator, num_samples):
-    generator_samples = list(itertools.islice(env.sample_hl(generator, discriminator), num_samples+1))
-    
-    generator.rollout_buffer.reset()
-    for s in generator_samples[:-1]:
-        generator.rollout_buffer.add(
-            obs=s['obs'],
-            action=s['action'].cpu(),
-            reward=s['reward'].cpu(),
-            episode_start=s['episode_start'],
-            value=s['value'],
-            log_prob=s['log_prob'],
-        )
-    
-    generator.rollout_buffer.compute_returns_and_advantage(
-        last_values=generator_samples[-1]['value'],
-        dones=generator_samples[-1]['done'],
-    )
-    
-    generator.train()
-
-def train(expert_data, env_class=NRasterizedRandomAgent, env_settings={}, epochs=10, discrim_batch_size=32, generator_steps=2048, discount=0.99):
-    """
-    Args: 
-        expert_data: list of transitions
-        env_class: environment class
-        env_settings: environment settings
-        epochs: number of epochs to train for
-        discrim_batch_size: discriminator batch size
-        generator_steps: number of steps taken in generator
-        discount: discount factor
-    Returns:
-        generator (stable_baselines3.PPO): options policy
-    """
-    env = env_class(**env_settings)
-    env.discount = discount
-
-    tempdir = tempfile.TemporaryDirectory(prefix="quickstart")
-    tempdir_path = pathlib.Path(tempdir.name)
-    logger.configure(tempdir_path / "GAIL/")
-    print(f"All Tensorboards and logging are being written inside {tempdir_path}/.")
-
-    venv = make_vec_env(env_class, n_envs=1, env_kwargs=env_settings)
-    discriminator = adversarial.GAIL(
-        expert_data=expert_data,
-        expert_batch_size=discrim_batch_size,
-        discrim_kwargs={'discrim_net': CnnDiscriminatorFlatAction(venv)},
-        #discrim_kwargs={'discrim_net': CnnDiscriminator(venv)},
-        venv=venv, # unused
-        gen_algo=stable_baselines3.PPO("CnnPolicy", venv), # unused
-    )
-
-    generator = stable_baselines3.PPO(
-        OptionsCnnPolicy,
-        OptionsEnv(env),
-        verbose=1,
-        n_steps=generator_steps,
-    )
-
-    # PPO.train requires logger as set up in
-    # PPO._setup_learn (called by PPO.learn)
-    generator._logger = stable_baselines3.common.utils.configure_logger(
-        generator.verbose,
-        generator.tensorboard_log,
-    )
-
-    for _ in tqdm(range(epochs)):
-        train_discriminator(LLOptions(env), generator, discriminator, num_samples=discrim_batch_size)
-        train_generator(HLOptions(env), generator, discriminator, num_samples=generator_steps)
-    
-    return generator
-
-# %%
-if __name__ == '__main__':
-    # %%
-    model_name = 'gail_options_image'
-    env_class = NRasterizedRandomAgent
-    env_settings = {'width': 36, 'height': 36, 'm_per_px': 2}
-
-    with open("data/NormalizedIntersimpleExpertMu.001_NRasterizedIncrementingAgentw36h36mppx2.pkl", "rb") as f:
-        trajectories = pickle.load(f)
-    #import pdb 
-    #pdb.set_trace()
-    transitions = rollout.flatten_trajectories(trajectories)
-    generator = train(
-        transitions,
-        env_class=env_class,
-        env_settings=env_settings,
-        epochs=2, 
-        discrim_batch_size=32, 
-        generator_steps=2048, 
-        discount=0.99
-        )
-
-    generator.save(model_name) # save ppo sb3 generator class
-
-    # %%
-    model = stable_baselines3.PPO.load(model_name) # not actually used
-
-    env = OptionsGail(NRasterizedRandomAgent(**env_settings), render=True)
-    for s in env.sample_ll(generator):
-        if s['dones']:
-            break
-
-    env.close(filestr='render/'+model_name)
-
-# %% Tests
-
-def test_ll_expert_data():
-    with open("data/NormalizedIntersimpleExpertMu.001_NRasterizedAgent51w36h36mppx2.pkl", "rb") as f:
-        expert_trajectories = pickle.load(f)
-    expert_transitions = rollout.flatten_trajectories(expert_trajectories)
-
-    env = LLOptions(NRasterized(agent=51, width=36, height=36, m_per_px=2))
-
-    gen_transitions = list(itertools.islice(env.sample_ll(
-        policy=stable_baselines3.PPO(
-            OptionsCnnPolicy,
-            OptionsEnv(env),
-            verbose=1,
-        )
-    ), 10))
-    gen_transitions = flatten_transitions(gen_transitions)
-
-    assert expert_transitions[:10].obs.shape == gen_transitions['obs'].shape
-    assert expert_transitions[:10].next_obs.shape == gen_transitions['next_obs'].shape
-    assert expert_transitions[:10].acts.shape == gen_transitions['acts'].shape
-    assert expert_transitions[:10].dones.shape == gen_transitions['dones'].shape
-
-def test_ll_states():
-    env = NRasterized()
-    policy = stable_baselines3.PPO(
-        OptionsCnnPolicy,
-        OptionsEnv(env),
-        verbose=1,
-    )
-    llenv = LLOptions(env)
-    transitions = list(itertools.islice(llenv.sample_ll(policy=policy), 100))
-
-    env2 = NRasterized()
-    s2 = env2.reset()
-    for i, t in enumerate(transitions):
-        assert i == 0 or np.array_equal(t['obs'], transitions[i-1]['next_obs'])
-        assert np.array_equal(t['obs'], s2)
-        assert t['acts'].shape == (1,)
-
-        nexts2, _, done2, _ = env2.step(t['acts'])
-        assert np.array_equal(t['next_obs'], nexts2)
-        assert np.array_equal(t['dones'], done2)
-
-        if done2:
-            break
-
-        s2 = nexts2
-
-def test_hl_transitions():
-    pass

scratch/arec/intersimple/plan.txt

@@ -1,55 +0,0 @@
-Environment 
--- each 'environment' follows a single roundabout and track id (recording of that roundabout)
--- on reset, the environment we will use changes the vehicle to control while having the other agents follow their true data (expert controller)
----- Note this can be problematic as it can lead to vehicles behind you crashing into you
-
-TRAINING
----------
-1. Load pre-trained massive set of transitions
--- For all roundabouts
-    -- For all tracks
-        -- For all vehicles
-            -- For all valid timesteps
-                -- Rasterized state (incl. path), action
-
-2. HGAIL
--- For each epoch
-    -- INSTANTIATE A NEW ENVIRONMENT (Roundabout + Track) w/ randomized agent, from set of all expert environments
-        -- Train discriminator off training data + yielded low-level transitions in replay buffer
-        -- Train generator off yielded high-level transitions + summed low-level discriminator rewards
-
-TESTING
-----------
-1. Save average vehicle velocities for all expert vehicles (loop roundabout + track + vehicle, average over time)
-
-2. Run test suite for: expert, BC, GAIL, RAIL, HGAIL, (and hopefully HRAIL)
--- For all roundabouts, tracks
-    -- Get expert velocities for track
-    -- Simulate incrementing agent environment (e.g. on reset, agent +=1)
-        -- Store low-level true joint states, actions, and controlled vehicle index
-    -- Per-vehicle statistics (v_all, v_mean, v_shortfall, a_all, jerk_all, n_collisions, T) 
--- Aggregate statistics + joint 
-
-Problems
------------
-Should train without stopping for collisions, however when doing so, end up with policy that always takes decelerate option
--- It seems safe at the start of each vehicles sim, but actually it isn't since a car will spawn and hit it
-Solutions:
-    -- Hold cars from spawning if their spawn location is full
-    -- Start simulations a few seconds later (after cars clear their spawn places) <- Preferred
-    
-Test could run indefinitely if stop_on_collision is off
-Solution:
-    -- Set maximum episode length in intersimple
-
-
-
-
-
-
-
-
-Save massive set of transition raw states beforehand (1 from training, but with raw states)
-# -- For all roundabouts, tracks
-#     -- For all vehicles, steps
-#         -- Raw vehicle state, action

scratch/arec/intersimple/render_env_from_model.py

@@ -1,58 +0,0 @@
-
-import stable_baselines3 as sb3
-from intersim.envs.intersimple import NRasterized
-
-
-def render_env(model_name='gail_image_multiagent_nocollision', agent=51, environment=NRasterized):
-    """
-    Render a video from an model, agent, and environment
-    Args:
-        model_name (str): name of the model
-        agent (int): agent to start the video from
-        environment (gym.Env): gym environment class to render environment on
-    """
-
-    model = sb3.PPO.load(model_name)
-
-    env = environment(stop_on_collision=False, width=36, height=36, m_per_px=2, agent=agent)
-
-    obs = env.reset()
-    i=0
-    while True and i < 600:
-        i+=1
-        action, _states = model.predict(obs)
-        obs, rewards, done, info = env.step(action)
-        env.render(mode='post')
-        if done:
-            break
-
-    env.close(filestr='render/'+model_name+'_agent%i'%(agent)) 
-
-def render_options_env(model_name='gail_image_multiagent_nocollision', agent=51, environment=NRasterized):
-    """
-    Render a video from an model, agent, and environment
-    Args:
-        model_name (str): name of the model
-        agent (int): agent to start the video from
-        environment (gym.Env): gym environment class to render environment on
-    """
-
-    model = sb3.PPO.load(model_name)
-
-    env = environment(stop_on_collision=False, width=36, height=36, m_per_px=2, agent=agent)
-
-    obs = env.reset()
-    i=0
-    while True and i < 600:
-        i+=1
-        action, _states = model.predict(obs)
-        obs, rewards, done, info = env.step(action)
-        env.render(mode='post')
-        if done:
-            break
-
-    env.close(filestr='render/'+model_name+'_agent%i'%(agent)) 
-
-if __name__ == '__main__':
-    import fire
-    fire.Fire(render_env)

scratch/arec/intersimple/render_options.py

@@ -1,11 +0,0 @@
-import sys
-sys.path.append('../../../')
-from src.util import render_env
-ALL_OPTIONS = [(v,t) for v in [0,2,4,6,8] for t in [5, 10]]
-
-def render_wrapper(**kwargs):
-    render_env(**kwargs, options_list=ALL_OPTIONS)
-
-if __name__=='__main__':
-    import fire
-    fire.Fire(render_wrapper)

scratch/etienne/intersimple/airl_flat.py

@@ -1,64 +0,0 @@
-# %%
-import pathlib
-import pickle
-import tempfile
-
-import stable_baselines3 as sb3
-from stable_baselines3.common.env_util import make_vec_env
-
-from imitation.algorithms import adversarial, bc
-from imitation.data import rollout
-from imitation.util import logger
-
-from intersim.envs.intersimple import IntersimpleReward
-
-model_name = 'airl_flat'
-
-# Load pickled test demonstrations.
-with open("data/NormalizedIntersimpleExpert_IntersimpleRewardAgent51.pkl", "rb") as f:
-    # This is a list of `imitation.data.types.Trajectory`, where
-    # every instance contains observations and actions for a single expert
-    # demonstration.
-    trajectories = pickle.load(f)
-
-# %%
-# Convert List[types.Trajectory] to an instance of `imitation.data.types.Transitions`.
-# This is a more general dataclass containing unordered
-# (observation, actions, next_observation) transitions.
-transitions = rollout.flatten_trajectories(trajectories)
-
-venv = make_vec_env(IntersimpleReward, n_envs=2, env_kwargs={'agent': 51})
-
-tempdir = tempfile.TemporaryDirectory(prefix="quickstart")
-tempdir_path = pathlib.Path(tempdir.name)
-print(f"All Tensorboards and logging are being written inside {tempdir_path}/.")
-
-# Train AIRL on expert data.
-# GAIL, and AIRL also accept as `expert_data` any Pytorch-style DataLoader that
-# iterates over dictionaries containing observations, actions, and next_observations.
-logger.configure(tempdir_path / "AIRL/")
-airl_trainer = adversarial.AIRL(
-    venv,
-    expert_data=transitions,
-    expert_batch_size=64,
-    gen_algo=sb3.PPO("MlpPolicy", venv, verbose=1, n_steps=1024), # n_steps = 2048 ?
-)
-airl_trainer.train(total_timesteps=100000)
-airl_trainer.gen_algo.save(model_name)
-
-del airl_trainer
-
-# %%
-model = sb3.PPO.load(model_name)
-
-env = IntersimpleReward(agent=51)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)

scratch/etienne/intersimple/bc_flat.py

@@ -1,59 +0,0 @@
-# %%
-import pathlib
-import pickle
-import tempfile
-
-import stable_baselines3 as sb3
-from stable_baselines3.common.env_util import make_vec_env
-
-from imitation.algorithms import adversarial, bc
-from imitation.data import rollout
-from imitation.util import logger
-
-from intersim.envs.intersimple import IntersimpleReward
-
-model_name = 'bc_flat'
-
-# Load pickled test demonstrations.
-with open("data/NormalizedIntersimpleExpert_IntersimpleRewardAgent51.pkl", "rb") as f:
-    # This is a list of `imitation.data.types.Trajectory`, where
-    # every instance contains observations and actions for a single expert
-    # demonstration.
-    trajectories = pickle.load(f)
-
-# %%
-# Convert List[types.Trajectory] to an instance of `imitation.data.types.Transitions`.
-# This is a more general dataclass containing unordered
-# (observation, actions, next_observation) transitions.
-transitions = rollout.flatten_trajectories(trajectories)
-
-venv = make_vec_env(IntersimpleReward, n_envs=2, env_kwargs={'agent': 51})
-
-tempdir = tempfile.TemporaryDirectory(prefix="quickstart")
-tempdir_path = pathlib.Path(tempdir.name)
-print(f"All Tensorboards and logging are being written inside {tempdir_path}/.")
-
-# Train BC on expert data.
-# BC also accepts as `expert_data` any PyTorch-style DataLoader that iterates over
-# dictionaries containing observations and actions.
-logger.configure(tempdir_path / "BC/")
-bc_trainer = bc.BC(venv.observation_space, venv.action_space, expert_data=transitions)
-bc_trainer.train(n_epochs=1000)
-bc_trainer.save_policy(model_name)
-
-del bc_trainer
-
-# %%
-model = bc.reconstruct_policy(model_name)
-
-env = IntersimpleReward(agent=51)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)

scratch/etienne/intersimple/data/expert.py

@@ -1,138 +0,0 @@
-from intersim.envs.intersimple import Intersimple, InfoFilter
-from stable_baselines3.common.policies import BasePolicy
-import gym
-from intersim.envs.intersimple import *
-from gail.envs import *
-import imitation.data.rollout as rollout
-from stable_baselines3.common.vec_env.dummy_vec_env import DummyVecEnv
-from imitation.data.wrappers import RolloutInfoWrapper
-
-class IntersimExpert(BasePolicy):
-    
-    def __init__(self, intersim_env, mu=0, *args, **kwargs):
-        super().__init__(
-            observation_space=gym.spaces.Space(),
-            action_space=gym.spaces.Space(),
-            *args, **kwargs
-        )
-        self._intersim = intersim_env
-        self._mu = mu
-
-    def forward(self, *args, **kwargs):
-        raise NotImplementedError()
-
-    def _predict(self, *args, **kwargs):
-        raise NotImplementedError()
-
-    def _action(self):
-        target_t = min(self._intersim._ind + 1, len(self._intersim._svt.simstate) - 1)
-        target_state = self._intersim._svt.simstate[target_t]
-        return self._intersim.target_state(target_state, mu=self._mu)
-
-    def predict(self, *args, **kwargs):
-        return self._action(), None
-
-class IntersimpleExpert(BasePolicy):
-
-    def __init__(self, intersimple_env, mu=0, *args, **kwargs):
-        super().__init__(
-            observation_space=intersimple_env.observation_space,
-            action_space=intersimple_env.action_space,
-            *args, **kwargs
-        )
-        self._intersimple = intersimple_env
-        self._intersim_expert = IntersimExpert(intersimple_env._env, mu=mu)
-    
-    def forward(self, *args, **kwargs):
-        raise NotImplementedError()
-
-    def _predict(self, *args, **kwargs):
-        raise NotImplementedError()
-
-    def _action(self):
-        # RandomLocation mixin re-initializes the intersim sub-env
-        self._intersim_expert._intersim = self._intersimple._env
-        return self._intersim_expert._action()[self._intersimple._agent]
-
-    def predict(self, *args, **kwargs):
-        return self._action(), None
-
-class NormalizedIntersimpleExpert(IntersimpleExpert):
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-    def predict(self, *args, **kwargs):
-        action, _ = super().predict(*args, **kwargs)
-        return self._intersimple._normalize(action), None
-
-class DummyVecEnvPolicy(BasePolicy):
-
-    def __init__(self, experts):
-        self._experts = [e() for e in experts]
-
-    def forward(self, *args, **kwargs):
-        raise NotImplementedError()
-
-    def _predict(self, *args, **kwargs):
-        raise NotImplementedError()
-    
-    def predict(self, *args, **kwargs):
-        predictions = [e.predict() for e in self._experts]
-        actions = [p[0] for p in predictions]
-        states = [p[1] for p in predictions]
-        return actions, states
-    
-    def forward(self, *args, **kwargs):
-        raise NotImplementedError()
-
-    def _predict(self, *args, **kwargs):
-        raise NotImplementedError()
-
-def save_video(env, expert):
-    env.reset()
-    env.render()
-    done = False
-    while not done:
-        actions, _ = expert.predict()
-        _, _, done, _ = env.step(actions)
-        env.render()
-    env.close()
-
-def demonstrations(expert='NormalizedIntersimpleExpert', env='NRasterizedRandomAgent', path=None, min_timesteps=25000, min_episodes=None, video=False, env_args={}, policy_args={}):
-    """Rollout and save expert demos.
-    
-    Usage:
-        python -m intersimple.expert <flags>
-
-    """
-    Env = globals()[env]
-    Expert = globals()[expert]
-
-    env = Env(**env_args)
-    info_env = RolloutInfoWrapper(env)
-    venv = DummyVecEnv([lambda: info_env])
-
-    policy = Expert(env, **policy_args)
-    venv_policy = DummyVecEnvPolicy([lambda: policy])
-
-    if video:
-        save_video(env, policy)
-    
-    path = path or (policy.__class__.__name__ + '_' + env.__class__.__name__ + '.pkl')
-    include_infos = isinstance(env, InfoFilter)
-
-    rollout.rollout_and_save(
-        path=path,
-        policy=venv_policy,
-        venv=venv,
-        sample_until=rollout.make_sample_until(
-            min_timesteps=min_timesteps,
-            min_episodes=min_episodes,
-        ),
-        exclude_infos=not include_infos,
-    )
-
-if __name__ == '__main__':
-    import fire
-    fire.Fire(demonstrations)

scratch/etienne/intersimple/data/generate.sh

@@ -1,15 +0,0 @@
-#python -m expert --env=IntersimpleReward --min_timesteps=200 --env_args='{agent:51}' --path='NormalizedIntersimpleExpert_IntersimpleRewardAgent51.pkl'
-#python -m expert --env=IntersimpleReward --min_timesteps=200 --env_args='{agent:51}' --policy_args='{mu:0.005}' --path='NormalizedIntersimpleExpert_IntersimpleRewardAgent51Mu.005.pkl'
-#python -m expert --env=IntersimpleReward --min_timesteps=200 --env_args='{agent:51}' --policy_args='{mu:0.001}' --path='NormalizedIntersimpleExpert_IntersimpleRewardAgent51Mu.001.pkl'
-#python -m expert --env=NRasterized --min_timesteps=200 --env_args='{agent:51,width:36,height:36,m_per_px:2}' --policy_args='{mu:0.001}' --path='NormalizedIntersimpleExpertMu.001_NRasterizedAgent51w36h36mppx2.pkl'
-#python -m expert --env=NRasterized --min_timesteps=200 --env_args='{agent:51,width:36,height:36,m_per_px:2}' --policy_args='{mu:0.001}' --path='NormalizedIntersimpleExpertMu.001_NRasterizedAgent51w36h36mppx2.pkl'
-#python -m expert --env=NRasterized --min_timesteps=3000 --video --env_args='{width:36,height:36,m_per_px:2}' --policy_args='{mu:0.001}' --path='NormalizedIntersimpleExpertMu.001_NRasterizedRandomAgentw36h36mppx2.pkl'
-#python -m expert --env=NRasterizedRandomAgent --min_timesteps=200 --env_args='{width:36,height:36,m_per_px:2}' --policy_args='{mu:0.001}' --path='NormalizedIntersimpleExpertMu.001_NRasterizedRandomAgentw36h36mppx2.pkl'
-#python -m expert --env=NRasterizedRandomAgent --min_timesteps=10000 --env_args='{width:36,height:36,m_per_px:2}' --policy_args='{mu:0.001}' --path='NormalizedIntersimpleExpertMu.001N10000_NRasterizedRandomAgentw36h36mppx2.pkl'
-#python -m expert --env=NRasterizedRouteRandomAgent --min_timesteps=10000 --env_args='{width:70,height:70,m_per_px:1}' --policy_args='{mu:0.001}' --path='NormalizedIntersimpleExpertMu.001N10000_NRasterizedRouteRandomAgentw70h70mppx1.pkl'
-#python -m expert --env=NRasterizedRouteRandomAgentLocation --min_timesteps=100000 --env_args='{width:70,height:70,m_per_px:1}' --policy_args='{mu:0.001}' --path='NormalizedIntersimpleExpertMu.001N100000_NRasterizedRouteRandomAgentLocationw70h70mppx1.pkl'
-#python -m expert --env=NRasterizedRouteRandomAgentLocation --min_timesteps=100000 --env_args='{width:70,height:70,m_per_px:1,map_color:128}' --policy_args='{mu:0.001}' --path='NormalizedIntersimpleExpertMu.001N100000_NRasterizedRouteRandomAgentLocationw70h70mppx1mapc128.pkl'
-#python -m expert --env=NRasterizedRouteSpeedRandomAgentLocation --min_timesteps=10000 --env_args='{width:70,height:70,m_per_px:1,map_color:128,mu:0.001}' --policy_args='{mu:0.001}' --path='NormalizedIntersimpleExpertMu.001N10000_NRasterizedRouteSpeedRandomAgentLocationw70h70mppx1mapc128mu.001.pkl'
-#python -m data.expert --env=NRasterizedRouteSpeedRandomAgentLocation --min_timesteps=10000 --env_args='{width:70,height:70,m_per_px:1,map_color:128,mu:0.001,skip_frames:5}' --policy_args='{mu:0.001}' --path='NormalizedIntersimpleExpertMu.001N10000_NRasterizedRouteSpeedRandomAgentLocationw70h70mppx1mapc128mu.001skip5.pkl'
-#python -m data.expert --env=TLNRasterizedRouteRandomAgentLocation --min_timesteps=100000 --env_args='{width:70,height:70,m_per_px:1,mu:0.001,random_skip:True,max_episode_steps:50}' --policy_args='{mu:0.001}' --path='NormalizedIntersimpleExpertMu.001N100000_TLNRasterizedRouteRandomAgentLocationw70h70mppx1mu.001rskips50.pkl'
-python -m data.expert --env=TLNRasterizedRouteRandomAgentLocation --min_timesteps=50000 --env_args='{width:70,height:70,m_per_px:1,mu:0.001,random_skip:True,max_episode_steps:50}' --policy_args='{mu:0.001}' --path='NormalizedIntersimpleExpertMu.001N50000_TLNRasterizedRouteRandomAgentLocationw70h70mppx1mu.001rskips50.pkl'

scratch/etienne/intersimple/gail/envs.py

@@ -1,46 +0,0 @@
-import gym
-from gym.wrappers.time_limit import TimeLimit
-import numpy as np
-from intersim.envs.intersimple import NRasterizedRouteRandomAgentLocation, RandomLocation, RandomAgent, RewardVisualization, Reward, \
-    ImageObservationAnimation, RasterizedRoute, NObservations, RasterizedObservation, \
-    NormalizedActionSpace, ActionVisualization, InteractionSimulatorMarkerViz, ImitationCompat, Intersimple
-
-class RasterizedSpeed:
-
-    def __init__(self, max_speed=12, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        channels, height, width = self.observation_space.shape
-        self.observation_space = gym.spaces.Box(
-            low=0,
-            high=255,
-            shape=(channels+1, height, width),
-            dtype=np.uint8
-        )
-        self._max_speed = max_speed
-
-    def _simple_obs(self, intersim_obs, intersim_info):
-        img = super()._simple_obs(intersim_obs, intersim_info)
-
-        ego_speed = intersim_obs['state'][self._agent, 2]
-        scaled_speed = (255 * ego_speed) // self._max_speed
-        speed_layer = scaled_speed * np.ones_like(img[:1], dtype=np.uint8)
-        speed_layer = speed_layer.clamp(0, 255)
-        
-        obs = np.concatenate((img, speed_layer), axis=0)
-        return obs
-
-class NRasterizedRouteSpeedRandomAgentLocation(RandomLocation, RandomAgent, RewardVisualization,
-        Reward, ImageObservationAnimation, RasterizedRoute, NObservations, RasterizedSpeed, RasterizedObservation,
-        NormalizedActionSpace, ActionVisualization, InteractionSimulatorMarkerViz, ImitationCompat, Intersimple):
-    pass
-
-class TransparentTimeLimit(TimeLimit):
-
-    def __getattr__(self, name):
-        return getattr(self.env, name)
-    
-    def close(self, *args, **kwargs):
-        return self.env.close(*args, **kwargs)
-
-def TLNRasterizedRouteRandomAgentLocation(max_episode_steps, *args, **kwargs):
-    return TransparentTimeLimit(NRasterizedRouteRandomAgentLocation(*args, **kwargs), max_episode_steps=max_episode_steps)

scratch/etienne/intersimple/gail/options2.py

@@ -1,127 +0,0 @@
-import gym
-import torch
-from src.util.collisions import feasible
-import numpy as np
-from collections import deque
-
-def imitation_discriminator(discriminator):
-    return lambda obs, action, next_obs, done: discriminator.discrim_net.predict_reward_train(
-        state=torch.tensor(obs).unsqueeze(0).to(discriminator.discrim_net.device()),
-        action=torch.tensor([[action]]).to(discriminator.discrim_net.device()),
-        next_state=torch.tensor(next_obs).unsqueeze(0).to(discriminator.discrim_net.device()), # unused
-        done=torch.tensor(done).unsqueeze(0).to(discriminator.discrim_net.device()), # unused
-    ).item()
-
-class OptionsEnv(gym.Wrapper):
-
-    def __init__(self, env, options, discriminator, discount, ll_buffer, *args, **kwargs):
-        super().__init__(env, *args, **kwargs)
-
-        self.options = options
-        num_hl_options = len(self.options)
-        self.action_space = gym.spaces.Discrete(num_hl_options)
-        self.observation_space = gym.spaces.Dict({
-            'obs': env.observation_space,
-            'mask': gym.spaces.Box(low=0, high=1, shape=(num_hl_options,)),
-        })
-
-        self.discriminator = discriminator
-        self.discount = discount
-        self.ll_buffer = ll_buffer
-    
-    @staticmethod
-    def _hl_observation(obs, mask):
-        return {
-            'obs': obs,
-            'mask': mask,
-        }
-    
-    def reset(self):
-        self.done = False
-        self.obs = self.env.reset()
-        self.m = available_actions(self.env, self.options)
-        return self._hl_observation(self.obs, self.m)
-    
-    def _ll_step(self, action):
-        return self.env.step(action)
-    
-    def step(self, action):
-        assert self.m[action]
-        assert not self.done
-
-        plan = list(map(float, generate_plan(self.env, action, self.options)))
-        reward = 0
-        steps = 0
-
-        while not self.done and plan and \
-            (feasible(self.env, safety_plan(self.env, plan)) or self.m.sum() == 1):
-            
-            a, plan = plan[0], plan[1:]
-            a = self.env._normalize(a)
-
-            next_obs, _, self.done, info = self._ll_step(a)
-
-            reward += self.discount**steps * self.discriminator(self.obs, a, next_obs, self.done)
-
-            self.ll_buffer.append({
-                'obs': self.obs,
-                'next_obs': next_obs,
-                'acts': np.array((a,)),
-                'dones': np.array(self.done),
-            })
-
-            steps += 1
-            self.obs = next_obs
-
-        self.m = available_actions(self.env, self.options)
-
-        return self._hl_observation(self.obs, self.m), reward, self.done, info
-
-class RenderOptions(OptionsEnv):
-
-    def __init__(self, env, options, *args, **kwargs):
-        super().__init__(env, options, discriminator=lambda s, a, n, d: 0, discount=1, ll_buffer=deque(maxlen=0), *args, **kwargs)
-
-    def _ll_step(self, action):
-        out = super()._ll_step(action)
-        self.env.render(mode='post')
-        return out
-    
-    def close(self, *args, **kwargs):
-        self.env.close(*args, **kwargs)
-
-def safety_plan(env, plan):
-    return np.concatenate((plan, np.array(5 * [env._env._min_acc])), axis=0)
-
-def available_actions(env, options):
-    """Return mask of available actions given current `env` state.
-    Action 0 is considered safe fallback.
-    """
-    plans = [generate_plan(env, i, options) for i, _ in enumerate(options)]
-    # is emergency braking still possible?
-    plans = list(map(lambda p: safety_plan(env, p), plans))
-
-    T = max(len(p) for p in plans)
-    plans = [np.pad(p, ((0, T-len(p)),), constant_values=np.nan) for p in plans]
-    plans = np.stack(plans, axis=0)
-    
-    valid = feasible(env, plans)
-    if not valid.any():
-        valid[0] = True
-
-    return valid
-
-def target_velocity_plan(current_v: float, target_v: float, t: int, dt: float):
-    """Smoothly target a velocity in a given number of steps"""
-    # for now, constant acceleration
-    a = (target_v - current_v)  / (t * dt)
-    return a*np.ones((t,))
-
-def generate_plan(env, i, options):
-    """Generate input profile for high-level action `i`."""
-    assert i < len(options), "Invalid option index {i}"
-    target_v, t = options[i]
-    current_v = env._env.state[env._agent, 1].item() # extract from env
-    plan = target_velocity_plan(current_v, target_v, t, env._env._dt)
-    assert len(plan) == t, "incorrect plan length"
-    return plan

scratch/etienne/intersimple/gail_flat.py

@@ -1,70 +0,0 @@
-# %%
-import pathlib
-import pickle
-import tempfile
-
-import stable_baselines3 as sb3
-from stable_baselines3.common.env_util import make_vec_env
-
-from imitation.algorithms import adversarial, bc
-from imitation.data import rollout
-from imitation.util import logger
-
-from intersim.envs.intersimple import IntersimpleReward
-
-from gail.discriminator import MlpDiscriminator
-
-model_name = 'gail_flat'
-
-# %%
-# Load pickled test demonstrations.
-with open("data/NormalizedIntersimpleExpert_IntersimpleRewardAgent51.pkl", "rb") as f:
-    # This is a list of `imitation.data.types.Trajectory`, where
-    # every instance contains observations and actions for a single expert
-    # demonstration.
-    trajectories = pickle.load(f)
-
-# %%
-# Convert List[types.Trajectory] to an instance of `imitation.data.types.Transitions`.
-# This is a more general dataclass containing unordered
-# (observation, actions, next_observation) transitions.
-transitions = rollout.flatten_trajectories(trajectories)
-
-venv = make_vec_env(IntersimpleReward, n_envs=2, env_kwargs={'agent': 51})
-
-tempdir = tempfile.TemporaryDirectory(prefix="quickstart")
-tempdir_path = pathlib.Path(tempdir.name)
-print(f"All Tensorboards and logging are being written inside {tempdir_path}/.")
-
-# Train GAIL on expert data.
-# GAIL, and AIRL also accept as `expert_data` any Pytorch-style DataLoader that
-# iterates over dictionaries containing observations, actions, and next_observations.
-logger.configure(tempdir_path / "GAIL/")
-gail_trainer = adversarial.GAIL(
-    venv,
-    expert_data=transitions,
-    expert_batch_size=150,
-    n_disc_updates_per_round=32,
-    discrim_kwargs={'discrim_net': MlpDiscriminator()},
-    gen_algo=sb3.PPO("MlpPolicy", venv, verbose=1, n_steps=4530),
-    allow_variable_horizon=True,
-)
-gail_trainer.train(total_timesteps=400000)
-gail_trainer.gen_algo.save(model_name)
-
-#del gail_trainer
-
-# %%
-model = sb3.PPO.load(model_name)
-
-env = IntersimpleReward(agent=51)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)

scratch/etienne/intersimple/gail_flat_ray.py

@@ -1,115 +0,0 @@
-# %%
-import pathlib
-import pickle
-import tempfile
-
-import stable_baselines3 as sb3
-from stable_baselines3.common.env_util import make_vec_env
-
-from imitation.algorithms import adversarial, bc
-from imitation.data import rollout
-from imitation.util import logger
-
-from intersim.envs.intersimple import IntersimpleReward, speed_reward
-
-from gail.discriminator import MlpDiscriminator
-import numpy as np
-import functools
-from stable_baselines3.common.evaluation import evaluate_policy
-from ray import tune
-import os
-import torch
-
-model_name = 'gail_flat'
-
-# %%
-# Load pickled test demonstrations.
-#with open("data/NormalizedIntersimpleExpert_IntersimpleRewardAgent51.pkl", "rb") as f:
-with open("data/NormalizedIntersimpleExpert_IntersimpleRewardAgent51Mu.001.pkl", "rb") as f:
-    # This is a list of `imitation.data.types.Trajectory`, where
-    # every instance contains observations and actions for a single expert
-    # demonstration.
-    trajectories = pickle.load(f)
-
-# %%
-# Convert List[types.Trajectory] to an instance of `imitation.data.types.Transitions`.
-# This is a more general dataclass containing unordered
-# (observation, actions, next_observation) transitions.
-transitions = rollout.flatten_trajectories(trajectories)
-
-venv = make_vec_env(IntersimpleReward, n_envs=2, env_kwargs={'agent': 51})
-
-tempdir = tempfile.TemporaryDirectory(prefix="quickstart")
-tempdir_path = pathlib.Path(tempdir.name)
-print(f"All Tensorboards and logging are being written inside {tempdir_path}/.")
-
-def training_function(config, checkpoint_dir=None):
-    logger.configure(tempdir_path / "GAIL/")
-
-    discriminator = MlpDiscriminator()
-    if checkpoint_dir:
-        discriminator.load_state_dict(torch.load(os.path.join(checkpoint_dir, 'disc_checkpoint')))
-        generator = sb3.PPO.load(os.path.join(checkpoint_dir, 'gen_checkpoint'))
-    else:
-        generator = sb3.PPO("MlpPolicy", venv, verbose=1, n_steps=config['n_steps'])
-
-    gail_trainer = adversarial.GAIL(
-        venv,
-        expert_data=transitions,
-        expert_batch_size=config['expert_batch_size'],
-        n_disc_updates_per_round=config['n_disc_updates_per_round'],
-        discrim_kwargs={'discrim_net': MlpDiscriminator()},
-        gen_algo=generator,
-        allow_variable_horizon=True,
-    )
-
-    def callback(epoch):
-        print("callback")
-        eval_env = IntersimpleReward(agent=51, reward=functools.partial(speed_reward, collision_penalty=0.))
-        #sync_envs_normalization(self.training_env, self.eval_env)
-        episode_rewards, episode_lengths = evaluate_policy(generator, eval_env, return_episode_rewards=True)
-        tune.report(
-            reward=np.mean(episode_rewards),
-            length=np.mean(episode_lengths),
-            training_iteration=epoch,
-        )
-
-        with tune.checkpoint_dir(step=epoch) as checkpoint_dir:
-            gail_trainer.gen_algo.save(os.path.join(checkpoint_dir, 'gen_checkpoint'))
-            torch.save(discriminator.state_dict(), os.path.join(checkpoint_dir, 'disc_checkpoint'))
-
-    gail_trainer.train(total_timesteps=40000, callback=callback)
-
-analysis = tune.run(
-    training_function,
-    config = {
-        'expert_batch_size': tune.randint(1, 22), #220,
-        'n_disc_updates_per_round': tune.randint(2, 100), #16,
-        'n_steps': tune.randint(1, 10000), #4096,
-    },
-    resources_per_trial={
-        'cpu': 1,
-        # 'gpu': 1,
-    },
-    local_dir='ray',
-    num_samples=10,
-)
-
-print('Best config', analysis.get_best_config(metric='progress', mode='max'))
-
-#del gail_trainer
-
-# %%
-model = sb3.PPO.load(model_name)
-
-env = IntersimpleReward(agent=51)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)

scratch/etienne/intersimple/gail_image.py

@@ -1,70 +0,0 @@
-# %%
-import pathlib
-import pickle
-import tempfile
-
-import stable_baselines3 as sb3
-from stable_baselines3.common.env_util import make_vec_env
-
-from imitation.algorithms import adversarial, bc
-from imitation.data import rollout
-from imitation.util import logger
-
-from intersim.envs.intersimple import NRasterized
-
-from gail.discriminator import CnnDiscriminator
-
-model_name = 'gail_image'
-
-# %%
-# Load pickled test demonstrations.
-with open("data/NormalizedIntersimpleExpertMu.001_NRasterizedAgent51w36h36mppx2.pkl", "rb") as f:
-    # This is a list of `imitation.data.types.Trajectory`, where
-    # every instance contains observations and actions for a single expert
-    # demonstration.
-    trajectories = pickle.load(f)
-
-# %%
-# Convert List[types.Trajectory] to an instance of `imitation.data.types.Transitions`.
-# This is a more general dataclass containing unordered
-# (observation, actions, next_observation) transitions.
-transitions = rollout.flatten_trajectories(trajectories)
-
-venv = make_vec_env(NRasterized, n_envs=2, env_kwargs={'agent': 51, 'width': 36, 'height': 36, 'm_per_px': 2})
-
-tempdir = tempfile.TemporaryDirectory(prefix="quickstart")
-tempdir_path = pathlib.Path(tempdir.name)
-print(f"All Tensorboards and logging are being written inside {tempdir_path}/.")
-
-# Train GAIL on expert data.
-# GAIL, and AIRL also accept as `expert_data` any Pytorch-style DataLoader that
-# iterates over dictionaries containing observations, actions, and next_observations.
-logger.configure(tempdir_path / "GAIL/")
-gail_trainer = adversarial.GAIL(
-    venv,
-    expert_data=transitions,
-    expert_batch_size=32,
-    #n_disc_updates_per_round=2048,
-    discrim_kwargs={'discrim_net': CnnDiscriminator(venv)},
-    gen_algo=sb3.PPO("CnnPolicy", venv, verbose=1, n_steps=1024),
-    allow_variable_horizon=True,
-)
-gail_trainer.train(total_timesteps=100000)
-gail_trainer.gen_algo.save(model_name)
-
-#del gail_trainer
-
-# %%
-model = sb3.PPO.load(model_name)
-
-env = NRasterized(agent=51, width=36, height=36, m_per_px=2)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)

scratch/etienne/intersimple/gail_image_multiagent_nocollision.py

@@ -1,70 +0,0 @@
-# %%
-import pathlib
-import pickle
-import tempfile
-
-import stable_baselines3 as sb3
-from stable_baselines3.common.env_util import make_vec_env
-
-from imitation.algorithms import adversarial, bc
-from imitation.data import rollout
-from imitation.util import logger
-
-from intersim.envs.intersimple import NRasterized
-
-from gail.discriminator import CnnDiscriminatorFlatAction
-
-model_name = 'gail_image_multiagent_nocollision'
-
-# %%
-# Load pickled test demonstrations.
-with open("data/NormalizedIntersimpleExpertMu.001_NRasterizedRandomAgentw36h36mppx2.pkl", "rb") as f:
-    # This is a list of `imitation.data.types.Trajectory`, where
-    # every instance contains observations and actions for a single expert
-    # demonstration.
-    trajectories = pickle.load(f)
-
-# %%
-# Convert List[types.Trajectory] to an instance of `imitation.data.types.Transitions`.
-# This is a more general dataclass containing unordered
-# (observation, actions, next_observation) transitions.
-transitions = rollout.flatten_trajectories(trajectories)
-
-venv = make_vec_env(NRasterized, n_envs=2, env_kwargs={'stop_on_collision':False, 'width': 36, 'height': 36, 'm_per_px': 2})
-
-tempdir = tempfile.TemporaryDirectory(prefix="quickstart")
-tempdir_path = pathlib.Path(tempdir.name)
-print(f"All Tensorboards and logging are being written inside {tempdir_path}/.")
-
-# Train GAIL on expert data.
-# GAIL, and AIRL also accept as `expert_data` any Pytorch-style DataLoader that
-# iterates over dictionaries containing observations, actions, and next_observations.
-logger.configure(tempdir_path / "GAIL/")
-gail_trainer = adversarial.GAIL(
-    venv,
-    expert_data=transitions,
-    expert_batch_size=32,
-    #n_disc_updates_per_round=2048,
-    discrim_kwargs={'discrim_net': CnnDiscriminatorFlatAction(venv)},
-    gen_algo=sb3.PPO("CnnPolicy", venv, verbose=1, n_steps=1024),
-    allow_variable_horizon=True,
-)
-gail_trainer.train(total_timesteps=100000)
-gail_trainer.gen_algo.save(model_name)
-
-#del gail_trainer
-
-# %%
-model = sb3.PPO.load(model_name)
-
-env = NRasterized(stop_on_collision=False, width=36, height=36, m_per_px=2)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)

scratch/etienne/intersimple/gail_image_random.py

@@ -1,79 +0,0 @@
-# %%
-import pathlib
-import pickle
-import tempfile
-
-import stable_baselines3 as sb3
-from stable_baselines3.common.env_util import make_vec_env
-
-from imitation.algorithms import adversarial, bc
-from imitation.data import rollout
-from imitation.util import logger
-
-from intersim.envs.intersimple import NRasterizedRandomAgent, IntersimpleReward, speed_reward
-import functools
-from stable_baselines3.common.evaluation import evaluate_policy
-
-from gail.discriminator import CnnDiscriminator
-
-model_name = 'gail_image_random'
-
-# %%
-# Load pickled test demonstrations.
-with open("data/NormalizedIntersimpleExpertMu.001_NRasterizedRandomAgentw36h36mppx2.pkl", "rb") as f:
-    # This is a list of `imitation.data.types.Trajectory`, where
-    # every instance contains observations and actions for a single expert
-    # demonstration.
-    trajectories = pickle.load(f)
-
-# %%
-# Convert List[types.Trajectory] to an instance of `imitation.data.types.Transitions`.
-# This is a more general dataclass containing unordered
-# (observation, actions, next_observation) transitions.
-transitions = rollout.flatten_trajectories(trajectories)
-
-env_kwargs = {'width': 36, 'height': 36, 'm_per_px': 2}
-venv = make_vec_env(NRasterizedRandomAgent, n_envs=2, env_kwargs=env_kwargs)
-
-tempdir = tempfile.TemporaryDirectory(prefix="quickstart")
-tempdir_path = pathlib.Path(tempdir.name)
-print(f"All Tensorboards and logging are being written inside {tempdir_path}/.")
-
-# Train GAIL on expert data.
-# GAIL, and AIRL also accept as `expert_data` any Pytorch-style DataLoader that
-# iterates over dictionaries containing observations, actions, and next_observations.
-logger.configure(tempdir_path / "GAIL/")
-generator = sb3.PPO("CnnPolicy", venv, verbose=1, n_steps=1024)
-gail_trainer = adversarial.GAIL(
-    venv,
-    expert_data=transitions,
-    expert_batch_size=32,
-    #n_disc_updates_per_round=2048,
-    discrim_kwargs={'discrim_net': CnnDiscriminator(venv)},
-    gen_algo=generator,
-    allow_variable_horizon=True,
-)
-def callback(round):
-    eval_env = NRasterizedRandomAgent(reward=functools.partial(speed_reward, collision_penalty=0.), **env_kwargs)
-    #sync_envs_normalization(self.training_env, self.eval_env)
-    episode_rewards, episode_lengths = evaluate_policy(generator, eval_env, return_episode_rewards=True)
-
-gail_trainer.train(total_timesteps=100000, callback=callback)
-gail_trainer.gen_algo.save(model_name)
-
-#del gail_trainer
-
-# %%
-model = sb3.PPO.load(model_name)
-
-env = NRasterizedRandomAgent(width=36, height=36, m_per_px=2)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)

scratch/etienne/intersimple/gail_image_random_ray.py

@@ -1,171 +0,0 @@
-# %%
-import pathlib
-import pickle
-import tempfile
-import os
-import random
-import numpy as np
-import torch
-
-# set up ray tune
-import ray
-from ray import tune
-from ray.tune import Analysis, ExperimentAnalysis
-from ray.tune.schedulers import ASHAScheduler
-from ray.tune.suggest.hyperopt import HyperOptSearch
-from ray.tune.suggest import ConcurrencyLimiter
-
-import stable_baselines3 as sb3
-from stable_baselines3.common.env_util import make_vec_env
-
-
-from imitation.algorithms import adversarial, bc
-from imitation.data import rollout
-from imitation.util import logger
-
-from intersim.envs.intersimple import NRasterizedRandomAgent, IntersimpleReward, speed_reward, NRasterized, NRasterizedRandomAgentVerbose
-import functools
-from stable_baselines3.common.evaluation import evaluate_policy
-from gym.wrappers import TimeLimit
-
-from gail.discriminator import CnnDiscriminator
-
-model_name = 'gail_image_random_ray'
-env_kwargs={'width': 36, 'height': 36, 'm_per_px': 2}
-
-# %%
-
-import argparse
-parser = argparse.ArgumentParser()
-parser.add_argument("--outdir", help="result directory", default='ray')
-parser.add_argument("--test", help="test run", default=False, action="store_true")
-args = parser.parse_args()
-outdir = args.outdir
-
-# %%
-# Load pickled test demonstrations.
-with open("data/NormalizedIntersimpleExpertMu.001N10000_NRasterizedRandomAgentw36h36mppx2.pkl", "rb") as f:
-    # This is a list of `imitation.data.types.Trajectory`, where
-    # every instance contains observations and actions for a single expert
-    # demonstration.
-    trajectories = pickle.load(f)
-
-# Convert List[types.Trajectory] to an instance of `imitation.data.types.Transitions`.
-# This is a more general dataclass containing unordered
-# (observation, actions, next_observation) transitions.
-transitions = rollout.flatten_trajectories(trajectories)
-# Store transitions in shared ray memory
-ray_transitions = ray.put(transitions)
-
-# %%
-venv = make_vec_env(NRasterizedRandomAgent, n_envs=2, env_kwargs=env_kwargs)
-
-tempdir = tempfile.TemporaryDirectory(prefix="quickstart")
-tempdir_path = pathlib.Path(tempdir.name)
-print(f"All Tensorboards and logging are being written inside {tempdir_path}/.")
-logger.configure(tempdir_path / "GAIL/")
-
-def get_ray_config(test=False):
-    if test:
-        return {
-            'expert_batch_size': 2,
-            'ppo_n_steps': 2,
-            'ppo_batch_size': 2,
-            'ppo_n_epochs': 1,
-            'total_timesteps': 10,
-        }
-    else:
-        return {
-            'expert_batch_size': tune.choice([2**x for x in range(6,10)]),
-            'ppo_n_steps': tune.choice([2048, 3072, 4096]),
-            'ppo_batch_size': tune.choice([2**x for x in range(9,13)]),
-            'ppo_n_epochs': tune.choice([6,10]),
-            'total_timesteps': 400_000,
-        }
-
-
-def ray_train(config, checkpoint_dir=None):
-    # Train GAIL on expert data.
-    # GAIL, and AIRL also accept as `expert_data` any Pytorch-style DataLoader that
-    # iterates over dictionaries containing observations, actions, and next_observations.
-
-    discriminator = CnnDiscriminator(venv)
-    if checkpoint_dir:
-        discriminator.load_state_dict(torch.load(os.path.join(checkpoint_dir, 'disc_checkpoint')))
-        generator = sb3.PPO.load(os.path.join(checkpoint_dir, 'gen_checkpoint'))
-    else:
-        generator = sb3.PPO(
-            "CnnPolicy", venv, verbose=0, 
-            n_steps=config["ppo_n_steps"], 
-            batch_size=config["ppo_batch_size"], 
-            n_epochs=config["ppo_n_epochs"]
-        )
-    gail_trainer = adversarial.GAIL(
-        venv,
-        expert_data=ray.get(ray_transitions),
-        expert_batch_size=config["expert_batch_size"],
-        #n_disc_updates_per_round=2048,
-        discrim_kwargs={'discrim_net': discriminator},
-        gen_algo=generator,
-        allow_variable_horizon=True,
-    )
-    def callback(round):
-        # eval_env = NRasterized(agent=51, reward=functools.partial(speed_reward, collision_penalty=0.), **env_kwargs)
-        eval_env = TimeLimit(NRasterizedRandomAgent(reward=functools.partial(speed_reward, collision_penalty=0.), **env_kwargs), max_episode_steps=1000)
-        episode_rewards, episode_lengths = evaluate_policy(generator, eval_env, return_episode_rewards=True)
-        tune.report(
-            reward=np.mean(episode_rewards),
-            length=np.mean(episode_lengths),
-            training_iteration=round,
-        )
-        with tune.checkpoint_dir(step=round) as checkpoint_dir:
-            gail_trainer.gen_algo.save(os.path.join(checkpoint_dir, 'gen_checkpoint'))
-            torch.save(discriminator.state_dict(), os.path.join(checkpoint_dir, 'disc_checkpoint'))
-
-    gail_trainer.train(total_timesteps=config['total_timesteps'], callback=callback)
-
-
-ray_config = get_ray_config(args.test)
-search = HyperOptSearch(ray_config, metric='length', mode="max",)
-search = ConcurrencyLimiter(search, max_concurrent=10)
-custom_scheduler = ASHAScheduler(time_attr='training_iteration', metric='length', mode="max", grace_period=15)
-
-analysis = tune.run(
-    ray_train, 
-    # config=ray_config,
-    search_alg=search,
-    scheduler=custom_scheduler,
-    local_dir=outdir,
-    resources_per_trial={"cpu":10, "gpu": 0.2},
-    num_samples=1 if args.test else 100,
-)
-
-del analysis
-
-# %%
-# outdir = "ray/ray_train_2021-09-20_13-33-50/ray_train_f06785b0_33_expert_batch_size=128,ppo_batch_size=1024,ppo_n_epochs=6,ppo_n_steps=2048,total_timesteps=400000_2021-09-20_15-52-05"
-
-# %%
-analysis = Analysis(outdir, default_metric="length", default_mode="max")
-filepath = analysis.get_best_logdir()
-print("Best ray experiment:", filepath)
-config = analysis.get_best_config()
-print("Best config:", config)
-
-# %%
-
-model = sb3.PPO.load(os.path.join(analysis.get_last_checkpoint(), 'gen_checkpoint'))
-
-# env = NRasterized(agent=51, **env_kwargs)
-env = TimeLimit(NRasterizedRandomAgent(reward=functools.partial(speed_reward, collision_penalty=0.), **env_kwargs), max_episode_steps=1000)
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.env.close(filestr='render/'+model_name)
-# %%
-

scratch/etienne/intersimple/gail_image_singleagent_nocollision.py

@@ -1,70 +0,0 @@
-# %%
-import pathlib
-import pickle
-import tempfile
-
-import stable_baselines3 as sb3
-from stable_baselines3.common.env_util import make_vec_env
-
-from imitation.algorithms import adversarial, bc
-from imitation.data import rollout
-from imitation.util import logger
-
-from intersim.envs.intersimple import NRasterized
-
-from gail.discriminator import CnnDiscriminator
-
-model_name = 'gail_image_singleagent_nocollision'
-
-# %%
-# Load pickled test demonstrations.
-with open("data/NormalizedIntersimpleExpertMu.001_NRasterizedAgent51w36h36mppx2.pkl", "rb") as f:
-    # This is a list of `imitation.data.types.Trajectory`, where
-    # every instance contains observations and actions for a single expert
-    # demonstration.
-    trajectories = pickle.load(f)
-
-# %%
-# Convert List[types.Trajectory] to an instance of `imitation.data.types.Transitions`.
-# This is a more general dataclass containing unordered
-# (observation, actions, next_observation) transitions.
-transitions = rollout.flatten_trajectories(trajectories)
-
-venv = make_vec_env(NRasterized, n_envs=2, env_kwargs={'agent':51, 'stop_on_collision':False, 'width': 36, 'height': 36, 'm_per_px': 2})
-
-tempdir = tempfile.TemporaryDirectory(prefix="quickstart")
-tempdir_path = pathlib.Path(tempdir.name)
-print(f"All Tensorboards and logging are being written inside {tempdir_path}/.")
-
-# Train GAIL on expert data.
-# GAIL, and AIRL also accept as `expert_data` any Pytorch-style DataLoader that
-# iterates over dictionaries containing observations, actions, and next_observations.
-logger.configure(tempdir_path / "GAIL/")
-gail_trainer = adversarial.GAIL(
-    venv,
-    expert_data=transitions,
-    expert_batch_size=32,
-    #n_disc_updates_per_round=2048,
-    discrim_kwargs={'discrim_net': CnnDiscriminator(venv)},
-    gen_algo=sb3.PPO("CnnPolicy", venv, verbose=1, n_steps=1024),
-    allow_variable_horizon=True,
-)
-gail_trainer.train(total_timesteps=100000)
-gail_trainer.gen_algo.save(model_name)
-
-#del gail_trainer
-
-# %%
-model = sb3.PPO.load(model_name)
-
-env = NRasterized(agent=51, width=36, height=36, m_per_px=2, stop_on_collision=False)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)

scratch/etienne/intersimple/gail_options_image.py

@@ -1,90 +0,0 @@
-# %%
-import sys
-sys.path.append('../../../')
-
-from src.discriminator import CnnDiscriminator, CnnDiscriminatorFlatAction
-from imitation.algorithms import adversarial
-import stable_baselines3
-import torch.utils.data
-import numpy as np
-from intersim.envs.intersimple import NRasterized
-import itertools
-from torch.distributions import Categorical
-import gym
-import torch
-import pickle
-import imitation.data.rollout as rollout
-import tempfile
-import pathlib
-from imitation.util import logger
-from stable_baselines3.common.env_util import make_vec_env
-from tqdm import tqdm
-from src.policies.options import OptionsCnnPolicy
-from src.gail.options import OptionsEnv, LLOptions, HLOptions, RenderOptions
-from src.gail.train import train_discriminator, train_generator
-
-model_name = 'gail_options_image'
-env_settings = {'agent': 51, 'width': 36, 'height': 36, 'm_per_px': 2}
-
-ALL_OPTIONS = [(v,t) for v in [0,2,4,6,8] for t in [5, 10]] # option 0 is safe fallback
-
-def train(expert_data, epochs=20, expert_batch_size=32, generator_steps=1024, discount=0.99):
-    env = NRasterized(**env_settings)
-    env.discount = discount
-
-    tempdir = tempfile.TemporaryDirectory(prefix="quickstart")
-    tempdir_path = pathlib.Path(tempdir.name)
-    logger.configure(tempdir_path / "GAIL/")
-    print(f"All Tensorboards and logging are being written inside {tempdir_path}/.")
-
-    venv = make_vec_env(NRasterized, n_envs=1, env_kwargs=env_settings)
-    discriminator = adversarial.GAIL(
-        expert_data=expert_data,
-        expert_batch_size=expert_batch_size,
-        discrim_kwargs={'discrim_net': CnnDiscriminatorFlatAction(venv)},
-        #discrim_kwargs={'discrim_net': CnnDiscriminator(venv)},
-        venv=venv, # unused
-        gen_algo=stable_baselines3.PPO("CnnPolicy", venv), # unused
-    )
-
-    generator = stable_baselines3.PPO(
-        OptionsCnnPolicy,
-        OptionsEnv(env, options=ALL_OPTIONS),
-        verbose=1,
-        n_steps=generator_steps,
-    )
-
-    # PPO.train requires logger as set up in
-    # PPO._setup_learn (called by PPO.learn)
-    generator._logger = stable_baselines3.common.utils.configure_logger(
-        generator.verbose,
-        generator.tensorboard_log,
-    )
-
-    for _ in tqdm(range(epochs)):
-        train_discriminator(LLOptions(env, options=ALL_OPTIONS), generator, discriminator, num_samples=expert_batch_size)
-        train_generator(HLOptions(env, options=ALL_OPTIONS), generator, discriminator, num_samples=generator_steps)
-    
-    return generator
-
-# %%
-if __name__ == '__main__':
-    # %%
- 
-    with open("data/NormalizedIntersimpleExpertMu.001_NRasterizedAgent51w36h36mppx2.pkl", "rb") as f:
-        trajectories = pickle.load(f)
-    transitions = rollout.flatten_trajectories(trajectories)
-    generator = train(transitions)
-
-    generator.save(model_name)
-
-    # %%
-    model = stable_baselines3.PPO.load(model_name)
-
-    env = RenderOptions(NRasterized(**env_settings), options=ALL_OPTIONS)
-
-    for s in env.sample_ll(model):
-        if s['dones']:
-            break
-
-    env.close(filestr='render/'+model_name)

scratch/etienne/intersimple/gail_options_image_alltracks.py

@@ -1,410 +0,0 @@
-# %%
-import sys
-sys.path.append('../../../')
-from src.discriminator import CnnDiscriminator, CnnDiscriminatorFlatAction
-from src.policies import OptionsCnnPolicy
-from src.util import feasible
-from src.data import load_experts
-
-from imitation.algorithms import adversarial
-from imitation.util import logger
-import imitation.data.rollout as rollout
-
-import stable_baselines3
-from stable_baselines3.common.env_util import make_vec_env
-
-import torch
-import torch.utils.data
-import numpy as np
-import itertools
-import gym
-import pickle
-import tempfile
-import pathlib
-from tqdm import tqdm
-
-from intersim.envs.intersimple import NRasterized, NRasterizedRandomAgent, NRasterizedIncrementingAgent
-
-ALL_OPTIONS = [(v,t) for v in [0,2,4,6,8] for t in [5, 10, 20]] # option 0 is safe fallback
-
-class OptionsEnv(gym.Wrapper):
-    """
-    Wrap an intersimple environment with an options generator
-    """
-    def __init__(self, env, *args, **kwargs):
-        """
-        Initialize wrapped environment and set high-level action and observation spaces
-        """
-        super().__init__(env, *args, **kwargs)
-        num_hl_options = len(ALL_OPTIONS)
-        self.action_space = gym.spaces.Discrete(num_hl_options)
-        self.observation_space = gym.spaces.Dict({
-            'obs': env.observation_space,
-            'mask': gym.spaces.Box(low=0, high=1, shape=(num_hl_options,)),
-        })
-
-    def _after_choice(self):
-        pass
-    
-    def _after_step(self):
-        pass
-
-    def _transitions(self):
-        raise NotImplementedError('Use `LLOptions` or `HLOptions` for sampling.')
-    
-    def sample(self, generator):
-        """
-        yield transitions using a generator
-        Args:
-            generator (sb3.PPO)
-        Yields:
-
-        """
-        self.done = True
-        while True:
-            self.episode_start = False
-            if self.done:
-                self.s = self.env.reset()
-                self.done = False
-                self.episode_start = True
-
-            self.m = available_actions(self.env)
-            self.ch, self.value, self.log_prob = generator.policy.predict({
-                'obs': torch.tensor(self.s).unsqueeze(0).to(generator.policy.device),
-                'mask': torch.tensor(self.m).unsqueeze(0).to(generator.policy.device),
-            })
-            self.plan = list(map(float, generate_plan(self.env, self.ch)))
-
-            self._after_choice()
-
-            assert not self.done
-            assert self.plan
-            #assert feasible(self.env, self.plan, self.ch)
-
-            while not self.done and self.plan and feasible(self.env, self.plan, self.ch):
-                self.a, self.plan = self.plan[0], self.plan[1:]
-                self.a = self.env._normalize(self.a)
-                self.nexts, _, self.done, _ = self.env.step(self.a)
-
-                self._after_step()
-
-                self.s = self.nexts
-            
-            yield from self._transitions()
-
-class LLOptions(OptionsEnv):
-    """Sample low-level (state, action) tuples for discriminator training."""
-    
-    def __init__(self, *args, **kwargs):
-        """
-        LLOption uses the true LL observations
-        """
-        super().__init__(*args, **kwargs)
-        # overwrite observation space to just output obs directly
-        self.observation_space = self.observation_space['obs']
-
-    def _after_choice(self):
-        """
-        After each option choice, initialize/reset the transition buffer
-        """
-        self._transition_buffer = []
-
-    def _after_step(self):
-        """
-        After each ll action, append s, s', a, done to transition buffer
-        """
-        self._transition_buffer.append({
-            'obs': self.s,
-            'next_obs': self.nexts,
-            'acts': np.array((self.a,)),
-            'dones': np.array(self.done),
-        })
-
-    def _transitions(self):
-        """
-        Yield from the transition buffer
-        """
-        yield from self._transition_buffer
-    
-    def sample_ll(self, policy):
-        """
-        Args:
-            policy
-        Returns:
-            gen: iterable which samples low-level transitions from the environment
-        """
-        return self.sample(policy)
-
-class HLOptions(OptionsEnv):
-    """Sample high-level (state, action, reward) tuples for generator training."""
-    
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-    def _after_choice(self):
-        """
-        After an option selection, initialize total reward and number of steps
-        """
-        self.obs = {'obs': np.copy(self.s), 'mask': np.copy(self.m)}
-        self.r = 0
-        self.steps = 0
-
-    def _after_step(self):
-        """
-        After each low-level action, add the discounted discriminated reward score (given a discriminator)
-        """
-        self.r += self.discount**self.steps * self.discriminator.discrim_net.reward_train(
-            state=torch.tensor(self.s).unsqueeze(0).to(self.discriminator.discrim_net.device()),
-            action=torch.tensor([[self.a]]).to(self.discriminator.discrim_net.device()),
-            next_state=torch.tensor(self.s).unsqueeze(0).to(self.discriminator.discrim_net.device()), # unused
-            done=torch.tensor(self.done).unsqueeze(0).to(self.discriminator.discrim_net.device()), # unused
-        )
-        self.steps += 1
-    
-    def _transitions(self):
-        """
-        Yield a single dictionary per high-level selected action
-        Fields:
-            obs: high-level state and mask at selection
-            action: chosen high-level action
-            reward: accumulated option reward
-            episode_start: whether the action was chosen at the episode start
-            value: the value estimate from the starting state
-            log_prob: the log_prob of the selected action from the starting state
-            done: whether the episode has ended
-
-        """
-        yield {
-            'obs': self.obs,
-            'action': self.ch,
-            'reward': self.r.detach(),
-            'episode_start': self.episode_start,
-            'value': self.value.detach(),
-            'log_prob': self.log_prob.detach(),
-            'done': self.done,
-        }
-    
-    def sample_hl(self, policy, discriminator):
-        """
-        Args:
-            policy
-            discriminator: function with which to score rewards
-        Returns:
-            gen: iterable which samples high-level transitions from the environment
-        """
-        self.discriminator = discriminator
-        return self.sample(policy)
-
-class RenderOptions(LLOptions):
-
-    def _after_step(self):
-        """
-        Render the environment after each low-level step
-        """
-        super()._after_step()
-        self.env.render()
-    
-    def close(self, *args, **kwargs):
-        """
-        On 'close', close the environment
-        """
-        self.env.close(*args, **kwargs)
-
-def available_actions(env):
-    """Return mask of available actions given current `env` state."""
-    valid = np.array([feasible(env, generate_plan(env, i), i) for i in range(len(ALL_OPTIONS))])
-    return valid
-
-def target_velocity_plan(current_v: float, target_v: float, t: int, dt: float):
-    """Smoothly target a velocity in a given number of steps"""
-    # for now, constant acceleration
-    a = (target_v - current_v)  / (t * dt)
-    return a*np.ones((t,))
-
-def generate_plan(env, i):
-    """Generate input profile for high-level action `i`.
-
-    Args:
-        env (gym.Env): current environment state
-        i (int): high-level action `i`
-    Returns:
-        plan (np.array): length T array of acceleration values
-    """
-    assert i < len(ALL_OPTIONS), "Invalid option index {i}"
-    target_v, t = ALL_OPTIONS[i]
-    current_v = env._env.state[env._agent, 1].item() # extract from env
-    plan = target_velocity_plan(current_v, target_v, t, env._env._dt)
-    assert len(plan) == t, "incorrect plan length"
-    return plan
-
-def flatten_transitions(transitions):
-    return {
-        'obs': np.stack(list(t['obs'] for t in transitions), axis=0),
-        'next_obs': np.stack(list(t['next_obs'] for t in transitions), axis=0),
-        'acts': np.stack(list(t['acts'] for t in transitions), axis=0),
-        'dones': np.stack(list(t['dones'] for t in transitions), axis=0),
-    }
-
-def train_discriminator(env, generator, discriminator, num_samples):
-    transitions = list(itertools.islice(env.sample_ll(generator), num_samples))
-    generator_samples = flatten_transitions(transitions)
-    discriminator.train_disc(gen_samples=generator_samples)
-
-def train_generator(env, generator, discriminator, num_samples):
-    generator_samples = list(itertools.islice(env.sample_hl(generator, discriminator), num_samples+1))
-    
-    generator.rollout_buffer.reset()
-    for s in generator_samples[:-1]:
-        generator.rollout_buffer.add(
-            obs=s['obs'],
-            action=s['action'].cpu(),
-            reward=s['reward'].cpu(),
-            episode_start=s['episode_start'],
-            value=s['value'],
-            log_prob=s['log_prob'],
-        )
-    
-    generator.rollout_buffer.compute_returns_and_advantage(
-        last_values=generator_samples[-1]['value'],
-        dones=generator_samples[-1]['done'],
-    )
-    
-    generator.train()
-
-def train(expert_data, env_class=NRasterizedRandomAgent, env_settings={}, epochs=10, discrim_batch_size=32, generator_steps=2048, discount=0.99):
-    """
-    Args: 
-        expert_data: list of transitions
-        env_class: environment class
-        env_settings: environment settings
-        epochs: number of epochs to train for
-        discrim_batch_size: discriminator batch size
-        generator_steps: number of steps taken in generator
-        discount: discount factor
-    Returns:
-        generator (stable_baselines3.PPO): options policy
-    """
-    env = env_class(**env_settings)
-    env.discount = discount
-
-    tempdir = tempfile.TemporaryDirectory(prefix="quickstart")
-    tempdir_path = pathlib.Path(tempdir.name)
-    logger.configure(tempdir_path / "GAIL/")
-    print(f"All Tensorboards and logging are being written inside {tempdir_path}/.")
-
-    venv = make_vec_env(env_class, n_envs=1, env_kwargs=env_settings)
-    discriminator = adversarial.GAIL(
-        expert_data=expert_data,
-        expert_batch_size=discrim_batch_size,
-        discrim_kwargs={'discrim_net': CnnDiscriminatorFlatAction(venv)},
-        #discrim_kwargs={'discrim_net': CnnDiscriminator(venv)},
-        venv=venv, # unused
-        gen_algo=stable_baselines3.PPO("CnnPolicy", venv), # unused
-    )
-
-    generator = stable_baselines3.PPO(
-        OptionsCnnPolicy,
-        OptionsEnv(env),
-        verbose=1,
-        n_steps=generator_steps,
-    )
-
-    # PPO.train requires logger as set up in
-    # PPO._setup_learn (called by PPO.learn)
-    generator._logger = stable_baselines3.common.utils.configure_logger(
-        generator.verbose,
-        generator.tensorboard_log,
-    )
-
-    for _ in tqdm(range(epochs)):
-        train_discriminator(LLOptions(env), generator, discriminator, num_samples=discrim_batch_size)
-        train_generator(HLOptions(env), generator, discriminator, num_samples=generator_steps)
-    
-    return generator
-
-# %%
-if __name__ == '__main__':
-    # %%
-    model_name = 'gail_options_image'
-    env_class = NRasterizedRandomAgent
-    env_settings = {'width': 36, 'height': 36, 'm_per_px': 2}
-    
-    #env_class = NRasterized
-    #env_settings = {'agent': 51, 'width': 36, 'height': 36, 'm_per_px': 2}
-    files = ['../../../expert_data/DR_USA_Roundabout_FT/track%04i/expert.pkl'%(i) for i in range(5)]
-    transitions=load_experts(files)
-
-    generator = train(
-        transitions,
-        env_class=env_class,
-        env_settings=env_settings,
-        epochs=10, 
-        discrim_batch_size=32, 
-        generator_steps=2048, 
-        discount=0.99
-        )
-
-    generator.save(model_name)
-
-    # %%
-    model = stable_baselines3.PPO.load(model_name)
-
-    env = RenderOptions(NRasterizedRandomAgent(**env_args))
-
-    for s in env.sample_ll(model):
-        if s['dones']:
-            break
-
-    env.close(filestr='render/'+model_name)
-
-# %% Tests
-
-def test_ll_expert_data():
-    with open("data/NormalizedIntersimpleExpertMu.001_NRasterizedAgent51w36h36mppx2.pkl", "rb") as f:
-        expert_trajectories = pickle.load(f)
-    expert_transitions = rollout.flatten_trajectories(expert_trajectories)
-
-    env = LLOptions(NRasterized(agent=51, width=36, height=36, m_per_px=2))
-
-    gen_transitions = list(itertools.islice(env.sample_ll(
-        policy=stable_baselines3.PPO(
-            OptionsCnnPolicy,
-            OptionsEnv(env),
-            verbose=1,
-        )
-    ), 10))
-    gen_transitions = flatten_transitions(gen_transitions)
-
-    assert expert_transitions[:10].obs.shape == gen_transitions['obs'].shape
-    assert expert_transitions[:10].next_obs.shape == gen_transitions['next_obs'].shape
-    assert expert_transitions[:10].acts.shape == gen_transitions['acts'].shape
-    assert expert_transitions[:10].dones.shape == gen_transitions['dones'].shape
-
-def test_ll_states():
-    env = NRasterized()
-    policy = stable_baselines3.PPO(
-        OptionsCnnPolicy,
-        OptionsEnv(env),
-        verbose=1,
-    )
-    llenv = LLOptions(env)
-    transitions = list(itertools.islice(llenv.sample_ll(policy=policy), 100))
-
-    env2 = NRasterized()
-    s2 = env2.reset()
-    for i, t in enumerate(transitions):
-        assert i == 0 or np.array_equal(t['obs'], transitions[i-1]['next_obs'])
-        assert np.array_equal(t['obs'], s2)
-        assert t['acts'].shape == (1,)
-
-        nexts2, _, done2, _ = env2.step(t['acts'])
-        assert np.array_equal(t['next_obs'], nexts2)
-        assert np.array_equal(t['dones'], done2)
-
-        if done2:
-            break
-
-        s2 = nexts2
-
-def test_hl_transitions():
-    pass

scratch/etienne/intersimple/gail_options_image_random.py

@@ -1,91 +0,0 @@
-# %%
-import sys
-sys.path.append('../../../')
-
-from src.discriminator import CnnDiscriminatorFlatAction
-from imitation.algorithms import adversarial
-import stable_baselines3
-import torch.utils.data
-import numpy as np
-from intersim.envs.intersimple import NRasterizedRouteRandomAgent
-import itertools
-from torch.distributions import Categorical
-import gym
-import torch
-import pickle
-import imitation.data.rollout as rollout
-import tempfile
-import pathlib
-from imitation.util import logger
-from stable_baselines3.common.env_util import make_vec_env
-from tqdm import tqdm
-from src.policies.options import OptionsCnnPolicy
-from src.gail.options import OptionsEnv, LLOptions, HLOptions, RenderOptions
-from src.gail.train import train_discriminator, train_generator
-
-model_name = 'gail_options_image_random'
-env_settings = {'width': 70, 'height': 70, 'm_per_px': 1}
-
-ALL_OPTIONS = [(v,t) for v in [0,2,4,6,8] for t in [5, 10, 20]] # option 0 is safe fallback
-
-def train(expert_data, epochs=100, expert_batch_size=64, generator_steps=1024, discount=0.99):
-    env = NRasterizedRouteRandomAgent(**env_settings)
-    env.discount = discount
-
-    tempdir = tempfile.TemporaryDirectory(prefix="quickstart")
-    tempdir_path = pathlib.Path(tempdir.name)
-    logger.configure(tempdir_path / "GAIL/")
-    print(f"All Tensorboards and logging are being written inside {tempdir_path}/.")
-
-    venv = make_vec_env(NRasterizedRouteRandomAgent, n_envs=1, env_kwargs=env_settings)
-    discriminator = adversarial.GAIL(
-        expert_data=expert_data,
-        expert_batch_size=expert_batch_size,
-        discrim_kwargs={'discrim_net': CnnDiscriminatorFlatAction(venv)},
-        #discrim_kwargs={'discrim_net': CnnDiscriminator(venv)},
-        venv=venv, # unused
-        gen_algo=stable_baselines3.PPO("CnnPolicy", venv), # unused
-    )
-
-    generator = stable_baselines3.PPO(
-        OptionsCnnPolicy,
-        OptionsEnv(env, options=ALL_OPTIONS),
-        verbose=1,
-        n_steps=generator_steps,
-    )
-
-    # PPO.train requires logger as set up in
-    # PPO._setup_learn (called by PPO.learn)
-    generator._logger = stable_baselines3.common.utils.configure_logger(
-        generator.verbose,
-        generator.tensorboard_log,
-    )
-
-    for _ in tqdm(range(epochs)):
-        train_discriminator(LLOptions(env, options=ALL_OPTIONS), generator, discriminator, num_samples=expert_batch_size)
-        train_generator(HLOptions(env, options=ALL_OPTIONS), generator, discriminator, num_samples=generator_steps)
-        generator.save(model_name)
-    
-    return generator
-
-def video(model_name, env):
-    model = stable_baselines3.PPO.load(model_name)
-    env = RenderOptions(env, options=ALL_OPTIONS)
-    for s in env.sample_ll(model):
-        if s['dones']:
-            break
-    env.close(filestr='render/'+model_name)
-
-def evaluate():
-    video(
-        model_name=model_name,
-        env=NRasterizedRouteRandomAgent(**env_settings)
-    )
-
-# %%
-if __name__ == '__main__':
-
-    with open("data/NormalizedIntersimpleExpertMu.001N10000_NRasterizedRouteRandomAgentw70h70mppx1.pkl", "rb") as f:
-        trajectories = pickle.load(f)
-    transitions = rollout.flatten_trajectories(trajectories)
-    train(transitions)

scratch/etienne/intersimple/gail_options_image_random_location.py

@@ -1,144 +0,0 @@
-# %%
-from collections import deque
-import sys
-sys.path.append('../../../')
-
-from src.discriminator import CnnDiscriminator, CnnDiscriminatorFlatAction
-from imitation.algorithms import adversarial
-import stable_baselines3
-import pickle
-import imitation.data.rollout as rollout
-import tempfile
-import pathlib
-from imitation.util import logger
-from tqdm import tqdm
-from src.policies.options import OptionsCnnPolicy
-from src.gail.train import flatten_transitions
-from gail.options2 import OptionsEnv, RenderOptions, imitation_discriminator
-from gail.envs import TLNRasterizedRouteRandomAgentLocation
-from stable_baselines3.common.vec_env.dummy_vec_env import DummyVecEnv
-from stable_baselines3.common.env_util import make_vec_env
-import torch
-import numpy as np
-
-model_name = 'gail_options_image_random_location'
-env_settings = {'width': 70, 'height': 70, 'm_per_px': 1, 'mu': 0.001, 'random_skip': True, 'max_episode_steps': 200}
-
-ALL_OPTIONS = [(v,t) for v in [0,2,4,8,10] for t in [5, 10, 20]] # option 0 is safe fallback
-
-class NoisyDiscriminator(CnnDiscriminatorFlatAction):
-
-    def __init__(self, *args, std=0.0, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.std = std
-
-    def forward(self, state, action):
-        noise = self.std * torch.randn(*action.shape, device=action.device)
-        return super().forward(state, action + noise)
-
-class LLBuffer(deque):
-
-    def sample(self, n):
-        assert n <= self.maxlen, f'Sample size of {n} exceeds buffer capacity of {self.maxlen}'
-        assert n <= len(self), f'Sample size of {n} exceeds buffer size of {len(self)}'
-        ind = np.random.randint(len(self), size=n)
-        return list(self[i] for i in ind)
-
-def train(
-        expert_data,
-        expert_batch_size=4096,
-        discriminator_updates_per_round=20,
-        generator_steps=1024,
-        generator_batch_size=1024,
-        generator_total_steps=8192,
-        generator_updates_per_round=10,
-        discount=1.0,
-        epochs=200,
-    ):
-    env = TLNRasterizedRouteRandomAgentLocation(**env_settings)
-
-    tempdir = tempfile.TemporaryDirectory(prefix="quickstart")
-    tempdir_path = pathlib.Path(tempdir.name)
-    logger.configure(tempdir_path / "GAIL/")
-    print(f"All Tensorboards and logging are being written inside {tempdir_path}/.")
-
-    venv = DummyVecEnv([lambda: env])
-    discriminator = adversarial.GAIL(
-        expert_data=expert_data,
-        expert_batch_size=expert_batch_size,
-        #discrim_kwargs={'discrim_net': NoisyDiscriminator(venv, std=0.25)},
-        disc_opt_cls=torch.optim.RMSprop,
-        disc_opt_kwargs={'lr': 0.0001, 'weight_decay': 0.003},
-        discrim_kwargs={'discrim_net': CnnDiscriminator(venv)},
-        venv=venv, # unused
-        gen_algo=stable_baselines3.PPO("CnnPolicy", venv), # unused
-    )
-
-    ll_buffer = LLBuffer(maxlen=expert_batch_size*10)
-
-    options_env = make_vec_env(
-        OptionsEnv,
-        n_envs=1,
-        #vec_env_cls=SubprocVecEnv,
-        env_kwargs={
-            'env': env,
-            'options': ALL_OPTIONS,
-            'discriminator': imitation_discriminator(discriminator),
-            'discount': discount,
-            'll_buffer': ll_buffer,
-        }
-    )
-
-    generator = stable_baselines3.PPO(
-        OptionsCnnPolicy,
-        options_env,
-        verbose=1,
-        batch_size=generator_batch_size,
-        n_steps=generator_steps,
-        n_epochs=generator_updates_per_round,
-        gamma=1.0,
-        learning_rate=1e-4,
-    )
-
-    for _ in tqdm(range(epochs)):
-        ll_buffer.clear()
-
-        # train generator
-        generator.learn(total_timesteps=generator_total_steps)
-
-        # train discriminator
-        for _ in range(discriminator_updates_per_round):
-            generator_samples = ll_buffer.sample(expert_batch_size)
-            generator_samples = flatten_transitions(generator_samples)
-            discriminator.train_disc(gen_samples=generator_samples)
-
-        generator.save(model_name)
-    
-    return generator
-
-def video(model_name, env):
-    model = stable_baselines3.PPO.load(model_name)
-
-    done = False
-    obs = env.reset()
-    while not done:
-        action, _ = model.predict(obs)
-        obs, _, done, _ = env.step(action)
-
-    env.close(filestr='render/'+model_name)
-
-def evaluate():
-    video_settings = { **env_settings, 'random_skip': False, 'max_episode_steps': 200 }
-    env = TLNRasterizedRouteRandomAgentLocation(**video_settings)
-    env = RenderOptions(env, options=ALL_OPTIONS)
-    video(
-        model_name=model_name,
-        env=env
-    )
-
-# %%
-if __name__ == '__main__':
-    with open("data/NormalizedIntersimpleExpertMu.001N50000_TLNRasterizedRouteRandomAgentLocationw70h70mppx1mu.001rskips50.pkl", "rb") as f:
-        trajectories = pickle.load(f)
-    transitions = rollout.flatten_trajectories(trajectories)
-    train(transitions)

scratch/etienne/intersimple/imitation_quickstart.py

@@ -1,63 +0,0 @@
-# %%
-import pathlib
-import pickle
-import tempfile
-
-import stable_baselines3 as sb3
-from stable_baselines3.common.env_util import make_vec_env
-
-from imitation.algorithms import adversarial, bc
-from imitation.data import rollout
-from imitation.util import logger
-
-from intersim.envs.intersimple import IntersimpleReward
-
-# Load pickled test demonstrations.
-with open("data/NormalizedIntersimpleExpert_IntersimpleRewardAgent51.pkl", "rb") as f:
-    # This is a list of `imitation.data.types.Trajectory`, where
-    # every instance contains observations and actions for a single expert
-    # demonstration.
-    trajectories = pickle.load(f)
-
-# %%
-# Convert List[types.Trajectory] to an instance of `imitation.data.types.Transitions`.
-# This is a more general dataclass containing unordered
-# (observation, actions, next_observation) transitions.
-transitions = rollout.flatten_trajectories(trajectories)
-
-venv = make_vec_env(IntersimpleReward, n_envs=2, env_kwargs={'agent': 51})
-
-tempdir = tempfile.TemporaryDirectory(prefix="quickstart")
-tempdir_path = pathlib.Path(tempdir.name)
-print(f"All Tensorboards and logging are being written inside {tempdir_path}/.")
-
-# Train BC on expert data.
-# BC also accepts as `expert_data` any PyTorch-style DataLoader that iterates over
-# dictionaries containing observations and actions.
-logger.configure(tempdir_path / "BC/")
-bc_trainer = bc.BC(venv.observation_space, venv.action_space, expert_data=transitions)
-bc_trainer.train(n_epochs=1)
-
-# Train GAIL on expert data.
-# GAIL, and AIRL also accept as `expert_data` any Pytorch-style DataLoader that
-# iterates over dictionaries containing observations, actions, and next_observations.
-logger.configure(tempdir_path / "GAIL/")
-gail_trainer = adversarial.GAIL(
-    venv,
-    expert_data=transitions,
-    expert_batch_size=32,
-    gen_algo=sb3.PPO("MlpPolicy", venv, verbose=1, n_steps=1024),
-)
-gail_trainer.train(total_timesteps=2048)
-
-# Train AIRL on expert data.
-logger.configure(tempdir_path / "AIRL/")
-airl_trainer = adversarial.AIRL(
-    venv,
-    expert_data=transitions,
-    expert_batch_size=32,
-    gen_algo=sb3.PPO("MlpPolicy", venv, verbose=1, n_steps=1024),
-)
-airl_trainer.train(total_timesteps=2048)
-
-# %%

scratch/etienne/intersimple/ppo_const.py

@@ -1,36 +0,0 @@
-# %%
-from stable_baselines3 import PPO
-from stable_baselines3.common.env_util import make_vec_env
-from intersim.envs.intersimple import IntersimpleReward, speed_reward
-
-model_name = "ppo_const"
-
-env = IntersimpleReward(
-    agent=51,
-    #reward=speed_reward,
-)
-
-# %%
-model = PPO(
-    "MlpPolicy", env,
-    verbose=1,
-)
-model.learn(total_timesteps=100000)
-model.save(model_name)
-
-print('Done training.')
-
-del model # remove to demonstrate saving and loading
-
-# %%
-model = PPO.load(model_name)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)

scratch/etienne/intersimple/ppo_const_collision.py

@@ -1,41 +0,0 @@
-# %%
-from stable_baselines3 import PPO
-from stable_baselines3.common.env_util import make_vec_env
-from intersim.envs.intersimple import ConstCollisionReward, IntersimpleFlatAgent
-
-model_name = "ppo_const_collision"
-
-class IntersimpleConstCollisionAgent(ConstCollisionReward, IntersimpleFlatAgent):
-    pass
-
-env = IntersimpleConstCollisionAgent(
-    agent=51,
-    speed_reward_weight=0.001,
-    collision_penalty=1000
-)
-
-# %%
-model = PPO(
-    "MlpPolicy", env,
-    learning_rate=3e-6,
-    verbose=1,
-)
-model.learn(total_timesteps=2e5)
-model.save(model_name)
-
-print('Done training.')
-
-del model # remove to demonstrate saving and loading
-
-# %%
-model = PPO.load(model_name)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)

scratch/etienne/intersimple/ppo_const_image.py

@@ -1,35 +0,0 @@
-# %%
-from stable_baselines3 import PPO
-from stable_baselines3.common.env_util import make_vec_env
-from intersim.envs.intersimple import NRasterized
-
-model_name = "ppo_const_image"
-
-env = NRasterized(
-    agent=51,
-)
-
-# %%
-model = PPO(
-    "CnnPolicy", env,
-    verbose=1,
-)
-model.learn(total_timesteps=100000)
-model.save(model_name)
-
-print('Done training.')
-
-del model # remove to demonstrate saving and loading
-
-# %%
-model = PPO.load(model_name)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)

scratch/etienne/intersimple/ppo_const_image_random.py

@@ -1,33 +0,0 @@
-# %%
-from stable_baselines3 import PPO
-from intersim.envs.intersimple import NRasterizedRandomAgent
-import functools
-
-model_name = "ppo_const_image_random"
-
-env = NRasterizedRandomAgent()
-
-# %%
-model = PPO(
-    "CnnPolicy", env,
-    verbose=1,
-)
-model.learn(total_timesteps=2e5)
-model.save(model_name)
-
-print('Done training.')
-
-del model # remove to demonstrate saving and loading
-
-# %%
-model = PPO.load(model_name)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)

scratch/etienne/intersimple/ppo_intersimple_tspeed.py

@@ -1,25 +0,0 @@
-from stable_baselines3 import PPO
-from stable_baselines3.common.env_util import make_vec_env
-from intersim.envs.intersimple import IntersimpleTargetSpeed
-
-env = IntersimpleTargetSpeed()
-
-model = PPO("MlpPolicy", env, verbose=1)
-model.learn(total_timesteps=25000)
-model.save("ppo_intersimple")
-
-print('Done training.')
-
-del model # remove to demonstrate saving and loading
-
-model = PPO.load("ppo_intersimple")
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close()

scratch/etienne/intersimple/ppo_speed.py

@@ -1,46 +0,0 @@
-# %%
-from stable_baselines3 import PPO
-from intersim.envs.intersimple import IntersimpleReward, speed_reward
-import functools
-
-model_name = "ppo_speed"
-
-#def reward(state, action, info):
-#    speed = state[2].item()
-#    r = speed if speed < 10 else (10 - 5 * (speed - 10))
-#    return 0.1 * r
-
-env = IntersimpleReward(
-    agent=51,
-    reward=functools.partial(
-        speed_reward,
-        collision_penalty=0
-    ),
-)
-
-# %%
-model = PPO(
-    "MlpPolicy", env,
-    verbose=1,
-)
-model.learn(total_timesteps=100000)
-model.save(model_name)
-
-print('Done training.')
-
-del model # remove to demonstrate saving and loading
-
-# %%
-model = PPO.load(model_name)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)
-
-# %%

scratch/etienne/intersimple/ppo_speed_image.py

@@ -1,46 +0,0 @@
-# %%
-from stable_baselines3 import PPO
-from intersim.envs.intersimple import NRasterized, speed_reward
-import functools
-
-model_name = "ppo_speed_image"
-
-#def reward(state, action, info):
-#    speed = state[2].item()
-#    r = speed if speed < 10 else (10 - 5 * (speed - 10))
-#    return 0.1 * r
-
-env = NRasterized(
-    agent=20,
-    reward=functools.partial(
-        speed_reward,
-        collision_penalty=0
-    ),
-)
-
-# %%
-model = PPO(
-    "CnnPolicy", env,
-    verbose=1,
-)
-model.learn(total_timesteps=100000)
-model.save(model_name)
-
-print('Done training.')
-
-del model # remove to demonstrate saving and loading
-
-# %%
-model = PPO.load(model_name)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)
-
-# %%

scratch/etienne/intersimple/ppo_speed_image_lowres.py

@@ -1,49 +0,0 @@
-# %%
-from stable_baselines3 import PPO
-from intersim.envs.intersimple import NRasterized, speed_reward
-import functools
-
-model_name = "ppo_speed_image_lowres"
-
-#def reward(state, action, info):
-#    speed = state[2].item()
-#    r = speed if speed < 10 else (10 - 5 * (speed - 10))
-#    return 0.1 * r
-
-env = NRasterized(
-    agent=51,
-    height=36,
-    width=36,
-    m_per_px=2,
-    reward=functools.partial(
-        speed_reward,
-        collision_penalty=0
-    ),
-)
-
-# %%
-model = PPO(
-    "CnnPolicy", env,
-    verbose=1,
-)
-model.learn(total_timesteps=100000)
-model.save(model_name)
-
-print('Done training.')
-
-del model # remove to demonstrate saving and loading
-
-# %%
-model = PPO.load(model_name)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)
-
-# %%

scratch/etienne/intersimple/ppo_speed_image_lowres_random.py

@@ -1,42 +0,0 @@
-# %%
-from stable_baselines3 import PPO
-from intersim.envs.intersimple import NRasterizedRandomAgent, speed_reward
-import functools
-
-model_name = "ppo_speed_image_lowres_random"
-
-env = NRasterizedRandomAgent(
-    height=36,
-    width=36,
-    m_per_px=2,
-    reward=functools.partial(
-        speed_reward,
-        collision_penalty=0
-    )
-)
-
-# %%
-model = PPO(
-    "CnnPolicy", env,
-    verbose=1,
-    batch_size=2048,
-)
-model.learn(total_timesteps=2e5)
-model.save(model_name)
-
-print('Done training.')
-
-del model # remove to demonstrate saving and loading
-
-# %%
-model = PPO.load(model_name)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)

scratch/etienne/intersimple/ppo_speed_image_random.py

@@ -1,39 +0,0 @@
-# %%
-from stable_baselines3 import PPO
-from intersim.envs.intersimple import NRasterizedRandomAgent, speed_reward
-import functools
-
-model_name = "ppo_speed_image_random"
-
-env = NRasterizedRandomAgent(
-    reward=functools.partial(
-        speed_reward,
-        collision_penalty=0
-    )
-)
-
-# %%
-model = PPO(
-    "CnnPolicy", env,
-    verbose=1,
-    batch_size=2048,
-)
-model.learn(total_timesteps=2e5)
-model.save(model_name)
-
-print('Done training.')
-
-del model # remove to demonstrate saving and loading
-
-# %%
-model = PPO.load(model_name)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)

scratch/etienne/intersimple/ppo_speed_lidar.py

@@ -1,50 +0,0 @@
-# %%
-from stable_baselines3 import PPO
-from intersim.envs import IntersimpleLidarFlat
-from intersim.envs.intersimple import speed_reward
-import functools
-
-model_name = "ppo_speed_lidar"
-
-#def reward(state, action, info):
-#    speed = state[2].item()
-#    r = speed if speed < 10 else (10 - 5 * (speed - 10))
-#    return 0.1 * r
-
-env = IntersimpleLidarFlat(
-    n_rays=5,
-    agent=51,
-    reward=functools.partial(
-        speed_reward,
-        collision_penalty=0
-    ),
-)
-
-# %%
-model = PPO(
-    "MlpPolicy", env,
-    learning_rate=1e-4,
-    verbose=1,
-    tensorboard_log='runs/'
-)
-model.learn(total_timesteps=100000)
-model.save(model_name)
-
-print('Done training.')
-
-del model # remove to demonstrate saving and loading
-
-# %%
-model = PPO.load(model_name)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)
-
-# %%

scratch/etienne/intersimple/ppo_speed_lidar_random.py

@@ -1,49 +0,0 @@
-# %%
-from stable_baselines3 import PPO
-from intersim.envs import IntersimpleLidarFlatRandom
-from intersim.envs.intersimple import speed_reward
-import functools
-
-model_name = "ppo_speed_lidar_random"
-
-#def reward(state, action, info):
-#    speed = state[2].item()
-#    r = speed if speed < 10 else (10 - 5 * (speed - 10))
-#    return 0.1 * r
-
-env = IntersimpleLidarFlatRandom(
-    n_rays=5,
-    reward=functools.partial(
-        speed_reward,
-        collision_penalty=0
-    ),
-)
-
-# %%
-model = PPO(
-    "MlpPolicy", env,
-    learning_rate=1e-4,
-    verbose=1,
-    tensorboard_log='runs/'
-)
-model.learn(total_timesteps=1000000)
-model.save(model_name)
-
-print('Done training.')
-
-del model # remove to demonstrate saving and loading
-
-# %%
-model = PPO.load(model_name)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)
-
-# %%

scratch/etienne/intersimple/ppo_speed_random.py

@@ -1,43 +0,0 @@
-# %%
-from stable_baselines3 import PPO
-from intersim.envs.intersimple import IntersimpleFlatRandomAgent, Reward, RewardVisualization, speed_reward
-import functools
-
-model_name = "ppo_speed_random"
-
-class IntersimpleRewardRandom(RewardVisualization, Reward, IntersimpleFlatRandomAgent):
-    """`IntersimpleFlatAgent` with rewards."""
-    pass
-
-env = IntersimpleRewardRandom(
-    reward=functools.partial(
-        speed_reward,
-        collision_penalty=0
-    )
-)
-
-# %%
-model = PPO(
-    "MlpPolicy", env,
-    verbose=1,
-    batch_size=2048,
-)
-model.learn(total_timesteps=2e5)
-model.save(model_name)
-
-print('Done training.')
-
-del model # remove to demonstrate saving and loading
-
-# %%
-model = PPO.load(model_name)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)

scratch/etienne/intersimple/ppo_tspeed.py

@@ -1,39 +0,0 @@
-# %%
-from stable_baselines3 import PPO
-from stable_baselines3.common.env_util import make_vec_env
-from intersim.envs.intersimple import IntersimpleTargetSpeedAgent
-
-model_name = "ppo_tspeed"
-
-env = IntersimpleTargetSpeedAgent(
-    agent=51,
-    target_speed=10,
-    speed_penalty_weight=0.001,
-    collision_penalty=1000
-)
-
-# %%
-model = PPO(
-    "MlpPolicy", env,
-    learning_rate=3e-6,
-    verbose=1,
-)
-model.learn(total_timesteps=2e5)
-model.save(model_name)
-
-print('Done training.')
-
-del model # remove to demonstrate saving and loading
-
-# %%
-model = PPO.load(model_name)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)

scratch/etienne/intersimple/ppo_tspeed_random.py

@@ -1,31 +0,0 @@
-# %%
-from stable_baselines3 import PPO
-from stable_baselines3.common.env_util import make_vec_env
-from intersim.envs.intersimple import IntersimpleTargetSpeedRandom
-
-model_name = "ppo_tspeed_random"
-
-# %%
-env = IntersimpleTargetSpeedRandom(target_speed=10)
-
-# %%
-model = PPO("MlpPolicy", env, verbose=1)
-model.learn(total_timesteps=250000)
-model.save(model_name)
-
-print('Done training.')
-
-del model # remove to demonstrate saving and loading
-
-# %%
-model = PPO.load(model_name)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close()

scratch/etienne/intersimple/render_env_from_model.py

@@ -1,33 +0,0 @@
-
-import stable_baselines3 as sb3
-from intersim.envs.intersimple import NRasterized
-
-
-def render_env(model_name='gail_image_multiagent_nocollision', agent=51, environment=NRasterized):
-    """
-    Render a video from an model, agent, and environment
-    Args:
-        model_name (str): name of the model
-        agent (int): agent to start the video from
-        environment (gym.Env): gym environment class to render environment on
-    """
-
-    model = sb3.PPO.load(model_name)
-
-    env = environment(stop_on_collision=False, width=36, height=36, m_per_px=2, agent=agent)
-
-    obs = env.reset()
-    i=0
-    while True and i < 600:
-        i+=1
-        action, _states = model.predict(obs)
-        obs, rewards, done, info = env.step(action)
-        env.render(mode='post')
-        if done:
-            break
-
-    env.close(filestr='render/'+model_name+'_agent%i'%(agent)) 
-
-if __name__ == '__main__':
-    import fire
-    fire.Fire(render_env)

scratch/etienne/intersimple/train_discrim.py

@@ -1,71 +0,0 @@
-# %%
-import sys
-sys.path.append('../../../')
-
-import pickle
-import imitation.data.rollout as rollout
-import imitation.data.types as types
-import torch
-from gail.envs import TLNRasterizedRouteRandomAgentLocation
-import tempfile
-import pathlib
-from imitation.util import logger
-from stable_baselines3.common.vec_env.dummy_vec_env import DummyVecEnv
-from imitation.algorithms import adversarial
-from src.discriminator import CnnDiscriminator
-import stable_baselines3
-from tqdm import tqdm
-
-with open("data/NormalizedIntersimpleExpertMu.001N50000_TLNRasterizedRouteRandomAgentLocationw70h70mppx1mu.001rskips50.pkl", "rb") as f:
-    trajectories = pickle.load(f)
-transitions = rollout.flatten_trajectories(trajectories)
-
-# %%
-env_settings = {'width': 70, 'height': 70, 'm_per_px': 1, 'mu': 0.001, 'random_skip': True, 'max_episode_steps': 200}
-env = TLNRasterizedRouteRandomAgentLocation(**env_settings)
-
-tempdir = tempfile.TemporaryDirectory(prefix="quickstart")
-tempdir_path = pathlib.Path(tempdir.name)
-logger.configure(tempdir_path / "GAIL/")
-print(f"All Tensorboards and logging are being written inside {tempdir_path}/.")
-
-expert_batch_size = 4096
-
-venv = DummyVecEnv([lambda: env])
-discriminator = adversarial.GAIL(
-    expert_data=transitions,
-    expert_batch_size=expert_batch_size,
-    #discrim_kwargs={'discrim_net': NoisyDiscriminator(venv, std=0.25)},
-    disc_opt_cls=torch.optim.RMSprop,
-    disc_opt_kwargs={'lr': 0.0001, 'weight_decay': 0.003},
-    discrim_kwargs={'discrim_net': CnnDiscriminator(venv)},
-    venv=venv, # unused
-    gen_algo=stable_baselines3.PPO("CnnPolicy", venv), # unused
-)
-
-expert_data_loader = torch.utils.data.DataLoader(
-    transitions,
-    batch_size=expert_batch_size,
-    collate_fn=types.transitions_collate_fn,
-    shuffle=True,
-    drop_last=True,
-)
-
-gen_data_loader = torch.utils.data.DataLoader(
-    transitions,
-    batch_size=expert_batch_size,
-    collate_fn=types.transitions_collate_fn,
-    shuffle=True,
-    drop_last=True,
-)
-
-# %%
-epochs = 1000
-for i in tqdm(range(epochs)):
-    for expert_samples, gen_samples in zip(expert_data_loader, gen_data_loader):
-        # randomly corrupt actions
-        gen_samples['acts'] = -1 + 2 * torch.rand(*gen_samples['acts'].shape)
-
-        discriminator.train_disc(expert_samples=expert_samples, gen_samples=gen_samples)
-
-    torch.save(discriminator.discrim_net.state_dict(), 'train_discrim.pt')

scratch/etienne/intersimple/trpo_speed_lidar.py

@@ -1,52 +0,0 @@
-# %%
-from sb3_contrib import TRPO
-from intersim.envs import IntersimpleLidarFlat
-from intersim.envs.intersimple import speed_reward
-import functools
-
-model_name = "trpo_speed_lidar"
-
-#def reward(state, action, info):
-#    speed = state[2].item()
-#    r = speed if speed < 10 else (10 - 5 * (speed - 10))
-#    return 0.1 * r
-
-env = IntersimpleLidarFlat(
-    n_rays=5,
-    agent=51,
-    reward=functools.partial(
-        speed_reward,
-        collision_penalty=0
-    ),
-)
-
-# %%
-model = TRPO(
-    "MlpPolicy", env,
-    learning_rate=1e-4,
-    verbose=1,
-    tensorboard_log='runs/',
-    #use_sde=True,
-    #sde_sample_freq=4,
-)
-model.learn(total_timesteps=1000000)
-model.save(model_name)
-
-print('Done training.')
-
-del model # remove to demonstrate saving and loading
-
-# %%
-model = TRPO.load(model_name)
-
-obs = env.reset()
-while True:
-    action, _states = model.predict(obs)
-    obs, rewards, done, info = env.step(action)
-    env.render(mode='post')
-    if done:
-        break
-
-env.close(filestr='render/'+model_name)
-
-# %%

scratch/etienne/pillbox/intersim_demos.ipynb

@@ -1,249 +0,0 @@
-{
- "cells": [
-  {
-   "cell_type": "code",
-   "execution_count": 2,
-   "source": [
-    "%cd learners"
-   ],
-   "outputs": [
-    {
-     "output_type": "stream",
-     "name": "stdout",
-     "text": [
-      "/home/buehrle/dev/InteractionImitation/scratch/etienne/pillbox/learners\n"
-     ]
-    }
-   ],
-   "metadata": {}
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 3,
-   "source": [
-    "%load_ext autoreload\n",
-    "%autoreload 2"
-   ],
-   "outputs": [],
-   "metadata": {}
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 1,
-   "source": [
-    "import torch\n",
-    "import numpy as np"
-   ],
-   "outputs": [],
-   "metadata": {}
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 9,
-   "source": [
-    "# save expert demos to ../experts/Intersim/demos.npz\n",
-    "# make sure to split different experts up\n",
-    "\n",
-    "from intersim.envs.simulator import InteractionSimulator\n",
-    "from intersim.utils import get_map_path, get_svt, SVT_to_stateactions\n",
-    "import gym\n",
-    "from tqdm import tqdm\n",
-    "\n",
-    "def pillbox_demo(observations, actions, rewards):\n",
-    "    demo = {\n",
-    "        'env': 'intersim:intersim-v0',\n",
-    "        'num_trajs': len(observations),\n",
-    "        'mean_reward': rewards.mean(),\n",
-    "        'std_reward': rewards.std(),\n",
-    "    }\n",
-    "    demo.update({\n",
-    "        str(i): {\n",
-    "            'states': o,\n",
-    "            'actions': a,\n",
-    "        } for i, (o, a) in enumerate(zip(observations, actions))\n",
-    "    })\n",
-    "    return demo\n",
-    "\n",
-    "def intersim_expert_demos(loc, track):\n",
-    "    svt, svt_path = get_svt(loc, track)\n",
-    "    osm = get_map_path(loc)\n",
-    "    \n",
-    "    n_actors = svt.simstate.size(1)\n",
-    "    observations = []\n",
-    "    actions = [] ##\n",
-    "    #states, actions = SVT_to_stateactions(svt) ##\n",
-    "    rewards = []\n",
-    "    \n",
-    "    print('Simulating')\n",
-    "    env = gym.make('intersim:intersim-v0', svt=svt, map_path=osm)\n",
-    "    obs, info = env.reset()\n",
-    "    for s in tqdm(svt.simstate[1:]): ##\n",
-    "    #for a in actions: ##\n",
-    "        relative_state = torch.stack((\n",
-    "            obs['relative_state'][..., 0],\n",
-    "            obs['relative_state'][..., 1],\n",
-    "            (obs['relative_state'][..., 2]**2 + obs['relative_state'][..., 3]**2).sqrt(),\n",
-    "            obs['relative_state'][..., 4],\n",
-    "            obs['relative_state'][..., 5],\n",
-    "        ), -1)\n",
-    "        observations.append(torch.cat((\n",
-    "            obs['state'].unsqueeze(1),\n",
-    "            relative_state,\n",
-    "        ), 1))\n",
-    "        obs, r, done, info = env.step(env.target_state(s, mu=.01))\n",
-    "        #obs, r, done, info = env.step(a) ##\n",
-    "        actions.append(info['action_taken'])\n",
-    "        rewards.append(r)\n",
-    "        assert not done, 'Episode terminated during expert demonstration.'\n",
-    "\n",
-    "    _except_idx = lambda o, i: torch.cat((o[:i], o[i+1:]))\n",
-    "        \n",
-    "    # transpose to per-agent observations and actions\n",
-    "    print('Transposing')\n",
-    "    observations = [torch.stack([_except_idx(o[i], i+1) for o in observations]) for i in range(n_actors)]\n",
-    "    actions = [torch.stack([a[i] for a in actions]) for i in range(n_actors)]\n",
-    "    \n",
-    "    print('Trimming')\n",
-    "    # trim observations and actions to start/end of trajectory\n",
-    "    _alive = lambda o: (~o.isnan().all(2).all(1)).nonzero()\n",
-    "    _start = lambda o: _alive(o).min()\n",
-    "    _end = lambda o: _alive(o).max() + 1\n",
-    "    start_end = [(_start(obs), _end(obs)) for obs in observations]\n",
-    "    observations = [obs[start:end] for obs, (start, end) in zip(observations, start_end)]\n",
-    "    actions = [act[start:end] for act, (start, end) in zip(actions, start_end)]\n",
-    "    \n",
-    "    #print('Cropping')\n",
-    "    ## crop observations to max number of observations\n",
-    "    #max_num_obs = max([(~obs.isnan().all(2)).sum(1).max() for obs in observations])\n",
-    "    #observations = [obs[:, :max_num_obs] for obs in observations]\n",
-    "    \n",
-    "    observations = [o.numpy() for o in observations]\n",
-    "    actions = [a.numpy() for a in actions]\n",
-    "    rewards = np.array(rewards)\n",
-    "    \n",
-    "    return pillbox_demo(observations, actions, rewards)"
-   ],
-   "outputs": [],
-   "metadata": {}
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 10,
-   "source": [
-    "demos = intersim_expert_demos(loc=0, track=0)"
-   ],
-   "outputs": [
-    {
-     "output_type": "stream",
-     "name": "stdout",
-     "text": [
-      "Simulating\n",
-      "Custom Vehicle Trajectory Paths\n",
-      "Map Path: datasets/maps/DR_USA_Roundabout_FT.osm\n",
-      "Environment Reset\n"
-     ]
-    },
-    {
-     "output_type": "stream",
-     "name": "stderr",
-     "text": [
-      "100%|███████████████████████████████████████████████████████████████████████████████████████████████████████| 3006/3006 [01:17<00:00, 38.87it/s]\n"
-     ]
-    },
-    {
-     "output_type": "stream",
-     "name": "stdout",
-     "text": [
-      "Transposing\n",
-      "Trimming\n"
-     ]
-    }
-   ],
-   "metadata": {
-    "scrolled": true,
-    "tags": [
-     "outputPrepend"
-    ]
-   }
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 6,
-   "source": [
-    "demos['num_trajs']"
-   ],
-   "outputs": [
-    {
-     "output_type": "execute_result",
-     "data": {
-      "text/plain": [
-       "151"
-      ]
-     },
-     "metadata": {},
-     "execution_count": 6
-    }
-   ],
-   "metadata": {}
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 7,
-   "source": [
-    "demos['25']['states'].shape"
-   ],
-   "outputs": [
-    {
-     "output_type": "execute_result",
-     "data": {
-      "text/plain": [
-       "(71, 151, 5)"
-      ]
-     },
-     "metadata": {},
-     "execution_count": 7
-    }
-   ],
-   "metadata": {}
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 8,
-   "source": [
-    "np.savez('../experts/intersim:intersim-v0/demos.npz', **demos)"
-   ],
-   "outputs": [],
-   "metadata": {}
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "source": [],
-   "outputs": [],
-   "metadata": {}
-  }
- ],
- "metadata": {
-  "kernelspec": {
-   "name": "python3",
-   "display_name": "Python 3.7.5 64-bit ('.venv': venv)"
-  },
-  "language_info": {
-   "codemirror_mode": {
-    "name": "ipython",
-    "version": 3
-   },
-   "file_extension": ".py",
-   "mimetype": "text/x-python",
-   "name": "python",
-   "nbconvert_exporter": "python",
-   "pygments_lexer": "ipython3",
-   "version": "3.7.5"
-  },
-  "interpreter": {
-   "hash": "56465d2ea10f338edb3d30adb010c5849fd826fffc543ba31360f3db8b47a703"
-  }
- },
- "nbformat": 4,
- "nbformat_minor": 4
-}

scratch/etienne/pillbox/intersim_expert.ipynb

@@ -1,168 +0,0 @@
-{
- "cells": [
-  {
-   "cell_type": "code",
-   "execution_count": 2,
-   "source": [
-    "%cd learners"
-   ],
-   "outputs": [
-    {
-     "output_type": "stream",
-     "name": "stdout",
-     "text": [
-      "[Errno 2] No such file or directory: 'learners'\n",
-      "/home/buehrle/dev/InteractionImitation/scratch/etienne/pillbox/learners\n"
-     ]
-    }
-   ],
-   "metadata": {}
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "source": [
-    "import gym\n",
-    "from tqdm import tqdm\n",
-    "\n",
-    "def rollout(pi, max_steps=1000):\n",
-    "    env = gym.make('intersim:intersim-v0')\n",
-    "    env.reset() # obs = env.reset()\n",
-    "    obs, _, done, _ = env.step(0 * env.action_space.sample())\n",
-    "    \n",
-    "    _except = lambda o, i: torch.cat((o[:i], o[i+1:]))\n",
-    "    \n",
-    "    _relative_state_v = lambda obs: torch.stack((\n",
-    "        obs[..., 0],\n",
-    "        obs[..., 1],\n",
-    "        (obs[..., 2]**2 + obs[..., 3]**2).sqrt(),\n",
-    "        obs[..., 4],\n",
-    "        obs[..., 5],\n",
-    "    ), -1)\n",
-    "    \n",
-    "    for _ in tqdm(range(max_steps)):\n",
-    "        pi_obs = [\n",
-    "            torch.cat((e.unsqueeze(0), _relative_state_v(_except(o, i)))).unsqueeze(0)\n",
-    "            for i, (e, o) in enumerate(zip(obs['state'], obs['relative_state']))\n",
-    "        ]\n",
-    "        \n",
-    "        actions = [pi(o).squeeze() for o in pi_obs]\n",
-    "        actions = torch.stack(actions).unsqueeze(1)\n",
-    "        obs, _, done, _ = env.step(actions)\n",
-    "        env.render(mode='post')\n",
-    "        if done:\n",
-    "            break\n",
-    "    env.close()"
-   ],
-   "outputs": [],
-   "metadata": {}
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 4,
-   "source": [
-    "import torch\n",
-    "import numpy as np\n",
-    "\n",
-    "def expert(obs):\n",
-    "    ego = obs[:, 0]\n",
-    "    rel = obs[:, 1:]\n",
-    "    front = torch.stack((torch.cos(ego[:, 3]), torch.sin(ego[:, 3])), -1)\n",
-    "    left = torch.stack((-torch.sin(ego[:, 3]), torch.cos(ego[:, 3])), -1)\n",
-    "    df = (rel[:, :, :2] * front.unsqueeze(1)).sum(-1)\n",
-    "    dl = (rel[:, :, :2] * left.unsqueeze(1)).sum(-1)\n",
-    "\n",
-    "    df = torch.where(df.isnan(), np.inf * torch.ones_like(df), df)\n",
-    "    dl = torch.where(dl.isnan(), np.inf * torch.ones_like(dl), dl)\n",
-    "    rel = torch.where(rel.isnan(), np.inf * torch.ones_like(rel), rel)\n",
-    "\n",
-    "    # relative speed in direction of position difference vector\n",
-    "    vrel = rel[:, :, 2] * (rel[:, :, :2] * torch.stack((\n",
-    "        torch.cos(ego[:, 3].unsqueeze(1) + rel[:, :, 3]),\n",
-    "        torch.sin(ego[:, 3].unsqueeze(1) + rel[:, :, 3])),\n",
-    "        -1)).sum(-1)\n",
-    "    vrel = torch.where(vrel.isnan(), np.inf * torch.ones_like(vrel), vrel)\n",
-    "    vrel = torch.maximum(vrel, torch.zeros_like(vrel))\n",
-    "    \n",
-    "    alpha = torch.atan2(dl, df)\n",
-    "    d = (rel[:, :, :2] ** 2).sum(-1)\n",
-    "    attn = torch.exp(-torch.where(alpha > 0, 0.8*alpha, 1*alpha)**2 - 0.01 * d - 0.1*vrel) \n",
-    "    \n",
-    "    act = 10 - ego[:, 2] - 20 * attn.sum(-1)\n",
-    "    \n",
-    "    return act"
-   ],
-   "outputs": [],
-   "metadata": {}
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 5,
-   "source": [
-    "rollout(expert, max_steps=500)"
-   ],
-   "outputs": [
-    {
-     "output_type": "stream",
-     "name": "stdout",
-     "text": [
-      "Vehicle Trajectory Paths: /home/buehrle/dev/InteractionImitation/InteractionSimulator/datasets/trackfiles/DR_USA_Roundabout_FT/vehicle_tracks_000.csv\n",
-      "Map Path: /home/buehrle/dev/InteractionImitation/InteractionSimulator/datasets/maps/DR_USA_Roundabout_FT.osm\n",
-      "Environment Reset\n"
-     ]
-    },
-    {
-     "output_type": "stream",
-     "name": "stderr",
-     "text": [
-      "  0%|          | 0/500 [00:00<?, ?it/s]\n"
-     ]
-    },
-    {
-     "output_type": "error",
-     "ename": "RuntimeError",
-     "evalue": "torch.cat(): Sizes of tensors must match except in dimension 0. Got 5 and 6 in dimension 1 (The offending index is 1)",
-     "traceback": [
-      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
-      "\u001b[0;31mRuntimeError\u001b[0m                              Traceback (most recent call last)",
-      "\u001b[0;32m/tmp/ipykernel_4266/633354333.py\u001b[0m in \u001b[0;36m<module>\u001b[0;34m\u001b[0m\n\u001b[0;32m----> 1\u001b[0;31m \u001b[0mrollout\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mexpert\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mmax_steps\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;36m500\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m",
-      "\u001b[0;32m/tmp/ipykernel_4266/1839273282.py\u001b[0m in \u001b[0;36mrollout\u001b[0;34m(pi, max_steps)\u001b[0m\n\u001b[1;32m     12\u001b[0m         pi_obs = [\n\u001b[1;32m     13\u001b[0m             \u001b[0mtorch\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mcat\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0me\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0munsqueeze\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;36m0\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0m_except_self\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mo\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mi\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0munsqueeze\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;36m0\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 14\u001b[0;31m             \u001b[0;32mfor\u001b[0m \u001b[0mi\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m(\u001b[0m\u001b[0me\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mo\u001b[0m\u001b[0;34m)\u001b[0m \u001b[0;32min\u001b[0m \u001b[0menumerate\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mzip\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mobs\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0;34m'state'\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mobs\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0;34m'relative_state'\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     15\u001b[0m         ]\n\u001b[1;32m     16\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n",
-      "\u001b[0;32m/tmp/ipykernel_4266/1839273282.py\u001b[0m in \u001b[0;36m<listcomp>\u001b[0;34m(.0)\u001b[0m\n\u001b[1;32m     12\u001b[0m         pi_obs = [\n\u001b[1;32m     13\u001b[0m             \u001b[0mtorch\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mcat\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0me\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0munsqueeze\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;36m0\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0m_except_self\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mo\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mi\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0munsqueeze\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;36m0\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 14\u001b[0;31m             \u001b[0;32mfor\u001b[0m \u001b[0mi\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m(\u001b[0m\u001b[0me\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mo\u001b[0m\u001b[0;34m)\u001b[0m \u001b[0;32min\u001b[0m \u001b[0menumerate\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mzip\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mobs\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0;34m'state'\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mobs\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0;34m'relative_state'\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     15\u001b[0m         ]\n\u001b[1;32m     16\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n",
-      "\u001b[0;31mRuntimeError\u001b[0m: torch.cat(): Sizes of tensors must match except in dimension 0. Got 5 and 6 in dimension 1 (The offending index is 1)"
-     ]
-    }
-   ],
-   "metadata": {}
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "source": [],
-   "outputs": [],
-   "metadata": {}
-  }
- ],
- "metadata": {
-  "kernelspec": {
-   "name": "python3",
-   "display_name": "Python 3.7.5 64-bit ('.venv': venv)"
-  },
-  "language_info": {
-   "codemirror_mode": {
-    "name": "ipython",
-    "version": 3
-   },
-   "file_extension": ".py",
-   "mimetype": "text/x-python",
-   "name": "python",
-   "nbconvert_exporter": "python",
-   "pygments_lexer": "ipython3",
-   "version": "3.7.5"
-  },
-  "interpreter": {
-   "hash": "56465d2ea10f338edb3d30adb010c5849fd826fffc543ba31360f3db8b47a703"
-  }
- },
- "nbformat": 4,
- "nbformat_minor": 5
-}

scratch/etienne/pillbox/learners/adril.py

@@ -1,140 +0,0 @@
-import gym
-from gym import spaces
-from sklearn.neighbors import KDTree
-from scipy.stats import norm
-import numpy as np
-import warnings
-from abc import ABC, abstractmethod
-from typing import Dict, Generator, Optional, Union
-import torch as th
-
-try:
-    # Check memory used by replay buffer when possible
-    import psutil
-except ImportError:
-    psutil = None
-
-from stable_baselines3.common.preprocessing import get_action_dim, get_obs_shape
-from stable_baselines3.common.type_aliases import ReplayBufferSamples, RolloutBufferSamples
-from stable_baselines3.common.vec_env import VecNormalize
-from stable_baselines3.common.buffers import ReplayBuffer
-
-
-class AdRILWrapper(gym.Env):
-  metadata = {'render.modes': ['human']}
-
-  def __init__(self, base_env):
-    super(AdRILWrapper, self).__init__()
-    self.base_env = base_env
-    self.iter = 0
-    self.observation_space = self.base_env.observation_space
-    self.action_space = self.base_env.action_space
-    self.trajs = list()
-    self.num_trajs = 0
-    self.curr_state = None
-  def step(self, action):
-    next_obs, _, done, info = self.base_env.step(action)
-    reward = self.iter # Transformed by replay buffer
-    self.trajs.append((self.curr_state, action, next_obs, done))
-    if done:
-        self.num_trajs += 1
-    self.curr_state = next_obs
-    return next_obs, reward, done, info
-  def reset(self):
-    obs = self.base_env.reset()
-    self.curr_state = obs
-    return obs
-  def render(self, mode='human'):
-    self.base_env.render(mode=mode)
-  def close (self):
-    self.base_env.close()
-  def get_learner_trajs(self):
-    return self.trajs
-  def set_iter(self, k):
-    self.iter = k
-    
-class AdRILReplayBuffer(ReplayBuffer):
-    def __init__(
-        self,
-        buffer_size: int,
-        observation_space: spaces.Space,
-        action_space: spaces.Space,
-        device: Union[th.device, str] = "cpu",
-        n_envs: int = 1,
-        optimize_memory_usage: bool = False,
-        expert_data: dict = dict(),
-            N_expert: int = 0,
-        balanced: bool = True,
-    ):
-        super(AdRILReplayBuffer, self).__init__(buffer_size, observation_space, action_space, device, n_envs=n_envs, optimize_memory_usage=optimize_memory_usage)
-
-        self.expert_states = expert_data['obs']
-        self.expert_actions = expert_data['acts']
-        self.expert_next_states = expert_data['next_obs']
-        self.expert_dones = expert_data['dones']
-        n_expert = len(expert_data["obs"])
-        self.iter = 0
-        self.N_expert = N_expert
-        self.N_learner = 0
-        self.normalizer = 1
-        self.balanced = balanced
-
-    def set_iter(self, k):
-      self.iter = k
-      normalizer = 0
-      for i in range(0, k):
-        normalizer += 1 ** (-i) # written to support decaying learning rate
-      self.normalizer = normalizer
-
-    def set_n_learner(self, n):
-        self.N_learner = n
-
-    def _get_samples(self, batch_inds: np.ndarray, env: Optional[VecNormalize] = None) -> ReplayBufferSamples:
-        num_samples = len(batch_inds)
-        if self.balanced:
-          num_expert_samples = int(num_samples / 2)
-          batch_inds = batch_inds[:num_expert_samples]
-          expert_inds = np.random.randint(0, len(self.expert_states), size=num_expert_samples)
-          # balanced sampling
-          if self.optimize_memory_usage:
-              next_obs = self._normalize_obs(self.observations[(batch_inds + 1) % self.buffer_size, 0, :], env)
-          else:
-              next_obs = self._normalize_obs(self.next_observations[batch_inds, 0, :], env)
-          next_obs = np.concatenate((next_obs, self._normalize_obs(self.expert_next_states[expert_inds], env)), axis=0)
-          obs = self._normalize_obs(self.observations[batch_inds, 0, :], env)
-          obs = np.concatenate((obs, self._normalize_obs(self.expert_states[expert_inds], env)), axis=0)
-          actions = self.actions[batch_inds, 0, :]
-          actions = np.concatenate((actions, self.expert_actions[expert_inds].reshape(num_expert_samples, -1)), axis=0)
-          dones = self.dones[batch_inds]
-          dones = np.concatenate((dones, self.expert_dones[expert_inds].reshape(num_expert_samples, -1)), axis=0)
-          # AdRIL Rewards (indicator kernel)
-          mask1 = (self.rewards[batch_inds] >= 0).astype(np.float32)
-          mask2 = (self.rewards[batch_inds] < self.iter).astype(np.float32)
-          r1 = - (1. ** (-self.rewards[batch_inds])) * mask1 * mask2 # Past iter
-          r2 = np.zeros_like(self.rewards[batch_inds]) * mask1 * (1 - mask2) # current iter
-          r3 = -self.rewards[batch_inds] * (1 - mask1) # Expert
-          if self.iter > 0:
-              rewards = (r1 / self.N_learner) + r2 + r3
-          else:
-              rewards = r1 + r2 + r3
-          rewards = np.concatenate((rewards, np.ones_like(rewards) / self.N_expert), axis=0)
-        else:
-          if self.optimize_memory_usage:
-              next_obs = self._normalize_obs(self.observations[(batch_inds + 1) % self.buffer_size, 0, :], env)
-          else:
-              next_obs = self._normalize_obs(self.next_observations[batch_inds, 0, :], env)
-          obs = self._normalize_obs(self.observations[batch_inds, 0, :], env)
-          actions = self.actions[batch_inds, 0, :]
-          dones = self.dones[batch_inds]
-          # AdRIL Rewards (indicator kernel)
-          mask1 = (self.rewards[batch_inds] >= 0).astype(np.float32)
-          mask2 = (self.rewards[batch_inds] < self.iter).astype(np.float32)
-          r1 = - (1. ** (-self.rewards[batch_inds])) * mask1 * mask2 # Past iter
-          r2 = np.zeros_like(self.rewards[batch_inds]) * mask1 * (1 - mask2) # current iter
-          r3 = -self.rewards[batch_inds] * (1 - mask1) / self.N_expert # Expert
-          if self.iter > 0:
-              rewards = (r1 * 1. / self.N_learner) + r2 + r3
-          else:
-              rewards = r1 + r2 + r3
-        data = (obs, actions, next_obs, dones, rewards)
-        return ReplayBufferSamples(*tuple(map(self.to_torch, data)))

scratch/etienne/pillbox/learners/advil.py

@@ -1,222 +0,0 @@
-import numpy as np
-
-import torch
-import torch.autograd as autograd
-import torch.nn as nn
-import torch.nn.functional as F
-import torch.optim as optim
-from gym.spaces import Discrete
-import gym
-from stable_baselines3.common.preprocessing import get_action_dim
-from tqdm import tqdm
-from torch.autograd import Variable
-from itertools import repeat
-from torch.autograd import grad as torch_grad
-from typing import List, Type
-import types
-
-# Infinite dataloader
-def repeater(data_loader):
-    for loader in repeat(data_loader):
-        for data in loader:
-            yield data
-
-def create_mlp(
-    input_dim: int, output_dim: int, net_arch: List[int], activation_fn: Type[nn.Module] = nn.ReLU) -> List[nn.Module]:
-
-    if len(net_arch) > 0:
-        modules = [nn.Linear(input_dim, net_arch[0]), activation_fn()]
-    else:
-        modules = []
-
-    for idx in range(len(net_arch) - 1):
-        modules.append(nn.Linear(net_arch[idx], net_arch[idx + 1]))
-        modules.append(activation_fn())
-
-    if output_dim > 0:
-        last_layer_dim = net_arch[-1] if len(net_arch) > 0 else input_dim
-        modules.append(nn.Linear(last_layer_dim, output_dim))
-    return modules
-
-def init_ortho(layer):
-    if type(layer) == nn.Linear:
-        nn.init.orthogonal_(layer.weight)
-
-        
-class AdVILPolicy(nn.Module):
-    def __init__(self, env, mean=None, std=None):
-        super(AdVILPolicy, self).__init__()
-        if isinstance(env.action_space, Discrete):
-            self.net_arch = [64, 64]
-            self.action_dim = env.action_space.n
-            self.discrete = True
-        else:
-            self.net_arch = [256, 256]
-            self.action_dim = int(np.prod(env.action_space.shape))
-            self.low = torch.as_tensor(env.action_space.low)
-            self.high = torch.as_tensor(env.action_space.high)
-            self.discrete = False
-        self.obs_dim = int(np.prod(env.observation_space.shape))
-        self.observation_space = env.observation_space
-        net = create_mlp(self.obs_dim, self.action_dim, self.net_arch, nn.ReLU)
-        if self.discrete:
-            net.append(nn.Softmax(dim=1))
-        self.net = nn.Sequential(*net)
-        self.net.apply(init_ortho)
-        if mean is not None and std is not None:
-            self.mean = mean
-            self.std = std
-            self.is_normalized = True
-        else:
-            self.is_normalized = False
-    def forward(self, obs):
-        action = self.net(obs)
-        return action
-    def predict(self, obs, state, mask, deterministic):
-        obs = obs.reshape((-1,) + (self.obs_dim,))
-        if self.is_normalized:
-            obs = (obs - self.mean) / self.std
-        obs = torch.as_tensor(obs)
-        with torch.no_grad():
-            actions = self.forward(obs)
-            if self.discrete:
-                actions = actions.argmax(dim=1).reshape(-1)
-            else:
-                actions = self.low + ((actions + 1.0) / 2.0) * (self.high - self.low)
-                actions = torch.max(torch.min(actions, self.high), self.low)
-            actions = actions.cpu().numpy()
-        return actions, state
-
-
-class AdVILDiscriminator(nn.Module):
-    def __init__(self, env):
-        super(AdVILDiscriminator, self).__init__()
-        if isinstance(env.action_space, Discrete):
-            self.net_arch = [64, 64]
-            self.action_dim = env.action_space.n
-        else:
-            self.net_arch = [256, 256]
-            self.action_dim = int(np.prod(env.action_space.shape))
-        self.obs_dim = int(np.prod(env.observation_space.shape))
-        net = create_mlp(self.obs_dim + self.action_dim, 1, self.net_arch, nn.ReLU)
-        self.net = nn.Sequential(*net)
-        self.net.apply(init_ortho)
-
-    def forward(self, inputs):
-        output = self.net(inputs)
-        return output.view(-1)
-
-def pi_update(obs, acts, pi, f, pi_opt, prog):
-    pi_opt.zero_grad()
-    obs_v = Variable(obs)
-    pi_acts = pi(obs_v)
-    #learner_sa = torch.cat((obs, pi_acts), axis=1)
-    f_learner = f(obs, acts)
-    pi_loss = f_learner.mean() + orthogonal_reg(pi) + 2e-1 * (pi_acts - acts).square().mean()
-    pi_loss.backward()
-    if prog > 0.1:
-        torch.nn.utils.clip_grad_norm(pi.parameters(), 40.0)
-    pi_opt.step()
-    return pi_loss.item(), (2e-1 * (pi_acts - acts).square().mean()).item()
-
-def orthogonal_reg(pi):
-    with torch.enable_grad():
-        reg = 1e-4
-        orth_loss = torch.zeros(1)
-        for name, param in pi.named_parameters():
-            if 'bias' not in name:
-                x = torch.mm(torch.t(param), param)
-                x = x * (1. - torch.eye(param.shape[-1]))
-                orth_loss = orth_loss + reg * (x.square().sum())
-        return orth_loss
-
-def f_update(obs, acts, pi, f, f_opt, prog):
-    obs_v = Variable(obs)
-    pi_acts = pi(obs_v)
-    #learner_sa = torch.cat((obs, pi_acts), axis=1)
-    #expert_sa = Variable(torch.cat((obs, acts), axis=1))
-    f_learner = f(obs, pi_acts)
-    f_expert = f(obs, acts)
-    #gp = gradient_penalty((obs, pi_acts), (obs, acts), f)
-    f_opt.zero_grad()
-    f_loss = f_expert.mean() - f_learner.mean()# + 10 * gp
-    f_loss.backward()
-    if prog > 0.1:
-        torch.nn.utils.clip_grad_norm(f.parameters(), 40.0)
-    f_opt.step()
-    return f_loss.item()
-
-def gradient_penalty(learner_sa, expert_sa, f):
-    batch_size = expert_sa[0].size()[0]
-
-    #alpha = torch.rand(batch_size, 1)
-    #alpha = alpha.expand_as(expert_sa)
-
-    salpha = torch.rand(batch_size, 1, 1)
-    salpha = salpha.expand_as(expert_sa[0])
-
-    aalpha = torch.rand(batch_size, 1)
-    aalpha = aalpha.expand_as(expert_sa[1])
-
-    #interpolated = alpha * expert_sa.data + (1 - alpha) * learner_sa.data
-    #interpolated = Variable(interpolated, requires_grad=True)
-    #f_interpolated = f(interpolated.float())
-
-    sinterpolated = salpha * expert_sa[0].data + (1 - salpha) * learner_sa[0].data
-    sinterpolated = Variable(sinterpolated, requires_grad=True)
-
-    ainterpolated = aalpha * expert_sa[1].data + (1 - aalpha) * learner_sa[1].data
-    ainterpolated = Variable(ainterpolated, requires_grad=True)
-    
-    f_interpolated = f(sinterpolated, ainterpolated)
-
-    #gradients = torch_grad(outputs=f_interpolated, inputs=interpolated,
-    #                       grad_outputs=torch.ones(f_interpolated.size()),
-    #                       create_graph=True, retain_graph=True)[0]
-    
-    sgradients = torch_grad(outputs=f_interpolated, inputs=sinterpolated,
-                           grad_outputs=torch.ones(f_interpolated.size()),
-                           create_graph=True, retain_graph=True)[0]
-
-    agradients = torch_grad(outputs=f_interpolated, inputs=ainterpolated,
-                           grad_outputs=torch.ones(f_interpolated.size()),
-                           create_graph=True, retain_graph=True)[0]
-
-    #gradients = gradients.view(batch_size, -1)
-    sgradients = sgradients.view(batch_size, -1)
-    agradients = agradients.view(batch_size, -1)
-    #norm = gradients.norm(2, dim=1).mean().item()
-    #gradients_norm = torch.sqrt(torch.sum(gradients ** 2, dim=1) + 1e-12)
-    gradients_norm = torch.sqrt(torch.sum(sgradients ** 2, dim=1) + torch.sum(agradients ** 2, dim=1) + 1e-12)
-    # 2 * |f'(x_0)|
-    return ((gradients_norm - 0.4) ** 2).mean()
-
-def advil_training(data_loader, env, iters=int(1e5), policy_class=AdVILPolicy, discriminator_class=AdVILDiscriminator, lr_pi=8e-6, lr_f=8e-4):
-    if not isinstance(env.action_space, Discrete):
-        low = torch.as_tensor(env.action_space.low)
-        high = torch.as_tensor(env.action_space.high)
-    if data_loader.dataset.is_normalized:
-        pi = policy_class(env, data_loader.dataset.mean, data_loader.dataset.std)
-    else:
-        pi = policy_class(env)
-    f = discriminator_class(env)
-    pi_opt = optim.Adam(pi.parameters(), lr=lr_pi)
-
-    last_loss = 0
-    f_opt = optim.Adam(f.parameters(), lr=lr_f)
-    data_loader = repeater(data_loader)
-    for t in tqdm(range(iters)):
-        data = next(data_loader)
-        obs = data['obs']
-        acts = data['acts']
-        #if isinstance(env.action_space, Discrete):
-        #    acts = nn.functional.one_hot(acts, env.action_space.n)
-        #else:
-        #    acts = (((acts - low) / (high - low)) * 2.0) - 1.0
-        pi_loss, mse_reg = pi_update(obs, acts, pi, f, pi_opt, t/iters)
-        f_loss = f_update(obs, acts, pi, f, f_opt, t/iters)
-        if t % 100 == 0:
-            print("pi loss:", pi_loss)
-            print("mse reg:", mse_reg)
-            print("f loss:", f_loss)
-    return pi

scratch/etienne/pillbox/learners/intersim_advil.py

@@ -1,155 +0,0 @@
-import torch
-import torch.nn as nn
-
-def unnormalize(val, mean, std):
-    val *= std or 1
-    val += mean or 0
-    return val
-
-def normalize(val, mean, std):
-    val -= mean or 0
-    val /= std or 1
-    return val
-
-class IntersimPolicy(nn.Module):
-    def __init__(self, env, mean=None, std=None):
-        # assert "intersim" in env.unwrapped.spec.id
-        super().__init__()
-
-        self._ego_encoder = nn.Sequential(
-            # in 5, out 5
-            nn.Linear(5, 10),
-            nn.ReLU(),
-            nn.Linear(10, 10),
-            nn.ReLU(),
-            nn.Linear(10, 10),
-            nn.ReLU(),
-            nn.Linear(10, 5),
-            nn.ReLU(),
-        )
-        self._state_encoder = nn.Sequential(
-            # in 5, out 5
-            nn.Linear(5, 10),
-            nn.ReLU(),
-            nn.Linear(10, 10),
-            nn.ReLU(),
-            nn.Linear(10, 10),
-            nn.ReLU(),
-            nn.Linear(10, 5),
-            nn.ReLU(),
-        )
-        self._deepset = lambda e: e.sum(-2)
-        self._action_decoder = nn.Sequential(
-            # in 5 + 5, out 1
-            nn.Linear(5 + 5, 10),
-            nn.ReLU(),
-            nn.Linear(10, 10),
-            nn.ReLU(),
-            nn.Linear(10, 10),
-            nn.ReLU(),
-            nn.Linear(10, 10),
-            nn.ReLU(),
-            nn.Linear(10, 1),
-        )
-    
-    def forward(self, obs):
-        # obs.shape = (batch=514, 1 + others=150, 5)
-        # act.shape = (batch=514, 1)
-        
-        ego = obs[:, 0]#.detach().clone()
-        rel = obs[:, 1:]#.detach().clone()
-        nan = rel.isnan().any(-1, keepdim=True)
-        rel = torch.where(nan, torch.zeros_like(rel), rel) # required because of https://github.com/pytorch/pytorch/issues/15506
-        
-        d = (rel[:, :, :2] ** 2).sum(-1).sqrt()
-        front = torch.stack((torch.cos(ego[:, 3]), torch.sin(ego[:, 3])), -1)
-        left = torch.stack((-torch.sin(ego[:, 3]), torch.cos(ego[:, 3])), -1)
-        df = (rel[:, :, :2] * front.unsqueeze(1)).sum(-1)
-        dl = (rel[:, :, :2] * left.unsqueeze(1)).sum(-1)
-        alpha = torch.atan2(dl, df)
-        
-        rel[:, :, 0] = d
-        rel[:, :, 1] = alpha
-
-        e = self._ego_encoder(ego)
-        x = self._state_encoder(rel)
-        x = torch.where(nan, torch.zeros_like(x), x)
-        x = self._deepset(x)
-        a = self._action_decoder(torch.cat((e, x), 1))
-        
-        return 10 * a
-
-    def predict(self, state, mask, deterministic):
-        #action_distribution = self.forward(obs)
-        #action = action_distribution.argmax()
-        #return action
-        return self.forward(obs)
-
-class IntersimDiscriminator(nn.Module):
-    def __init__(self, env):
-        # assert "intersim" in env.unwrapped.spec.id
-        super().__init__()
-        
-        self._ego_encoder = nn.Sequential(
-            # in 5, out 5
-            nn.Linear(5, 10),
-            nn.ReLU(),
-            nn.Linear(10, 10),
-            nn.ReLU(),
-            nn.Linear(10, 10),
-            nn.ReLU(),
-            nn.Linear(10, 5),
-            nn.ReLU(),
-        )
-        self._state_encoder = nn.Sequential(
-            # in 5, out 5
-            nn.Linear(5, 10),
-            nn.ReLU(),
-            nn.Linear(10, 10),
-            nn.ReLU(),
-            nn.Linear(10, 10),
-            nn.ReLU(),
-            nn.Linear(10, 5),
-            nn.ReLU(),
-        )
-        self._deepset = lambda e: e.sum(-2)
-        self._discriminator = nn.Sequential(
-            # in 5 + 5 + 1, out 1
-            nn.Linear(5 + 5 + 1, 10),
-            nn.ReLU(),
-            nn.Linear(10, 10),
-            nn.ReLU(),
-            nn.Linear(10, 10),
-            nn.ReLU(),
-            nn.Linear(10, 10),
-            nn.ReLU(),
-            nn.Linear(10, 1),
-        )
-    
-    def forward(self, obs, acts):
-        # obs.shape = (batch=514, 1 + others=150, 5)
-        # acts.shape = (batch=514, 1)
-        # val.shape = (batch=514,)
-
-        ego = obs[:, 0]
-        rel = obs[:, 1:]
-        nan = rel.isnan().any(-1, keepdim=True)
-        rel = torch.where(nan, torch.zeros_like(rel), rel) # required because of https://github.com/pytorch/pytorch/issues/15506
-        
-        d = (rel[:, :, :2] ** 2).sum(-1).sqrt()
-        front = torch.stack((torch.cos(ego[:, 3]), torch.sin(ego[:, 3])), -1)
-        left = torch.stack((-torch.sin(ego[:, 3]), torch.cos(ego[:, 3])), -1)
-        df = (rel[:, :, :2] * front.unsqueeze(1)).sum(-1)
-        dl = (rel[:, :, :2] * left.unsqueeze(1)).sum(-1)
-        alpha = torch.atan2(dl, df)
-        
-        rel[:, :, 0] = d
-        rel[:, :, 1] = alpha
-        
-        e = self._ego_encoder(ego)
-        x = self._state_encoder(rel)
-        x = torch.where(nan, torch.zeros_like(x), x)
-        x = self._deepset(x)
-        v = self._discriminator(torch.cat((e, x, acts), 1))
-        
-        return v.squeeze(1)