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

Optionally save videos of evaluated policies

.gitignore

@@ -1,6 +1,9 @@
+*.png
 *.pkl
 *.pt
 *.zip
+**/ray/*
+**/runs/*
 
 # Byte-compiled / optimized / DLL files
 __pycache__/
@@ -149,3 +152,4 @@ output/
 
 # Dependencies
 InteractionSimulator/
+imitation/

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

@@ -0,0 +1,218 @@
+# %%
+import os
+
+from tqdm import tqdm
+from src.core.sampling import rollout
+from src.core.gail import gail_ppo, Buffer
+from src.core.value import SetValue
+from src.core.policy import SetPolicy
+from src.core.discriminator import DeepsetDiscriminator
+import torch
+
+from intersim.envs import IntersimpleLidarFlatRandom
+from intersim.envs.intersimple import speed_reward
+import functools
+from src.util.wrappers import CollisionPenaltyWrapper, TransformObservation, Setobs
+import numpy as np
+from torch.utils.tensorboard import SummaryWriter
+from ray import tune
+from datetime import datetime
+import json
+
+DIR = os.path.dirname(os.path.abspath(__file__))
+activations = [torch.nn.Tanh, torch.nn.LeakyReLU]
+
+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 training_function(config):
+    np.random.seed(config['seed'])
+    torch.manual_seed(config['seed'])
+
+    # choose validation environment
+    if config['experiment'] == 'A': 
+        envs = [Setobs(TransformObservation(CollisionPenaltyWrapper(
+            IntersimpleLidarFlatRandom(
+                n_rays=5,
+                reward=functools.partial(
+                    speed_reward,
+                    collision_penalty=0
+                ),
+                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)]
+    elif config['experiment'] == 'B': 
+        envs = sum([[Setobs(TransformObservation(CollisionPenaltyWrapper(
+            IntersimpleLidarFlatRandom(
+                n_rays=5,
+                reward=functools.partial(
+                    speed_reward,
+                    collision_penalty=0
+                ),
+                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)],[])
+    else:
+        raise NotImplementedError
+
+    env_fn = lambda i: envs[i]
+
+    # load expert data
+
+    if config['experiment'] == 'A': 
+        expert_data = torch.load(os.path.join(DIR, 'intersimple-expert-data-setobs2-loc0-track0.pt'))
+    elif config['experiment'] == 'B': 
+        expert_data = [
+            torch.load(os.path.join(DIR, 'intersimple-expert-data-setobs2-loc0-track0.pt')),
+            torch.load(os.path.join(DIR, 'intersimple-expert-data-setobs2-loc0-track1.pt')),
+            torch.load(os.path.join(DIR, 'intersimple-expert-data-setobs2-loc0-track2.pt')),
+            torch.load(os.path.join(DIR, 'intersimple-expert-data-setobs2-loc0-track3.pt')),
+        ]
+        d0 = [d[0] for d in expert_data]
+        d1 = [d[1] for d in expert_data]
+        d2 = [d[2] for d in expert_data]
+        d3 = [d[3] for d in expert_data]
+        expert_data = (torch.cat(d0), torch.cat(d1), torch.cat(d2), torch.cat(d3))
+    
+    expert_data = Buffer(*expert_data)
+
+    # configure and train policy
+    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+
+    policy = SetPolicy(expert_data.actions.shape[-1], 
+        n_hidden_layers=config['policy']['n_hidden_layers'],
+        hidden_layer_size=config['policy']['hidden_layer_size'],
+        activation=activations[config['policy']['activation']] ) # config net architecture
+    policy = policy.to(device)
+    
+    pi_opt = torch.optim.Adam(policy.parameters(), lr=config['policy']['learning_rate'])
+    pi_lr_scheduler = torch.optim.lr_scheduler.ExponentialLR(pi_opt, gamma=config['policy']['learning_rate_decay'])
+
+    expert_states = expert_data.states[~expert_data.dones].to(device)
+    expert_actions = expert_data.actions[~expert_data.dones].to(device)
+
+    for epoch in range(config['train_epochs']):
+        pi_opt.zero_grad()
+        loss = -policy.log_prob(policy(expert_states), expert_actions).mean()
+        loss.backward()
+        pi_opt.step()
+        pi_lr_scheduler.step()
+
+        if epoch % 25 == 0:
+            gen_states, gen_actions, gen_rewards, gen_dones, gen_collisions = rollout(env_fn, policy.cpu(), n_episodes=60, max_steps_per_episode=200)
+            gen_mean_episode_length = (~gen_dones).sum() / gen_states.shape[0]
+            gen_mean_reward_per_episode = gen_rewards[~gen_dones].sum() / gen_states.shape[0]
+            gen_collision_rate = (1. * gen_collisions.any(-1)).mean()
+
+            tune.report(
+                gen_mean_reward_per_episode=gen_mean_reward_per_episode.item(),
+                mean_episode_length=gen_mean_episode_length.item(),
+                gen_collision_rate=gen_collision_rate.item(),
+                loss=loss.item(),
+            )
+        
+        # save model checkpoints
+        ep = epoch + 1
+        if (ep % 50 == 0):
+            torch.save(policy.state_dict(), f'policy_epoch{ep}.pt')
+
+    # save model
+    torch.save(policy.state_dict(), 'policy_final.pt')
+
+if __name__ == '__main__':
+    import argparse
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--train', choices=['A', 'B'])
+    parser.add_argument('--epochs', type=int, default=500)
+    parser.add_argument('--test', type=str, help='path to config file to run final training on')
+    parser.add_argument('--test_seeds', type=int, default=5)
+    parser.add_argument('--test_cpus', type=int, help='number of cpus available to split test seed training over')
+    args = parser.parse_args()
+
+    assert (args.train is None) ^ (args.test is None), 'Must either train on an experiment or test with a config file'
+
+    # if no test config specified, train
+    if args.test is None:
+        print('Running Tuning for Experiment %s'%(args.train))
+        analysis = tune.run(
+            training_function,
+            config={
+                'experiment': args.train,
+                'trainenv': {
+                    'stop_on_collision': False, 
+                    'use_idm':True,
+                },
+                'policy': {
+                    'learning_rate': 3e-4, 
+                    'learning_rate_decay': tune.grid_search([0.999, 1.0]), 
+                    'hidden_layer_size': tune.grid_search([10, 20, 40]),
+                    'n_hidden_layers': tune.grid_search([2, 3]), 
+                    'activation':tune.grid_search([0, 1]),
+                },
+                'train_epochs': args.epochs,
+                'seed': 0,
+            }
+            # TODO resources_per_trial={'gpu': 1}
+        )
+        best_config = analysis.get_best_config(metric='gen_collision_rate', mode='min')
+        print('Best config: ', best_config)
+
+        # safe best_config
+        if not os.path.isdir(os.path.join(DIR, 'best_configs')):
+            os.mkdir(os.path.join(DIR, 'best_configs'))
+        
+        # save gail
+        with open(os.path.join(DIR, 'best_configs',f'bc_exp{args.train}.json'), 'w', encoding='utf-8') as f:
+            json.dump(best_config, f, ensure_ascii=False, indent=4)
+
+    # if config file specified, rerun it with appropriate number of seeds
+    else:
+        with open(args.test, 'rb') as f:
+            config = json.load(f)
+
+        print(f'Retraining {args.test} with {args.test_seeds} seeds on experiment {config["experiment"]}')
+
+        # rerun with appropriate number of seeds
+        rpt = {'cpu': int(args.test_cpus/args.test_seeds)} if (args.test_cpus is not None) else None
+        config['seed'] = tune.grid_search(list(range(1,args.test_seeds+1)))
+        analysis = tune.run(training_function, config=config, resources_per_trial=rpt)
+
+        # move final policies to appropriate directory
+        split_ = os.path.basename(args.test).split('_')
+        model = split_[0]
+        exper = split_[-1].split('.')[0]
+        savepath = os.path.join('test_policies',model,exper)
+        
+        if not os.path.isdir(savepath):
+            os.makedirs(savepath)
+        
+        import shutil
+        for i in range(args.test_seeds):
+            s = analysis._checkpoints[i]['config']['seed']
+            check_dir = analysis._checkpoints[i]['logdir']
+            shutil.copyfile(os.path.join(check_dir,'policy_final.pt'), 
+                os.path.join(savepath, f'policy_seed{s}.pt'))
+            shutil.copyfile(os.path.join(check_dir,'params.json'),
+                os.path.join(savepath, 'config.json')) # copy config automatically

best_configs/bc_expA.json

@@ -0,0 +1,16 @@
+{
+    "experiment": "A",
+    "trainenv": {
+        "stop_on_collision": false,
+        "use_idm": true
+    },
+    "policy": {
+        "learning_rate": 0.0003,
+        "learning_rate_decay": 1.0,
+        "hidden_layer_size": 40,
+        "n_hidden_layers": 2,
+        "activation": 0
+    },
+    "train_epochs": 300,
+    "seed": 0
+}

best_configs/bc_expB.json

@@ -0,0 +1,16 @@
+{
+    "experiment": "B",
+    "trainenv": {
+        "stop_on_collision": false,
+        "use_idm": true
+    },
+    "policy": {
+        "learning_rate": 0.0003,
+        "learning_rate_decay": 1.0,
+        "hidden_layer_size": 40,
+        "n_hidden_layers": 2,
+        "activation": 0
+    },
+    "train_epochs": 300,
+    "seed": 0
+}

best_configs/gail_expA.json

@@ -0,0 +1,32 @@
+{
+    "experiment": "A",
+    "trainenv": {
+        "stop_on_collision": false,
+        "use_idm": true
+    },
+    "policy": {
+        "learning_rate": 0.0003,
+        "learning_rate_decay": 1.0,
+        "clip_ratio": 0.2,
+        "iterations_per_epoch": 100,
+        "hidden_layer_size": 40,
+        "n_hidden_layers": 2,
+        "activation": 0
+    },
+    "value": {
+        "learning_rate": 0.0001,
+        "weight_decay": 0.001,
+        "iterations_per_epoch": 1000
+    },
+    "discriminator": {
+        "learning_rate": 0.001,
+        "weight_decay": 0.0001,
+        "iterations_per_epoch": 100,
+        "n_hidden_layers_element": 4,
+        "n_hidden_layers_global": 1,
+        "hidden_layer_size": 10,
+        "activation": 0
+    },
+    "train_epochs": 100,
+    "seed": 0
+}

best_configs/gail_expB.json

@@ -0,0 +1,32 @@
+{
+    "experiment": "B",
+    "trainenv": {
+        "stop_on_collision": false,
+        "use_idm": true
+    },
+    "policy": {
+        "learning_rate": 0.0003,
+        "learning_rate_decay": 1.0,
+        "clip_ratio": 0.2,
+        "iterations_per_epoch": 100,
+        "hidden_layer_size": 40,
+        "n_hidden_layers": 2,
+        "activation": 0
+    },
+    "value": {
+        "learning_rate": 0.0001,
+        "weight_decay": 0.001,
+        "iterations_per_epoch": 1000
+    },
+    "discriminator": {
+        "learning_rate": 0.001,
+        "weight_decay": 0.0001,
+        "iterations_per_epoch": 100,
+        "n_hidden_layers_element": 4,
+        "n_hidden_layers_global": 1,
+        "hidden_layer_size": 10,
+        "activation": 0
+    },
+    "train_epochs": 100,
+    "seed": 0
+}

best_configs/hail_expA.json

@@ -0,0 +1,34 @@
+{
+    "experiment": "A",
+    "trainenv": {
+        "stop_on_collision": false,
+        "safe_actions_collision_method": null,
+        "abort_unsafe_collision_method": null,
+        "use_idm": true
+    },
+    "policy": {
+        "learning_rate": 0.0003,
+        "learning_rate_decay": 1.0,
+        "clip_ratio": 0.2,
+        "iterations_per_epoch": 100,
+        "hidden_layer_size": 40,
+        "n_hidden_layers": 2,
+        "activation": 0,
+        "option": 0
+    },
+    "value": {
+        "learning_rate": 0.001,
+        "iterations_per_epoch": 1000
+    },
+    "discriminator": {
+        "learning_rate": 0.001,
+        "weight_decay": 0.0001,
+        "iterations_per_epoch": 100,
+        "n_hidden_layers_element": 4,
+        "n_hidden_layers_global": 1,
+        "hidden_layer_size": 10,
+        "activation": 0
+    },
+    "train_epochs": 90,
+    "seed": 0
+}

best_configs/hail_expB.json

@@ -0,0 +1,34 @@
+{
+    "experiment": "B",
+    "trainenv": {
+        "stop_on_collision": false,
+        "safe_actions_collision_method": null,
+        "abort_unsafe_collision_method": null,
+        "use_idm": true
+    },
+    "policy": {
+        "learning_rate": 0.0003,
+        "learning_rate_decay": 1.0,
+        "clip_ratio": 0.2,
+        "iterations_per_epoch": 100,
+        "hidden_layer_size": 20,
+        "n_hidden_layers": 2,
+        "activation": 0,
+        "option": 0
+    },
+    "value": {
+        "learning_rate": 0.001,
+        "iterations_per_epoch": 1000
+    },
+    "discriminator": {
+        "learning_rate": 0.001,
+        "weight_decay": 0.0001,
+        "iterations_per_epoch": 100,
+        "n_hidden_layers_element": 4,
+        "n_hidden_layers_global": 2,
+        "hidden_layer_size": 10,
+        "activation": 0
+    },
+    "train_epochs": 85,
+    "seed": 0
+}

best_configs/shail_expA.json

@@ -0,0 +1,34 @@
+{
+    "experiment": "A",
+    "trainenv": {
+        "stop_on_collision": false,
+        "safe_actions_collision_method": "circle",
+        "abort_unsafe_collision_method": "circle",
+        "use_idm": true
+    },
+    "policy": {
+        "learning_rate": 0.0003,
+        "learning_rate_decay": 1.0,
+        "clip_ratio": 0.2,
+        "iterations_per_epoch": 100,
+        "hidden_layer_size": 40,
+        "n_hidden_layers": 2,
+        "activation": 0,
+        "option": 0
+    },
+    "value": {
+        "learning_rate": 0.001,
+        "iterations_per_epoch": 1000
+    },
+    "discriminator": {
+        "learning_rate": 0.001,
+        "weight_decay": 0.0001,
+        "iterations_per_epoch": 100,
+        "n_hidden_layers_element": 4,
+        "n_hidden_layers_global": 1,
+        "hidden_layer_size": 10,
+        "activation": 0
+    },
+    "train_epochs": 90,
+    "seed": 0
+}

best_configs/shail_expB.json

@@ -0,0 +1,34 @@
+{
+    "experiment": "B",
+    "trainenv": {
+        "stop_on_collision": false,
+        "safe_actions_collision_method": "circle",
+        "abort_unsafe_collision_method": "circle",
+        "use_idm": true
+    },
+    "policy": {
+        "learning_rate": 0.0003,
+        "learning_rate_decay": 1.0,
+        "clip_ratio": 0.2,
+        "iterations_per_epoch": 100,
+        "hidden_layer_size": 20,
+        "n_hidden_layers": 2,
+        "activation": 0,
+        "option": 0
+    },
+    "value": {
+        "learning_rate": 0.001,
+        "iterations_per_epoch": 1000
+    },
+    "discriminator": {
+        "learning_rate": 0.001,
+        "weight_decay": 0.0001,
+        "iterations_per_epoch": 100,
+        "n_hidden_layers_element": 4,
+        "n_hidden_layers_global": 2,
+        "hidden_layer_size": 10,
+        "activation": 0
+    },
+    "train_epochs": 85,
+    "seed": 0
+}

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

@@ -0,0 +1,108 @@
+import os
+from src.eval_main import eval_main
+from src.evaluation.utils import load_and_average
+import torch
+import json
+
+activations = [torch.nn.Tanh, torch.nn.LeakyReLU]
+
+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', '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, '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, 'use_idm':True}
+    elif method in ['shail']:
+        env = 'NormalizedSafeOptionsEvalEnv'
+        env_kwargs={'stop_on_collision':True, 'max_episode_steps':1000, 'use_idm':True}
+    else:
+        raise NotImplementedError
+
+    files = ['']
+
+    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'])
+
+        policy_config = {k: v for k, v in config['policy'].items() if k not in exclude_keys_from_policy_kwargs}
+        policy_config['activation'] = activations[policy_config['activation']]
+        print('final policy config', policy_config)
+
+        policy_kwargs.update(policy_config)
+        print('final policy kwargs', policy_kwargs)
+
+    if not skip_running:
+        for policy_file in files:
+            # run metrics on that file
+            outbase = eval_main(locations=locations,
+                                method=method, 
+                                policy_file=policy_file, 
+                                policy_kwargs=policy_kwargs,
+                                env=env, 
+                                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])
+        if folder is None:
+            outfolder = os.path.join('out',method,locstr)
+        else:
+            path_items = folder.split('/')
+            outfolder = os.path.join('out', '/'.join(path_items[1:]), locstr)
+    
+    # load metrics from save_path
+    average_metrics = load_and_average(outfolder)
+    if method in ['expert', 'idm']:
+        latex_print(average_metrics, light=True)
+    else:
+        latex_print(average_metrics)
+
+def latex_print(am, light=False):
+    """
+    print latex line
+
+    am (Dict[str,tuple]): dict mapping metric_name to (mean, std)
+    """
+
+    print('success rate, distance travelled, RWSE_10, |DeltaV|, AccelJSD')
+    if light:
+        if 'rwse_10s' in am.keys():
+            print("%2.1f& %2.1f & %2.1f & %1.2f& "
+            "%0.3f \\\\" %( 100*am['success rate'][0], am['mean travel distance'][0], am['rwse_10s'][0],
+                            am['average absolute average velocity'][0],am['acceleration distribution divergence'][0] ))
+            return
+
+
+        print("%2.1f& %2.1f & $---$ & $---$ & "
+            "$---$ \\\\" %( 100*am['success rate'][0], am['mean travel distance'][0]))
+        return
+    
+    print("%2.1f \\scriptstyle\\pm %2.1f & %2.1f \\scriptstyle\\pm %2.1f & "
+        "%2.1f \\scriptstyle\\pm %1.1f & %1.2f \\scriptstyle\\pm %1.2f & "
+        "%0.3f \\scriptstyle\\pm %0.3f \\\\" %( 100*am['success rate'][0], 100*am['success rate'][1],
+               am['mean travel distance'][0] , am['mean travel distance'][1]  ,
+               am['rwse_10s'][0] , am['rwse_10s'][1]  ,
+               am['average absolute average velocity'][0] , am['average absolute average velocity'][1]  ,
+               am['acceleration distribution divergence'][0] , am['acceleration distribution divergence'][1]  ))
+
+if __name__=='__main__':
+    import fire 
+    fire.Fire(main)

evaluate_models.sh

@@ -0,0 +1,18 @@
+# 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 idm 
+python -m eval_experiments --method bc --folder='test_policies/bc/expA' 
+python -m eval_experiments --method gail --folder='test_policies/gail/expA' 
+python -m eval_experiments --method hail --folder='test_policies/hail/expA'
+python -m eval_experiments --method shail --folder='test_policies/shail/expA' 
+
+# Experiment B
+python -m eval_experiments --locations='[(0,4)]' 
+python -m eval_experiments --method idm --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

@@ -0,0 +1,243 @@
+# %%
+import os
+
+import gym
+from src.core.gail import gail_ppo, Buffer
+from src.core.value import SetValue
+from src.core.policy import SetPolicy
+from src.core.discriminator import DeepsetDiscriminator
+import torch
+
+from intersim.envs import IntersimpleLidarFlatRandom
+from intersim.envs.intersimple import speed_reward
+import functools
+from src.util.wrappers import CollisionPenaltyWrapper, TransformObservation, Setobs
+import numpy as np
+from torch.utils.tensorboard import SummaryWriter
+from ray import tune
+from datetime import datetime
+import json
+
+DIR = os.path.dirname(os.path.abspath(__file__))
+activations = [torch.nn.Tanh, torch.nn.LeakyReLU]
+
+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 training_function(config):
+    np.random.seed(config['seed'])
+    torch.manual_seed(config['seed'])
+
+    if config['experiment'] == 'A':
+        envs = [Setobs(TransformObservation(CollisionPenaltyWrapper(
+            IntersimpleLidarFlatRandom(
+                n_rays=5,
+                reward=functools.partial(
+                    speed_reward,
+                    collision_penalty=0
+                ),
+                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)]
+
+    elif config['experiment'] == 'B': 
+        envs = sum([[Setobs(TransformObservation(CollisionPenaltyWrapper(
+            IntersimpleLidarFlatRandom(
+                n_rays=5,
+                reward=functools.partial(
+                    speed_reward,
+                    collision_penalty=0
+                ),
+                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)],[])
+    
+    else:
+        raise NotImplementedError
+
+    env_fn = lambda i: envs[i]
+
+    policy = SetPolicy(env_fn(0).action_space.shape[0], 
+        n_hidden_layers=config['policy']['n_hidden_layers'],
+        hidden_layer_size=config['policy']['hidden_layer_size'],
+        activation=activations[config['policy']['activation']] ) # config net architecture
+    pi_opt = torch.optim.Adam(policy.parameters(), lr=config['policy']['learning_rate'])
+    pi_lr_scheduler = torch.optim.lr_scheduler.ExponentialLR(pi_opt, gamma=config['policy']['learning_rate_decay'])
+
+    value = SetValue() # config net architecture
+    v_opt = torch.optim.Adam(value.parameters(), lr=config['value']['learning_rate'], weight_decay=config['value']['weight_decay'])
+
+    discriminator = DeepsetDiscriminator(
+        n_hidden_layers_element=config['discriminator']['n_hidden_layers_element'],
+        n_hidden_layers_global=config['discriminator']['n_hidden_layers_global'],
+        hidden_layer_size=config['discriminator']['hidden_layer_size'],
+        activation=activations[config['discriminator']['activation']],
+    )
+    disc_opt = torch.optim.Adam(discriminator.parameters(), lr=config['discriminator']['learning_rate'], weight_decay=config['discriminator']['weight_decay'])
+
+    if config['experiment'] == 'A': 
+        expert_data = torch.load(os.path.join(DIR, 'intersimple-expert-data-setobs2-loc0-track0.pt'))
+    elif config['experiment'] == 'B': 
+        expert_data = [
+            torch.load(os.path.join(DIR, 'intersimple-expert-data-setobs2-loc0-track0.pt')),
+            torch.load(os.path.join(DIR, 'intersimple-expert-data-setobs2-loc0-track1.pt')),
+            torch.load(os.path.join(DIR, 'intersimple-expert-data-setobs2-loc0-track2.pt')),
+            torch.load(os.path.join(DIR, 'intersimple-expert-data-setobs2-loc0-track3.pt')),
+        ]
+        d0 = [d[0] for d in expert_data]
+        d1 = [d[1] for d in expert_data]
+        d2 = [d[2] for d in expert_data]
+        d3 = [d[3] for d in expert_data]
+        expert_data = (torch.cat(d0), torch.cat(d1), torch.cat(d2), torch.cat(d3))
+    
+    expert_data = Buffer(*expert_data)
+
+    def callback(info):
+        tune.report(gen_mean_reward_per_episode=info['gen/mean_reward_per_episode'],
+                    disc_mean_reward_per_episode=info['disc/mean_reward_per_episode'], 
+                    mean_episode_length=info['gen/mean_episode_length'],
+                    gen_collision_rate=info['gen/collision_rate'])
+        
+        # save model checkpoints
+        ep = info['epoch'] + 1
+        if (ep % 25 == 0):
+            torch.save(info['policy'].state_dict(), f'policy_epoch{ep}.pt')
+
+    value, policy = gail_ppo(
+        env_fn=env_fn,
+        expert_data=expert_data,
+        discriminator=discriminator,
+        disc_opt=disc_opt,
+        disc_iters=config['discriminator']['iterations_per_epoch'],
+        policy=policy,
+        value=value,
+        v_opt=v_opt,
+        v_iters=config['value']['iterations_per_epoch'],
+        epochs=config['train_epochs'],
+        rollout_episodes=60, 
+        rollout_steps=200,
+        gamma=0.99,
+        gae_lambda=0.9,
+        clip_ratio=config['policy']['clip_ratio'],
+        pi_opt=pi_opt,
+        pi_iters=config['policy']['iterations_per_epoch'],
+        logger=SummaryWriter(comment='gail-ppo-options-setobs2'),
+        callback=callback,
+        lr_schedulers=[pi_lr_scheduler],
+    )
+
+    # save model
+    torch.save(policy.state_dict(), 'policy_final.pt')
+
+if __name__ == '__main__':
+    import argparse
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--train', choices=['A', 'B'])
+    parser.add_argument('--epochs', type=int, default=200)
+    parser.add_argument('--test', type=str, help='path to config file to run final training on')
+    parser.add_argument('--test_seeds', type=int, default=5)
+    parser.add_argument('--test_cpus', type=int, help='number of cpus available to split test seed training over')
+    args = parser.parse_args()
+
+    assert (args.train is None) ^ (args.test is None), 'Must either train on an experiment or test with a config file'
+
+    # if no test config specified, train
+    if args.test is None:
+        print('Running Tuning for Experiment %s'%(args.train))
+        analysis = tune.run(
+            training_function,
+            config={
+                'experiment': args.train,
+                'trainenv': {
+                    'stop_on_collision': False,
+                    'use_idm': True, 
+                },
+                'policy': {
+                    'learning_rate': 3e-4, 
+                    'learning_rate_decay': 1.0, 
+                    'clip_ratio': 0.2, 
+                    'iterations_per_epoch': 100,
+                    'hidden_layer_size': tune.grid_search([20, 40]),
+                    'n_hidden_layers': tune.grid_search([2, 3]), 
+                    'activation':0,
+                },
+                'value': {
+                    'learning_rate': 1e-4,
+                    'weight_decay': 1e-3, 
+                    'iterations_per_epoch': 1000, 
+                },
+                'discriminator': {
+                    'learning_rate': 1e-3, 
+                    'weight_decay': 1e-4, 
+                    'iterations_per_epoch': 100, 
+                    'n_hidden_layers_element': tune.grid_search([3,4]),
+                    'n_hidden_layers_global': tune.grid_search([1,2]),
+                    'hidden_layer_size': 10, 
+                    'activation': 0,
+                },
+                'train_epochs': args.epochs,
+                'seed': 0,
+            }
+        )
+        best_config = analysis.get_best_config(metric='gen_collision_rate', mode='min')
+        print('Best config: ', best_config)
+
+        # safe best_config
+        if not os.path.isdir(os.path.join(DIR, 'best_configs')):
+            os.mkdir(os.path.join(DIR, 'best_configs'))
+        
+        # save gail
+        with open(os.path.join(DIR, 'best_configs',f'gail_exp{args.train}.json'), 'w', encoding='utf-8') as f:
+            json.dump(best_config, f, ensure_ascii=False, indent=4)
+
+    # if config file specified, rerun it with appropriate number of seeds
+    else:
+        with open(args.test, 'rb') as f:
+            config = json.load(f)
+
+        print(f'Retraining {args.test} with {args.test_seeds} seeds on experiment {config["experiment"]}')
+
+        # rerun with appropriate number of seeds
+        rpt = {'cpu': int(args.test_cpus/args.test_seeds)} if (args.test_cpus is not None) else None
+        config['seed'] = tune.grid_search(list(range(1,args.test_seeds+1)))
+        analysis = tune.run(training_function, config=config, resources_per_trial=rpt)
+
+        # move final policies to appropriate directory
+        split_ = os.path.basename(args.test).split('_')
+        model = split_[0]
+        exper = split_[-1].split('.')[0]
+        savepath = os.path.join('test_policies',model,exper)
+        
+        if not os.path.isdir(savepath):
+            os.makedirs(savepath)
+        
+        import shutil
+        for i in range(args.test_seeds):
+            s = analysis._checkpoints[i]['config']['seed']
+            check_dir = analysis._checkpoints[i]['logdir']
+            shutil.copyfile(os.path.join(check_dir,'policy_final.pt'), 
+                os.path.join(savepath, f'policy_seed{s}.pt'))
+            shutil.copyfile(os.path.join(check_dir,'params.json'),
+                os.path.join(savepath, 'config.json')) # copy config automatically

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)