Merge branch 'main' into dev

scratch/etienne/intersimple/data/expert.py

@@ -1,4 +1,4 @@
-from intersim.envs.intersimple import Intersimple
+from intersim.envs.intersimple import Intersimple, InfoFilter
 from stable_baselines3.common.policies import BasePolicy
 import gym
 import intersim.envs.intersimple
@@ -119,6 +119,7 @@ def demonstrations(expert='NormalizedIntersimpleExpert', env='NRasterizedRandomA
         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,
@@ -127,7 +128,8 @@ def demonstrations(expert='NormalizedIntersimpleExpert', env='NRasterizedRandomA
         sample_until=rollout.make_sample_until(
             min_timesteps=min_timesteps,
             min_episodes=min_episodes,
-        )
+        ),
+        exclude_infos=not include_infos,
     )
 
 if __name__ == '__main__':

scratch/etienne/intersimple/gail_options_image_alltracks.py

@@ -331,7 +331,7 @@ if __name__ == '__main__':
     
     #env_class = NRasterized
     #env_settings = {'agent': 51, 'width': 36, 'height': 36, 'm_per_px': 2}
-    files = ['../../../expert_data/DR_USA_Roundabout_FT0/track%04i/expert.pkl'%(i) for i in range(5)]
+    files = ['../../../expert_data/DR_USA_Roundabout_FT/track%04i/expert.pkl'%(i) for i in range(5)]
     transitions=load_experts(files)
 
     generator = train(

scratch/johannes/evaluation.py

@@ -0,0 +1,90 @@
+
+def evaluate_policy_simple(
+    model,
+    env: gym.Env,
+    n_eval_episodes: int = 10,
+    deterministic: bool = True,
+    render: bool = False,
+    callback = None,
+    reward_threshold = None,
+    return_episode_rewards: bool = False,
+    warn: bool = True,
+):
+    """
+    Runs policy for ``n_eval_episodes`` episodes and returns average reward.
+    If a vector env is passed in, this divides the episodes to evaluate onto the
+    different elements of the vector env. This static division of work is done to
+    remove bias. See https://github.com/DLR-RM/stable-baselines3/issues/402 for more
+    details and discussion.
+
+    .. note::
+        If environment has not been wrapped with ``Monitor`` wrapper, reward and
+        episode lengths are counted as it appears with ``env.step`` calls. If
+        the environment contains wrappers that modify rewards or episode lengths
+        (e.g. reward scaling, early episode reset), these will affect the evaluation
+        results as well. You can avoid this by wrapping environment with ``Monitor``
+        wrapper before anything else.
+
+    :param model: The RL agent you want to evaluate.
+    :param env: The gym environment or ``VecEnv`` environment.
+    :param n_eval_episodes: Number of episode to evaluate the agent
+    :param deterministic: Whether to use deterministic or stochastic actions
+    :param render: Whether to render the environment or not
+    :param callback: callback function to do additional checks,
+        called after each step. Gets locals() and globals() passed as parameters.
+    :param reward_threshold: Minimum expected reward per episode,
+        this will raise an error if the performance is not met
+    :param return_episode_rewards: If True, a list of rewards and episode lengths
+        per episode will be returned instead of the mean.
+    :param warn: If True (default), warns user about lack of a Monitor wrapper in the
+        evaluation environment.
+    :return: Mean reward per episode, std of reward per episode.
+        Returns ([float], [int]) when ``return_episode_rewards`` is True, first
+        list containing per-episode rewards and second containing per-episode lengths
+        (in number of steps).
+    """
+    episode_rewards = []
+    episode_lengths = []
+
+    episode_counts = 0
+
+    current_rewards = 0
+    current_lengths = 0
+    observations = env.reset()
+    states = None
+    while (episode_counts < n_eval_episodes):
+        actions, states = model.predict(observations, state=states, deterministic=deterministic)
+        observations, rewards, dones, infos = env.step(actions)
+        print(env._env.t)
+        current_rewards += rewards
+        current_lengths += 1
+
+        # unpack values so that the callback can access the local variables
+        reward = rewards
+        done = dones
+        info = infos
+        if info['collision']:
+            print("COLLISION")
+
+        if callback is not None:
+            callback(locals(), globals())
+
+        if dones:
+            episode_rewards.append(current_rewards)
+            episode_lengths.append(current_lengths)
+            episode_counts += 1
+            current_rewards = 0
+            current_lengths = 0
+            if states is not None:
+                states *= 0
+
+        if render:
+            env.render()
+
+    mean_reward = np.mean(episode_rewards)
+    std_reward = np.std(episode_rewards)
+    if reward_threshold is not None:
+        assert mean_reward > reward_threshold, "Mean reward below threshold: " f"{mean_reward:.2f} < {reward_threshold:.2f}"
+    if return_episode_rewards:
+        return episode_rewards, episode_lengths
+    return mean_reward, std_reward

scratch/johannes/gail_options_image.py

@@ -0,0 +1,96 @@
+# %%
+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, speed_reward
+import itertools
+import functools
+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
+from src.evaluation.evaluation import Evaluation
+
+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 epoch 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)
+
+        eval_env = env
+        ev = Evaluation(eval_env, n_eval_episodes=100)
+        ev.evaluate(epoch, generator, discriminator, expert_data)
+    
+    return generator
+
+# %%
+if __name__ == '__main__':
+    # %%
+ 
+    with open("data/NormalizedIntersimpleExpertMu.001_NRasterizedInfoAgent51w36h36mppx2.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/johannes/raytune_simple.py

@@ -0,0 +1,69 @@
+"""This example demonstrates basic Ray Tune random search and grid search."""
+import time
+
+import ray
+from ray import tune
+
+
+def evaluation_fn(step, width, height):
+    time.sleep(0.1)
+    return (0.1 + width * step / 100)**(-1) + height * 0.1
+
+def easy_objective(config):
+    # Hyperparameters
+    width, height = config["width"], config["height"]
+
+    mydata = ray.get(ray_data)
+    print(mydata)
+
+    for step in range(config["steps"]):
+        # Iterative training function - can be any arbitrary training procedure
+        intermediate_score = evaluation_fn(step, width, height)
+        # Feed the score back back to Tune.
+        tune.report(iterations=step, mean_loss=intermediate_score)
+
+
+if __name__ == "__main__":
+    import argparse
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--smoke-test", action="store_true", help="Finish quickly for testing")
+    parser.add_argument(
+        "--server-address",
+        type=str,
+        default=None,
+        required=False,
+        help="The address of server to connect to if using "
+        "Ray Client.")
+    args, _ = parser.parse_known_args()
+    if args.server_address is not None:
+        ray.init(f"ray://{args.server_address}")
+    else:
+        ray.init(configure_logging=False)
+
+    # This will do a grid search over the `activation` parameter. This means
+    # that each of the two values (`relu` and `tanh`) will be sampled once
+    # for each sample (`num_samples`). We end up with 2 * 50 = 100 samples.
+    # The `width` and `height` parameters are sampled randomly.
+    # `steps` is a constant parameter.
+
+    import numpy as np
+    N = 3
+    data = np.random.rand(N,N,N)
+    ray_data = ray.put(data)
+
+
+    analysis = tune.run(
+        easy_objective,
+        metric="mean_loss",
+        mode="min",
+        num_samples=5 if args.smoke_test else 50,
+        config={
+            "steps": 5 if args.smoke_test else 100,
+            "width": tune.uniform(0, 20),
+            "height": tune.uniform(-100, 100),
+            "activation": tune.grid_search(["relu", "tanh"])
+        })
+
+    print("Best hyperparameters found were: ", analysis.best_config)

src/__init__.py

@@ -1,3 +1,3 @@
 from src.data.expert_data import generate_expert_data, load_expert_data
 from src.data.data_utils import InteractionDatasetSingleAgent
-from src.metrics import metrics
+from src.evaluation.metrics import metrics

src/evaluation/evaluation.py

@@ -0,0 +1,87 @@
+import torch
+import numpy as np
+from stable_baselines3.common.vec_env import VecEnv
+from stable_baselines3.common.evaluation import evaluate_policy
+
+from intersim.envs.intersimple import Intersimple
+from src.evaluation.metrics import nanmean, divergence, visualize_distribution
+
+class Evaluation:
+    def __init__(self, eval_env, n_eval_episodes=10):
+        # if env is a VecEnv, the code needs to be adapted, since the callback will be called after each step, 
+        # so transitions of different envs will be mixed and the total number of episodes could be larger than n_eval_episodes!
+        assert not isinstance(eval_env, VecEnv)
+        self.env = eval_env
+        self.n_eval_episodes = n_eval_episodes
+        self.reset()
+
+    def reset(self):
+        self._n_collisions = 0
+        self._trajectories = []
+        self._episode_done = True
+        self._accelerations = []
+
+    def evaluate(self, epoch, generator, discriminator, expert_data):
+        self.reset()
+        metrics = {}
+        
+        episode_rewards, episode_lengths = evaluate_policy(
+            generator, 
+            self.env,
+            n_eval_episodes=self.n_eval_episodes,
+            callback=self.evaluate_policy_callback, 
+            return_episode_rewards=True
+        )
+
+        collision_rate = self._n_collisions / self.n_eval_episodes
+        metrics['collision_rate'] = collision_rate
+
+        assert len(self._trajectories) >= self.n_eval_episodes
+
+        # velocities produced by generator
+        policy_velocities = torch.cat([torch.stack(t)[:,2] for t in self._trajectories])
+        # if episodes terminate without collisions, then the state is fully nan
+        policy_velocities = policy_velocities[~torch.isnan(policy_velocities)]
+
+        # expert velocities
+        extract_state = lambda info: info['projected_state'][info['agent']]
+        expert_velocities = torch.stack([extract_state(info) for info in expert_data.infos])[:,2]
+        expert_velocities = expert_velocities[~torch.isnan(expert_velocities)]
+
+        metrics['avg_velocity_loss'] = (expert_velocities.mean() - policy_velocities.mean()).item()
+        metrics['velocity_divergence'] = divergence(policy_velocities, expert_velocities, type='js')
+        
+
+        # accelerations produced by generator
+        policy_accelerations = torch.tensor(self._accelerations)
+        # expert accelerations
+        extract_accel = lambda info: info['action_taken'][info['agent']]
+        expert_accelerations = torch.cat([extract_accel(info) for info in expert_data.infos])
+
+        metrics['acceleration_divergence'] = divergence(policy_accelerations, expert_accelerations, type='js')
+        visualize_distribution(expert_accelerations, policy_accelerations, 'output/_action_viz{:02}'.format(epoch)) 
+
+        print(metrics)
+        return metrics
+
+    def evaluate_policy_callback(self, local_vars, global_vars):
+        venv_i = local_vars['i']
+        info = local_vars['info']
+        done = local_vars['done']
+        _agent = info['agent']
+        env = local_vars['env'].envs[venv_i]
+        assert isinstance(env, Intersimple)
+
+        # Increase collision counter if episode terminated with a collision
+        if info['collision']:
+            assert done
+            self._n_collisions += 1
+
+        # if last episode is done, start new trajectory
+        # this is currently not necessary, only if velocity is to be averaged over individual trajectories first
+        # and then averaging over all trajectories
+        if self._episode_done:
+            self._trajectories.append([])
+        self._trajectories[-1].append(info['projected_state'][_agent])
+        self._accelerations.append(info['action_taken'][_agent])
+        self._episode_done = done

src/gail/options.py

@@ -40,7 +40,7 @@ class OptionsEnv(gym.Wrapper):
                 # action 0 is considered safe fallback
                 self.m[0] = True
             
-            self.ch, self.value, self.log_prob = generator.policy.predict({
+            self.ch, self.value, self.log_prob = generator.policy.forward({
                 'obs': torch.tensor(self.s).unsqueeze(0).to(generator.policy.device),
                 'mask': self.m.unsqueeze(0).to(generator.policy.device),
             })

src/gail/train.py

@@ -9,9 +9,10 @@ def flatten_transitions(transitions):
         'dones': np.stack(list(t['dones'] for t in transitions), axis=0),
     }
 
-def train_discriminator(env, generator, discriminator, num_samples):
+def train_discriminator(env, generator, discriminator, num_samples, n_updates=1):
     transitions = list(itertools.islice(env.sample_ll(generator), num_samples))
     generator_samples = flatten_transitions(transitions)
+    for _ in range(n_updates):
         discriminator.train_disc(gen_samples=generator_samples)
 
 def train_generator(env, generator, discriminator, num_samples):

src/policies/options.py

@@ -22,7 +22,7 @@ class OptionsCnnPolicy(stable_baselines3.common.policies.ActorCriticCnnPolicy):
         values = self.value_net(latent_vf)
         return values, distribution.distribution
 
-    def predict(self, obs, eps=1e-6):
+    def forward(self, obs, eps=1e-6):
         """
         Will mask invalid states before making action selections
         Args: