def test_box(self):
env = Gym(env_name="CartPole-v0")
obs_preprocessor = env.get_obs_preprocessor()
obs = env.reset()
state = obs_preprocessor(obs)
self.assertTrue(state.has_float_features_only)
self.assertEqual(state.float_features.shape, (1, obs.shape[0]))
self.assertEqual(state.float_features.dtype, torch.float32)
self.assertEqual(state.float_features.device, torch.device("cpu"))
npt.assert_array_almost_equal(obs, state.float_features.squeeze(0))
def run_test_offline(
env_name: str,
model: ModelManager__Union,
replay_memory_size: int,
num_batches_per_epoch: int,
num_train_epochs: int,
passing_score_bar: float,
num_eval_episodes: int,
use_gpu: bool,
):
env = Gym(env_name=env_name)
env.seed(SEED)
env.action_space.seed(SEED)
normalization = build_normalizer(env)
logger.info(f"Normalization is: \n{pprint.pformat(normalization)}")
manager = model.value
trainer = manager.initialize_trainer(
use_gpu=use_gpu,
reward_options=RewardOptions(),
normalization_data_map=normalization,
)
# first fill the replay buffer to burn_in
replay_buffer = ReplayBuffer(replay_capacity=replay_memory_size,
batch_size=trainer.minibatch_size)
# always fill full RB
fill_replay_buffer(env=env,
replay_buffer=replay_buffer,
desired_size=replay_memory_size)
device = torch.device("cuda") if use_gpu else None
# pyre-fixme[6]: Expected `device` for 2nd param but got `Optional[torch.device]`.
trainer_preprocessor = make_replay_buffer_trainer_preprocessor(
trainer, device, env)
writer = SummaryWriter()
with summary_writer_context(writer):
for epoch in range(num_train_epochs):
logger.info(f"Evaluating before epoch {epoch}: ")
eval_rewards = evaluate_cem(env, manager, 1)
for _ in tqdm(range(num_batches_per_epoch)):
train_batch = replay_buffer.sample_transition_batch()
preprocessed_batch = trainer_preprocessor(train_batch)
trainer.train(preprocessed_batch)
logger.info(f"Evaluating after training for {num_train_epochs} epochs: ")
eval_rewards = evaluate_cem(env, manager, num_eval_episodes)
mean_rewards = np.mean(eval_rewards)
assert (mean_rewards >= passing_score_bar
), f"{mean_rewards} doesn't pass the bar {passing_score_bar}."
def test_box_cuda(self):
env = Gym(env_name="CartPole-v0")
device = torch.device("cuda")
obs_preprocessor = env.get_obs_preprocessor(device=device)
obs = env.reset()
state = obs_preprocessor(obs)
self.assertTrue(state.has_float_features_only)
self.assertEqual(state.float_features.shape, (1, obs.shape[0]))
self.assertEqual(state.float_features.dtype, torch.float32)
# `device` doesn't have index. So we need this.
x = torch.zeros(1, device=device)
self.assertEqual(state.float_features.device, x.device)
npt.assert_array_almost_equal(obs,
state.float_features.cpu().squeeze(0))
def test_cartpole_reinforce(self):
# TODO(@badri) Parameterize this test
env = Gym("CartPole-v0")
norm = build_normalizer(env)
from reagent.net_builder.discrete_dqn.fully_connected import FullyConnected
net_builder = FullyConnected(sizes=[8], activations=["linear"])
cartpole_scorer = net_builder.build_q_network(
state_feature_config=None,
state_normalization_data=norm["state"],
output_dim=len(norm["action"].dense_normalization_parameters),
)
from reagent.gym.policies.samplers.discrete_sampler import SoftmaxActionSampler
policy = Policy(scorer=cartpole_scorer, sampler=SoftmaxActionSampler())
from reagent.training.reinforce import Reinforce, ReinforceParams
from reagent.optimizer.union import classes
trainer = Reinforce(
policy,
ReinforceParams(gamma=0.995,
optimizer=classes["Adam"](lr=5e-3,
weight_decay=1e-3)),
)
run_test_episode_buffer(
env,
policy,
trainer,
num_train_episodes=500,
passing_score_bar=180,
num_eval_episodes=100,
)
def evaluate_gym(
env_name: str,
model: ModelManager__Union,
publisher: ModelPublisher__Union,
num_eval_episodes: int,
passing_score_bar: float,
max_steps: Optional[int] = None,
):
publisher_manager = publisher.value
assert isinstance(
publisher_manager, FileSystemPublisher
), f"publishing manager is type {type(publisher_manager)}, not FileSystemPublisher"
env = Gym(env_name=env_name)
torchscript_path = publisher_manager.get_latest_published_model(
model.value)
jit_model = torch.jit.load(torchscript_path)
policy = create_predictor_policy_from_model(jit_model)
agent = Agent.create_for_env_with_serving_policy(env, policy)
rewards = evaluate_for_n_episodes(n=num_eval_episodes,
env=env,
agent=agent,
max_steps=max_steps)
avg_reward = np.mean(rewards)
logger.info(f"Average reward over {num_eval_episodes} is {avg_reward}.\n"
f"List of rewards: {rewards}")
assert (avg_reward >= passing_score_bar
), f"{avg_reward} fails to pass the bar of {passing_score_bar}!"
return
def offline_gym(
env_name: str,
pkl_path: str,
num_train_transitions: int,
max_steps: Optional[int],
seed: Optional[int] = None,
):
"""
Generate samples from a DiscreteRandomPolicy on the Gym environment and
saves results in a pandas df parquet.
"""
initialize_seed(seed)
env = Gym(env_name=env_name)
replay_buffer = ReplayBuffer(replay_capacity=num_train_transitions,
batch_size=1)
fill_replay_buffer(env, replay_buffer, num_train_transitions)
if isinstance(env.action_space, gym.spaces.Discrete):
is_discrete_action = True
else:
assert isinstance(env.action_space, gym.spaces.Box)
is_discrete_action = False
df = replay_buffer_to_pre_timeline_df(is_discrete_action, replay_buffer)
logger.info(f"Saving dataset with {len(df)} samples to {pkl_path}")
df.to_pickle(pkl_path)
def test_random_vs_lqr(self):
"""
Test random actions vs. a LQR controller. LQR controller should perform
much better than random actions in the linear dynamics environment.
"""
env = Gym(env_name="LinearDynamics-v0")
num_test_episodes = 500
def random_policy(env, state):
return np.random.uniform(
env.action_space.low, env.action_space.high, env.action_dim
)
def lqr_policy(env, state):
# Four matrices that characterize the environment
A, B, Q, R = env.A, env.B, env.Q, env.R
# Solve discrete algebraic Riccati equation:
M = linalg.solve_discrete_are(A, B, Q, R)
K = np.dot(
linalg.inv(np.dot(np.dot(B.T, M), B) + R), (np.dot(np.dot(B.T, M), A))
)
state = state.reshape((-1, 1))
action = -K.dot(state).squeeze()
return action
mean_acc_rws_random = self.run_n_episodes(env, num_test_episodes, random_policy)
mean_acc_rws_lqr = self.run_n_episodes(env, num_test_episodes, lqr_policy)
logger.info(f"Mean acc. reward of random policy: {mean_acc_rws_random}")
logger.info(f"Mean acc. reward of LQR policy: {mean_acc_rws_lqr}")
assert mean_acc_rws_lqr > mean_acc_rws_random
def test_pocman(self):
env = Gym(env_name="Pocman-v0")
env.seed(313)
mean_acc_reward = self._test_env(env)
assert -80 <= mean_acc_reward <= -70
def test_string_game(self):
env = Gym(env_name="StringGame-v0")
env.seed(313)
mean_acc_reward = self._test_env(env)
assert 0.1 >= mean_acc_reward
def train_seq2reward_and_compute_reward_mse(
env_name: str,
model: ModelManager__Union,
num_train_transitions: int,
num_test_transitions: int,
seq_len: int,
batch_size: int,
num_train_epochs: int,
use_gpu: bool,
saved_seq2reward_path: Optional[str] = None,
):
""" Train Seq2Reward Network and compute reward mse. """
env = Gym(env_name=env_name)
env.seed(SEED)
manager = model.value
trainer = manager.initialize_trainer(
use_gpu=use_gpu,
reward_options=RewardOptions(),
normalization_data_map=build_normalizer(env),
)
device = "cuda" if use_gpu else "cpu"
# pyre-fixme[6]: Expected `device` for 2nd param but got `str`.
trainer_preprocessor = make_replay_buffer_trainer_preprocessor(
trainer, device, env)
test_replay_buffer = ReplayBuffer(
replay_capacity=num_test_transitions,
batch_size=batch_size,
stack_size=seq_len,
return_everything_as_stack=True,
)
fill_replay_buffer(env, test_replay_buffer, num_test_transitions)
if saved_seq2reward_path is None:
# train from scratch
trainer = train_seq2reward(
env=env,
trainer=trainer,
trainer_preprocessor=trainer_preprocessor,
num_train_transitions=num_train_transitions,
seq_len=seq_len,
batch_size=batch_size,
num_train_epochs=num_train_epochs,
test_replay_buffer=test_replay_buffer,
)
else:
# load a pretrained model, and just evaluate it
trainer.seq2reward_network.load_state_dict(
torch.load(saved_seq2reward_path))
state_dim = env.observation_space.shape[0]
with torch.no_grad():
trainer.seq2reward_network.eval()
test_batch = test_replay_buffer.sample_transition_batch(
batch_size=test_replay_buffer.size)
preprocessed_test_batch = trainer_preprocessor(test_batch)
adhoc_action_padding(preprocessed_test_batch, state_dim=state_dim)
losses = trainer.get_loss(preprocessed_test_batch)
detached_losses = losses.cpu().detach().item()
trainer.seq2reward_network.train()
return detached_losses
def train_mdnrnn_and_train_on_embedded_env(
env_name: str,
embedding_model: ModelManager__Union,
num_embedding_train_transitions: int,
seq_len: int,
batch_size: int,
num_embedding_train_epochs: int,
train_model: ModelManager__Union,
num_state_embed_transitions: int,
num_agent_train_epochs: int,
num_agent_eval_epochs: int,
use_gpu: bool,
passing_score_bar: float,
# pyre-fixme[9]: saved_mdnrnn_path has type `str`; used as `None`.
saved_mdnrnn_path: str = None,
):
""" Train an agent on embedded states by the MDNRNN. """
env = Gym(env_name=env_name)
env.seed(SEED)
embedding_manager = embedding_model.value
embedding_trainer = embedding_manager.initialize_trainer(
use_gpu=use_gpu,
reward_options=RewardOptions(),
normalization_data_map=build_normalizer(env),
)
device = "cuda" if use_gpu else "cpu"
embedding_trainer_preprocessor = make_replay_buffer_trainer_preprocessor(
embedding_trainer,
# pyre-fixme[6]: Expected `device` for 2nd param but got `str`.
device,
env,
)
if saved_mdnrnn_path is None:
# train from scratch
embedding_trainer = train_mdnrnn(
env=env,
trainer=embedding_trainer,
trainer_preprocessor=embedding_trainer_preprocessor,
num_train_transitions=num_embedding_train_transitions,
seq_len=seq_len,
batch_size=batch_size,
num_train_epochs=num_embedding_train_epochs,
)
else:
# load a pretrained model, and just evaluate it
embedding_trainer.memory_network.mdnrnn.load_state_dict(
torch.load(saved_mdnrnn_path))
# create embedding dataset
embed_rb, state_min, state_max = create_embed_rl_dataset(
env=env,
memory_network=embedding_trainer.memory_network,
num_state_embed_transitions=num_state_embed_transitions,
batch_size=batch_size,
seq_len=seq_len,
hidden_dim=embedding_trainer.params.hidden_size,
use_gpu=use_gpu,
)
embed_env = StateEmbedEnvironment(
gym_env=env,
mdnrnn=embedding_trainer.memory_network,
max_embed_seq_len=seq_len,
state_min_value=state_min,
state_max_value=state_max,
)
agent_manager = train_model.value
agent_trainer = agent_manager.initialize_trainer(
use_gpu=use_gpu,
reward_options=RewardOptions(),
# pyre-fixme[6]: Expected `EnvWrapper` for 1st param but got
# `StateEmbedEnvironment`.
normalization_data_map=build_normalizer(embed_env),
)
device = "cuda" if use_gpu else "cpu"
agent_trainer_preprocessor = make_replay_buffer_trainer_preprocessor(
agent_trainer,
# pyre-fixme[6]: Expected `device` for 2nd param but got `str`.
device,
env,
)
num_batch_per_epoch = embed_rb.size // batch_size
for epoch in range(num_agent_train_epochs):
for _ in tqdm(range(num_batch_per_epoch), desc=f"epoch {epoch}"):
batch = embed_rb.sample_transition_batch(batch_size=batch_size)
preprocessed_batch = agent_trainer_preprocessor(batch)
agent_trainer.train(preprocessed_batch)
# evaluate model
rewards = []
policy = agent_manager.create_policy(serving=False)
# pyre-fixme[6]: Expected `EnvWrapper` for 1st param but got
# `StateEmbedEnvironment`.
agent = Agent.create_for_env(embed_env, policy=policy, device=device)
# num_processes=1 needed to avoid workers from dying on CircleCI tests
rewards = evaluate_for_n_episodes(
n=num_agent_eval_epochs,
# pyre-fixme[6]: Expected `EnvWrapper` for 2nd param but got
# `StateEmbedEnvironment`.
env=embed_env,
agent=agent,
num_processes=1,
)
assert (np.mean(rewards) >= passing_score_bar
), f"average reward doesn't pass our bar {passing_score_bar}"
return rewards
def train_mdnrnn_and_compute_feature_stats(
env_name: str,
model: ModelManager__Union,
num_train_transitions: int,
num_test_transitions: int,
seq_len: int,
batch_size: int,
num_train_epochs: int,
use_gpu: bool,
saved_mdnrnn_path: Optional[str] = None,
):
""" Train MDNRNN Memory Network and compute feature importance/sensitivity. """
env: gym.Env = Gym(env_name=env_name)
env.seed(SEED)
manager = model.value
trainer = manager.initialize_trainer(
use_gpu=use_gpu,
reward_options=RewardOptions(),
normalization_data_map=build_normalizer(env),
)
device = "cuda" if use_gpu else "cpu"
# pyre-fixme[6]: Expected `device` for 2nd param but got `str`.
trainer_preprocessor = make_replay_buffer_trainer_preprocessor(
trainer, device, env)
test_replay_buffer = ReplayBuffer(
replay_capacity=num_test_transitions,
batch_size=batch_size,
stack_size=seq_len,
return_everything_as_stack=True,
)
fill_replay_buffer(env, test_replay_buffer, num_test_transitions)
if saved_mdnrnn_path is None:
# train from scratch
trainer = train_mdnrnn(
env=env,
trainer=trainer,
trainer_preprocessor=trainer_preprocessor,
num_train_transitions=num_train_transitions,
seq_len=seq_len,
batch_size=batch_size,
num_train_epochs=num_train_epochs,
test_replay_buffer=test_replay_buffer,
)
else:
# load a pretrained model, and just evaluate it
trainer.memory_network.mdnrnn.load_state_dict(
torch.load(saved_mdnrnn_path))
with torch.no_grad():
trainer.memory_network.mdnrnn.eval()
test_batch = test_replay_buffer.sample_transition_batch(
batch_size=test_replay_buffer.size)
preprocessed_test_batch = trainer_preprocessor(test_batch)
feature_importance = calculate_feature_importance(
env=env,
trainer=trainer,
use_gpu=use_gpu,
test_batch=preprocessed_test_batch,
)
feature_sensitivity = calculate_feature_sensitivity(
env=env,
trainer=trainer,
use_gpu=use_gpu,
test_batch=preprocessed_test_batch,
)
trainer.memory_network.mdnrnn.train()
return feature_importance, feature_sensitivity