def main(model_path, temperature):
model_path = glob.glob(model_path)[0]
predictor = DiscreteDqnTorchPredictor(torch.jit.load(model_path))
predictor.softmax_temperature = temperature
env = OpenAIGymEnvironment(gymenv=ENV)
avg_rewards, avg_discounted_rewards = env.run_ep_n_times(
AVG_OVER_NUM_EPS, predictor, test=True
)
logger.info(
"Achieved an average reward score of {} over {} evaluations.".format(
avg_rewards, AVG_OVER_NUM_EPS
)
)
def test_open_ai_gym_generate_samples_multi_step(self):
env = OpenAIGymEnvironment(
"CartPole-v0",
epsilon=1.0, # take random actions to collect training data
softmax_policy=False,
gamma=0.9,
)
num_samples = 1000
num_steps = 5
samples = env.generate_random_samples(
num_samples,
use_continuous_action=True,
epsilon=1.0,
multi_steps=num_steps,
include_shorter_samples_at_start=True,
include_shorter_samples_at_end=True,
)
self._check_samples(samples, num_samples, num_steps, True)
def run_gym(
params: OpenAiGymParameters,
score_bar,
embed_rl_dataset: RLDataset,
gym_env: Env,
mdnrnn: MemoryNetwork,
max_embed_seq_len: int,
):
assert params.rl is not None
rl_parameters = params.rl
env_type = params.env
model_type = params.model_type
epsilon, epsilon_decay, minimum_epsilon = create_epsilon(
offline_train=True, rl_parameters=rl_parameters, params=params)
replay_buffer = OpenAIGymMemoryPool(params.max_replay_memory_size)
for row in embed_rl_dataset.rows:
replay_buffer.insert_into_memory(**row)
assert replay_buffer.memory_buffer is not None
state_mem = replay_buffer.memory_buffer.state
state_min_value = torch.min(state_mem).item()
state_max_value = torch.max(state_mem).item()
state_embed_env = StateEmbedGymEnvironment(gym_env, mdnrnn,
max_embed_seq_len,
state_min_value,
state_max_value)
open_ai_env = OpenAIGymEnvironment(
state_embed_env,
epsilon,
rl_parameters.softmax_policy,
rl_parameters.gamma,
epsilon_decay,
minimum_epsilon,
)
rl_trainer = create_trainer(params, open_ai_env)
rl_predictor = create_predictor(rl_trainer, model_type, params.use_gpu,
open_ai_env.action_dim)
assert (params.run_details.max_steps is not None
and params.run_details.offline_train_epochs is not None
), "Missing data required for offline training: {}".format(
str(params.run_details))
return train_gym_offline_rl(
gym_env=open_ai_env,
replay_buffer=replay_buffer,
model_type=model_type,
trainer=rl_trainer,
predictor=rl_predictor,
test_run_name="{} offline rl state embed".format(env_type),
score_bar=score_bar,
max_steps=params.run_details.max_steps,
avg_over_num_episodes=params.run_details.avg_over_num_episodes,
offline_train_epochs=params.run_details.offline_train_epochs,
num_batch_per_epoch=None,
)
def mdnrnn_gym(
params: OpenAiGymParameters,
feature_importance: bool = False,
feature_sensitivity: bool = False,
save_embedding_to_path: Optional[str] = None,
seed: Optional[int] = None,
):
assert params.mdnrnn is not None
use_gpu = params.use_gpu
logger.info("Running gym with params")
logger.info(params)
env_type = params.env
env = OpenAIGymEnvironment(env_type,
epsilon=1.0,
softmax_policy=False,
gamma=0.99,
random_seed=seed)
# create test data once
assert params.run_details.max_steps is not None
test_replay_buffer = get_replay_buffer(
params.run_details.num_test_episodes,
params.run_details.seq_len,
params.run_details.max_steps,
env,
)
test_batch = test_replay_buffer.sample_memories(
test_replay_buffer.memory_size, use_gpu=use_gpu, batch_first=True)
trainer = create_trainer(params, env, use_gpu)
_, _, trainer = train_sgd(
env,
trainer,
use_gpu,
"{} test run".format(env_type),
params.mdnrnn.minibatch_size,
params.run_details,
test_batch=test_batch,
)
feature_importance_map, feature_sensitivity_map, dataset = None, None, None
if feature_importance:
feature_importance_map = calculate_feature_importance(
env, trainer, use_gpu, params.run_details, test_batch=test_batch)
if feature_sensitivity:
feature_sensitivity_map = calculate_feature_sensitivity_by_actions(
env, trainer, use_gpu, params.run_details, test_batch=test_batch)
if save_embedding_to_path:
dataset = RLDataset(save_embedding_to_path)
create_embed_rl_dataset(env, trainer, dataset, use_gpu,
params.run_details)
dataset.save()
return env, trainer, feature_importance_map, feature_sensitivity_map, dataset
def run_gym(
params: OpenAiGymParameters,
offline_train,
score_bar,
seed=None,
save_timesteps_to_dataset=None,
start_saving_from_score=None,
path_to_pickled_transitions=None,
warm_trainer=None,
reward_shape_func=None,
):
use_gpu = params.use_gpu
logger.info("Running gym with params")
logger.info(params)
assert params.rl is not None
rl_parameters = params.rl
env_type = params.env
model_type = params.model_type
epsilon, epsilon_decay, minimum_epsilon = create_epsilon(
offline_train, rl_parameters, params
)
env = OpenAIGymEnvironment(
env_type,
epsilon,
rl_parameters.softmax_policy,
rl_parameters.gamma,
epsilon_decay,
minimum_epsilon,
seed,
)
replay_buffer = create_replay_buffer(
env, params, model_type, offline_train, path_to_pickled_transitions
)
trainer = warm_trainer if warm_trainer else create_trainer(params, env)
predictor = create_predictor(trainer, model_type, use_gpu, env.action_dim)
return train(
env,
offline_train,
replay_buffer,
model_type,
trainer,
predictor,
"{} test run".format(env_type),
score_bar,
params.run_details,
save_timesteps_to_dataset=save_timesteps_to_dataset,
start_saving_from_score=start_saving_from_score,
reward_shape_func=reward_shape_func,
)
def multi_step_sample_generator(
gym_env: OpenAIGymEnvironment,
num_transitions: int,
max_steps: Optional[int],
multi_steps: int,
include_shorter_samples_at_start: bool,
include_shorter_samples_at_end: bool,
):
"""
Convert gym env multi-step sample format to mdn-rnn multi-step sample format
:param gym_env: The environment used to generate multi-step samples
:param num_transitions: # of samples to return
:param max_steps: An episode terminates when the horizon is beyond max_steps
:param multi_steps: # of steps of states and actions per sample
:param include_shorter_samples_at_start: Whether to keep samples of shorter steps
which are generated at the beginning of an episode
:param include_shorter_samples_at_end: Whether to keep samples of shorter steps
which are generated at the end of an episode
"""
samples = gym_env.generate_random_samples(
num_transitions=num_transitions,
use_continuous_action=True,
max_step=max_steps,
multi_steps=multi_steps,
include_shorter_samples_at_start=include_shorter_samples_at_start,
include_shorter_samples_at_end=include_shorter_samples_at_end,
)
for j in range(num_transitions):
sample_steps = len(samples.terminals[j]) # type: ignore
state = dict_to_np(samples.states[j],
np_size=gym_env.state_dim,
key_offset=0)
action = dict_to_np(samples.actions[j],
np_size=gym_env.action_dim,
key_offset=gym_env.state_dim)
next_actions = np.float32( # type: ignore
[
dict_to_np(
samples.next_actions[j][k],
np_size=gym_env.action_dim,
key_offset=gym_env.state_dim,
) for k in range(sample_steps)
])
next_states = np.float32( # type: ignore
[
dict_to_np(samples.next_states[j][k],
np_size=gym_env.state_dim,
key_offset=0) for k in range(sample_steps)
])
rewards = np.float32(samples.rewards[j]) # type: ignore
terminals = np.float32(samples.terminals[j]) # type: ignore
not_terminals = np.logical_not(terminals)
ordered_states = np.vstack((state, next_states))
ordered_actions = np.vstack((action, next_actions))
mdnrnn_states = ordered_states[:-1]
mdnrnn_actions = ordered_actions[:-1]
mdnrnn_next_states = ordered_states[-multi_steps:]
mdnrnn_next_actions = ordered_actions[-multi_steps:]
# Padding zeros so that all samples have equal steps
# The general rule is to pad zeros at the end of sequences.
# In addition, if the sequence only has one step (i.e., the
# first state of an episode), pad one zero row ahead of the
# sequence, which enables embedding generated properly for
# one-step samples
num_padded_top_rows = 1 if multi_steps > 1 and sample_steps == 1 else 0
num_padded_bottom_rows = multi_steps - sample_steps - num_padded_top_rows
sample_steps_next = len(mdnrnn_next_states)
num_padded_top_rows_next = 0
num_padded_bottom_rows_next = multi_steps - sample_steps_next
yield (
np.pad(
mdnrnn_states,
((num_padded_top_rows, num_padded_bottom_rows), (0, 0)),
"constant",
constant_values=0.0,
),
np.pad(
mdnrnn_actions,
((num_padded_top_rows, num_padded_bottom_rows), (0, 0)),
"constant",
constant_values=0.0,
),
np.pad(
rewards,
((num_padded_top_rows, num_padded_bottom_rows)),
"constant",
constant_values=0.0,
),
np.pad(
mdnrnn_next_states,
((num_padded_top_rows_next, num_padded_bottom_rows_next),
(0, 0)),
"constant",
constant_values=0.0,
),
np.pad(
mdnrnn_next_actions,
((num_padded_top_rows_next, num_padded_bottom_rows_next),
(0, 0)),
"constant",
constant_values=0.0,
),
np.pad(
not_terminals,
((num_padded_top_rows, num_padded_bottom_rows)),
"constant",
constant_values=0.0,
),
sample_steps,
sample_steps_next,
)
def train_gym_offline_rl(
gym_env: OpenAIGymEnvironment,
replay_buffer: OpenAIGymMemoryPool,
model_type: str,
trainer: RLTrainer,
predictor: OnPolicyPredictor,
test_run_name: str,
score_bar: Optional[float],
max_steps: int,
avg_over_num_episodes: int,
offline_train_epochs: int,
num_batch_per_epoch: Optional[int],
bcq_imitator_hyper_params: Optional[Dict[str, Any]] = None,
):
if num_batch_per_epoch is None:
num_batch_per_epoch = replay_buffer.size // trainer.minibatch_size
assert num_batch_per_epoch > 0, "The size of replay buffer is not sufficient"
logger.info(
"{} offline transitions in replay buffer.\n"
"Training will take {} epochs, with each epoch having {} mini-batches"
" and each mini-batch having {} samples".format(
replay_buffer.size,
offline_train_epochs,
num_batch_per_epoch,
trainer.minibatch_size,
))
avg_reward_history, epoch_history = [], []
# Pre-train a GBDT imitator if doing batch constrained q-learning in Gym
if getattr(trainer, "bcq", None):
assert bcq_imitator_hyper_params is not None
gbdt = GradientBoostingClassifier(
n_estimators=bcq_imitator_hyper_params["gbdt_trees"],
max_depth=bcq_imitator_hyper_params["max_depth"],
)
samples = replay_buffer.sample_memories(replay_buffer.size, model_type)
X, y = samples.states.numpy(), torch.max(samples.actions,
dim=1)[1].numpy()
X_train, X_test, y_train, y_test = train_test_split(X,
y,
test_size=0.1)
logger.info("Fitting GBDT...")
gbdt.fit(X_train, y_train)
train_score = round(gbdt.score(X_train, y_train) * 100, 1)
test_score = round(gbdt.score(X_test, y_test) * 100, 1)
logger.info("GBDT train accuracy {}% || test accuracy {}%".format(
train_score, test_score))
trainer.bcq_imitator = gbdt.predict_proba # type: ignore
# Offline training
for i_epoch in range(offline_train_epochs):
for _ in range(num_batch_per_epoch):
samples = replay_buffer.sample_memories(trainer.minibatch_size,
model_type)
samples.set_device(trainer.device)
trainer.train(samples)
batch_td_loss = float(
torch.mean(
torch.tensor([
stat.td_loss
for stat in trainer.loss_reporter.incoming_stats
])))
trainer.loss_reporter.flush()
logger.info("Average TD loss: {} in epoch {}".format(
batch_td_loss, i_epoch + 1))
# test model performance for this epoch
avg_rewards, avg_discounted_rewards = gym_env.run_ep_n_times(
avg_over_num_episodes, predictor, test=True, max_steps=max_steps)
avg_reward_history.append(avg_rewards)
# For offline training, use epoch number as timestep history since
# we have a fixed batch of data to count epochs over.
epoch_history.append(i_epoch)
logger.info(
"Achieved an average reward score of {} over {} evaluations"
" after epoch {}.".format(avg_rewards, avg_over_num_episodes,
i_epoch))
if score_bar is not None and avg_rewards > score_bar:
logger.info("Avg. reward history for {}: {}".format(
test_run_name, avg_reward_history))
return avg_reward_history, epoch_history, trainer, predictor, gym_env
logger.info("Avg. reward history for {}: {}".format(
test_run_name, avg_reward_history))
return avg_reward_history, epoch_history, trainer, predictor, gym_env