in rlalgos/a2c_gae/a2c.py [0:0]
def run(self):
# Instantiate the learning model abd the baseline model
self.learning_model = self._create_model()
self.agent = self._create_agent(
n_actions=self.n_actions, model=self.learning_model
)
# We create a batcher dedicated to evaluation
model = copy.deepcopy(self.learning_model)
self.evaluation_batcher = RL_Batcher(
n_timesteps=self.config["max_episode_steps"],
create_agent=self._create_agent,
create_env=self._create_env,
env_args={
"n_envs": self.config["n_evaluation_envs"],
"max_episode_steps": self.config["max_episode_steps"],
"env_name": self.config["env_name"],
},
agent_args={"n_actions": self.n_actions, "model": model},
n_processes=self.config["n_evaluation_processes"],
seeds=[
self.config["env_seed"] + k * 10
for k in range(self.config["n_evaluation_processes"])
],
agent_info=DictTensor({"stochastic": torch.tensor([True])}),
env_info=DictTensor({}),
)
# Creation of the batcher for sampling complete pieces of trajectories (i.e Batcher)
# The batcher will sample n_threads*n_envs trajectories at each call
# To have a fast batcher, we have to configure it with n_timesteps=self.config["max_episode_steps"]
model = copy.deepcopy(self.learning_model)
self.train_batcher = RL_Batcher(
n_timesteps=self.config["a2c_timesteps"],
create_agent=self._create_agent,
create_env=self._create_train_env,
env_args={
"n_envs": self.config["n_envs"],
"max_episode_steps": self.config["max_episode_steps"],
"env_name": self.config["env_name"],
},
agent_args={"n_actions": self.n_actions, "model": model},
n_processes=self.config["n_processes"],
seeds=[
self.config["env_seed"] + k * 10
for k in range(self.config["n_processes"])
],
agent_info=DictTensor({"stochastic": torch.tensor([True])}),
env_info=DictTensor({}),
)
# Creation of the optimizer
self.learning_model.to(self.config["learner_device"])
optimizer = getattr(torch.optim, self.config["optim"])(
self.learning_model.parameters(), lr=self.config["lr"]
)
# Training Loop:
_start_time = time.time()
self.iteration = 0
# #We launch the evaluation batcher (in deterministic mode)
n_episodes = (
self.config["n_evaluation_processes"] * self.config["n_evaluation_envs"]
)
agent_info = DictTensor(
{"stochastic": torch.tensor([False]).repeat(n_episodes)}
)
self.evaluation_batcher.reset(agent_info=agent_info)
self.evaluation_batcher.execute()
self.evaluation_iteration = self.iteration
# Initialize the training batcher such that agents will start to acqire pieces of episodes
self.train_batcher.update(self._state_dict(self.learning_model))
n_episodes = self.config["n_envs"] * self.config["n_processes"]
agent_info = DictTensor({"stochastic": torch.tensor([True]).repeat(n_episodes)})
self.train_batcher.reset(agent_info=agent_info)
while time.time() - _start_time < self.config["time_limit"]:
self.train_batcher.execute()
trajectories, n = self.train_batcher.get(blocking=True)
assert n == self.config["n_envs"] * self.config["n_processes"]
# print(trajectories.trajectories["_observation/reward"].sum(1))
dt = self.get_loss(trajectories)
[
self.logger.add_scalar("loss/" + k, dt[k].item(), self.iteration)
for k in dt.keys()
]
ld = self.config["critic_coef"] * dt["critic_loss"]
lr = self.config["a2c_coef"] * dt["a2c_loss"]
le = self.config["entropy_coef"] * dt["entropy_loss"]
floss = ld - le - lr
optimizer.zero_grad()
floss.backward()
if self.config["clip_grad"] > 0:
n = torch.nn.utils.clip_grad_norm_(
self.learning_model.parameters(), self.config["clip_grad"]
)
self.logger.add_scalar("grad_norm", n.item(), self.iteration)
optimizer.step()
# Update the train batcher with the updated model
self.train_batcher.update(self._state_dict(self.learning_model))
self.iteration += 1
# We check the evaluation batcher
evaluation_trajectories, n = self.evaluation_batcher.get(blocking=False)
if not evaluation_trajectories is None: # trajectories are available
# Compute the cumulated reward
cumulated_reward = (
(
evaluation_trajectories.trajectories["_observation/reward"]
* evaluation_trajectories.trajectories.mask()
)
.sum(1)
.mean()
)
self.logger.add_scalar(
"evaluation_reward",
cumulated_reward.item(),
self.evaluation_iteration,
)
# print("At iteration %d, reward is %f"%(self.evaluation_iteration,cumulated_reward.item()))
# We reexecute the evaluation batcher (with same value of agent_info and same number of episodes)
self.evaluation_batcher.update(self._state_dict(self.learning_model))
self.evaluation_iteration = self.iteration
n_episodes = (
self.config["n_evaluation_processes"]
* self.config["n_evaluation_envs"]
)
agent_info = DictTensor(
{"stochastic": torch.tensor([False]).repeat(n_episodes)}
)
self.evaluation_batcher.reset(agent_info=agent_info)
self.evaluation_batcher.execute()
if time.time() - _start_time > 600 and self.iteration % 1000 == 0:
self.logger.update_csv()
self.train_batcher.close()
self.evaluation_batcher.get() # To wait for the last trajectories
self.evaluation_batcher.close()
self.logger.update_csv() # To save as a CSV file in logdir
self.logger.close()