def initialise_policy_storage(self):
return OnlineStorage(args=self.args,
num_steps=self.args.policy_num_steps,
num_processes=self.args.num_processes,
state_dim=self.args.state_dim,
latent_dim=self.args.latent_dim,
belief_dim=self.args.belief_dim,
task_dim=self.args.task_dim,
action_space=self.args.action_space,
hidden_size=self.args.encoder_gru_hidden_size,
normalise_rewards=self.args.norm_rew_for_policy,
)
def initialise_policy_storage(self):
return OnlineStorage(
args=self.args,
num_steps=self.args.policy_num_steps,
num_processes=self.args.num_processes,
state_dim=self.args.state_dim,
latent_dim=0, # use metalearner.py if you want to use the VAE
belief_dim=self.args.belief_dim,
task_dim=self.args.task_dim,
action_space=self.args.action_space,
hidden_size=0,
normalise_rewards=self.args.norm_rew_for_policy,
)
def initialise_policy(self):
# initialise rollout storage for the policy
self.policy_storage = OnlineStorage(
self.args,
self.args.policy_num_steps,
self.args.num_processes,
self.args.obs_dim,
self.args.act_space,
hidden_size=self.args.aggregator_hidden_size,
latent_dim=self.args.latent_dim,
normalise_observations=self.args.norm_obs_for_policy,
normalise_rewards=self.args.norm_rew_for_policy,
)
# initialise policy network
input_dim = self.args.obs_dim * int(
self.args.condition_policy_on_state)
input_dim += (
1 + int(not self.args.sample_embeddings)) * self.args.latent_dim
if hasattr(self.envs.action_space, 'low'):
action_low = self.envs.action_space.low
action_high = self.envs.action_space.high
else:
action_low = action_high = None
policy_net = Policy(
state_dim=input_dim,
action_space=self.args.act_space,
init_std=self.args.policy_init_std,
hidden_layers=self.args.policy_layers,
activation_function=self.args.policy_activation_function,
normalise_actions=self.args.normalise_actions,
action_low=action_low,
action_high=action_high,
).to(device)
# initialise policy trainer
if self.args.policy == 'a2c':
self.policy = A2C(
policy_net,
self.args.policy_value_loss_coef,
self.args.policy_entropy_coef,
optimiser_vae=self.vae.optimiser_vae,
lr=self.args.lr_policy,
eps=self.args.policy_eps,
alpha=self.args.a2c_alpha,
)
elif self.args.policy == 'ppo':
self.policy = PPO(
policy_net,
self.args.policy_value_loss_coef,
self.args.policy_entropy_coef,
optimiser_vae=self.vae.optimiser_vae,
lr=self.args.lr_policy,
eps=self.args.policy_eps,
ppo_epoch=self.args.ppo_num_epochs,
num_mini_batch=self.args.ppo_num_minibatch,
use_huber_loss=self.args.ppo_use_huberloss,
use_clipped_value_loss=self.args.ppo_use_clipped_value_loss,
clip_param=self.args.ppo_clip_param,
)
else:
raise NotImplementedError
class MetaLearner:
"""
Meta-Learner class with the main training loop for variBAD.
"""
def __init__(self, args):
self.args = args
utl.seed(self.args.seed, self.args.deterministic_execution)
# count number of frames and number of meta-iterations
self.frames = 0
self.iter_idx = 0
# initialise tensorboard logger
self.logger = TBLogger(self.args, self.args.exp_label)
# initialise environments
self.envs = make_vec_envs(
env_name=args.env_name,
seed=args.seed,
num_processes=args.num_processes,
gamma=args.policy_gamma,
log_dir=args.agent_log_dir,
device=device,
allow_early_resets=False,
episodes_per_task=self.args.max_rollouts_per_task,
obs_rms=None,
ret_rms=None,
)
# calculate what the maximum length of the trajectories is
args.max_trajectory_len = self.envs._max_episode_steps
args.max_trajectory_len *= self.args.max_rollouts_per_task
# calculate number of meta updates
self.args.num_updates = int(
args.num_frames) // args.policy_num_steps // args.num_processes
# get action / observation dimensions
if isinstance(self.envs.action_space, gym.spaces.discrete.Discrete):
self.args.action_dim = 1
else:
self.args.action_dim = self.envs.action_space.shape[0]
self.args.obs_dim = self.envs.observation_space.shape[0]
self.args.num_states = self.envs.num_states if str.startswith(
self.args.env_name, 'Grid') else None
self.args.act_space = self.envs.action_space
self.vae = VaribadVAE(self.args, self.logger, lambda: self.iter_idx)
self.initialise_policy()
def initialise_policy(self):
# initialise rollout storage for the policy
self.policy_storage = OnlineStorage(
self.args,
self.args.policy_num_steps,
self.args.num_processes,
self.args.obs_dim,
self.args.act_space,
hidden_size=self.args.aggregator_hidden_size,
latent_dim=self.args.latent_dim,
normalise_observations=self.args.norm_obs_for_policy,
normalise_rewards=self.args.norm_rew_for_policy,
)
# initialise policy network
input_dim = self.args.obs_dim * int(
self.args.condition_policy_on_state)
input_dim += (
1 + int(not self.args.sample_embeddings)) * self.args.latent_dim
if hasattr(self.envs.action_space, 'low'):
action_low = self.envs.action_space.low
action_high = self.envs.action_space.high
else:
action_low = action_high = None
policy_net = Policy(
state_dim=input_dim,
action_space=self.args.act_space,
init_std=self.args.policy_init_std,
hidden_layers=self.args.policy_layers,
activation_function=self.args.policy_activation_function,
normalise_actions=self.args.normalise_actions,
action_low=action_low,
action_high=action_high,
).to(device)
# initialise policy trainer
if self.args.policy == 'a2c':
self.policy = A2C(
policy_net,
self.args.policy_value_loss_coef,
self.args.policy_entropy_coef,
optimiser_vae=self.vae.optimiser_vae,
lr=self.args.lr_policy,
eps=self.args.policy_eps,
alpha=self.args.a2c_alpha,
)
elif self.args.policy == 'ppo':
self.policy = PPO(
policy_net,
self.args.policy_value_loss_coef,
self.args.policy_entropy_coef,
optimiser_vae=self.vae.optimiser_vae,
lr=self.args.lr_policy,
eps=self.args.policy_eps,
ppo_epoch=self.args.ppo_num_epochs,
num_mini_batch=self.args.ppo_num_minibatch,
use_huber_loss=self.args.ppo_use_huberloss,
use_clipped_value_loss=self.args.ppo_use_clipped_value_loss,
clip_param=self.args.ppo_clip_param,
)
else:
raise NotImplementedError
def train(self):
"""
Given some stream of environments and a logger (tensorboard),
(meta-)trains the policy.
"""
start_time = time.time()
# reset environments
(prev_obs_raw, prev_obs_normalised) = self.envs.reset()
prev_obs_raw = prev_obs_raw.to(device)
prev_obs_normalised = prev_obs_normalised.to(device)
# insert initial observation / embeddings to rollout storage
self.policy_storage.prev_obs_raw[0].copy_(prev_obs_raw)
self.policy_storage.prev_obs_normalised[0].copy_(prev_obs_normalised)
self.policy_storage.to(device)
vae_is_pretrained = False
for self.iter_idx in range(self.args.num_updates):
# First, re-compute the hidden states given the current rollouts (since the VAE might've changed)
# compute latent embedding (will return prior if current trajectory is empty)
with torch.no_grad():
latent_sample, latent_mean, latent_logvar, hidden_state = self.encode_running_trajectory(
)
# check if we flushed the policy storage
assert len(self.policy_storage.latent_mean) == 0
# add this initial hidden state to the policy storage
self.policy_storage.hidden_states[0].copy_(hidden_state)
self.policy_storage.latent_samples.append(latent_sample.clone())
self.policy_storage.latent_mean.append(latent_mean.clone())
self.policy_storage.latent_logvar.append(latent_logvar.clone())
# rollout policies for a few steps
for step in range(self.args.policy_num_steps):
# sample actions from policy
with torch.no_grad():
value, action, action_log_prob = utl.select_action(
args=self.args,
policy=self.policy,
obs=prev_obs_normalised
if self.args.norm_obs_for_policy else prev_obs_raw,
deterministic=False,
latent_sample=latent_sample,
latent_mean=latent_mean,
latent_logvar=latent_logvar,
)
# observe reward and next obs
(next_obs_raw, next_obs_normalised), (
rew_raw, rew_normalised), done, infos = utl.env_step(
self.envs, action)
tasks = torch.FloatTensor([info['task']
for info in infos]).to(device)
done = torch.from_numpy(np.array(
done, dtype=int)).to(device).float().view((-1, 1))
# create mask for episode ends
masks_done = torch.FloatTensor([[0.0] if done_ else [1.0]
for done_ in done]).to(device)
# bad_mask is true if episode ended because time limit was reached
bad_masks = torch.FloatTensor(
[[0.0] if 'bad_transition' in info.keys() else [1.0]
for info in infos]).to(device)
# compute next embedding (for next loop and/or value prediction bootstrap)
latent_sample, latent_mean, latent_logvar, hidden_state = utl.update_encoding(
encoder=self.vae.encoder,
next_obs=next_obs_raw,
action=action,
reward=rew_raw,
done=done,
hidden_state=hidden_state)
# before resetting, update the embedding and add to vae buffer
# (last state might include useful task info)
if not (self.args.disable_decoder
and self.args.disable_stochasticity_in_latent):
self.vae.rollout_storage.insert(prev_obs_raw.clone(),
action.detach().clone(),
next_obs_raw.clone(),
rew_raw.clone(),
done.clone(),
tasks.clone())
# add the obs before reset to the policy storage
# (only used to recompute embeddings if rlloss is backpropagated through encoder)
self.policy_storage.next_obs_raw[step] = next_obs_raw.clone()
self.policy_storage.next_obs_normalised[
step] = next_obs_normalised.clone()
# reset environments that are done
done_indices = np.argwhere(
done.cpu().detach().flatten()).flatten()
if len(done_indices) == self.args.num_processes:
[next_obs_raw, next_obs_normalised] = self.envs.reset()
if not self.args.sample_embeddings:
latent_sample = latent_sample
else:
for i in done_indices:
[next_obs_raw[i],
next_obs_normalised[i]] = self.envs.reset(index=i)
if not self.args.sample_embeddings:
latent_sample[i] = latent_sample[i]
# # add experience to policy buffer
self.policy_storage.insert(
obs_raw=next_obs_raw,
obs_normalised=next_obs_normalised,
actions=action,
action_log_probs=action_log_prob,
rewards_raw=rew_raw,
rewards_normalised=rew_normalised,
value_preds=value,
masks=masks_done,
bad_masks=bad_masks,
done=done,
hidden_states=hidden_state.squeeze(0).detach(),
latent_sample=latent_sample.detach(),
latent_mean=latent_mean.detach(),
latent_logvar=latent_logvar.detach(),
)
prev_obs_normalised = next_obs_normalised
prev_obs_raw = next_obs_raw
self.frames += self.args.num_processes
# --- UPDATE ---
if self.args.precollect_len <= self.frames:
# check if we are pre-training the VAE
if self.args.pretrain_len > 0 and not vae_is_pretrained:
for _ in range(self.args.pretrain_len):
self.vae.compute_vae_loss(update=True)
vae_is_pretrained = True
# otherwise do the normal update (policy + vae)
else:
train_stats = self.update(
obs=prev_obs_normalised
if self.args.norm_obs_for_policy else prev_obs_raw,
latent_sample=latent_sample,
latent_mean=latent_mean,
latent_logvar=latent_logvar)
# log
run_stats = [action, action_log_prob, value]
if train_stats is not None:
self.log(run_stats, train_stats, start_time)
# clean up after update
self.policy_storage.after_update()
def encode_running_trajectory(self):
"""
(Re-)Encodes (for each process) the entire current trajectory.
Returns sample/mean/logvar and hidden state (if applicable) for the current timestep.
:return:
"""
# for each process, get the current batch (zero-padded obs/act/rew + length indicators)
prev_obs, next_obs, act, rew, lens = self.vae.rollout_storage.get_running_batch(
)
# get embedding - will return (1+sequence_len) * batch * input_size -- includes the prior!
all_latent_samples, all_latent_means, all_latent_logvars, all_hidden_states = self.vae.encoder(
actions=act,
states=next_obs,
rewards=rew,
hidden_state=None,
return_prior=True)
# get the embedding / hidden state of the current time step (need to do this since we zero-padded)
latent_sample = (torch.stack([
all_latent_samples[lens[i]][i] for i in range(len(lens))
])).detach().to(device)
latent_mean = (torch.stack([
all_latent_means[lens[i]][i] for i in range(len(lens))
])).detach().to(device)
latent_logvar = (torch.stack([
all_latent_logvars[lens[i]][i] for i in range(len(lens))
])).detach().to(device)
hidden_state = (torch.stack([
all_hidden_states[lens[i]][i] for i in range(len(lens))
])).detach().to(device)
return latent_sample, latent_mean, latent_logvar, hidden_state
def get_value(self, obs, latent_sample, latent_mean, latent_logvar):
obs = utl.get_augmented_obs(self.args, obs, latent_sample, latent_mean,
latent_logvar)
return self.policy.actor_critic.get_value(obs).detach()
def update(self, obs, latent_sample, latent_mean, latent_logvar):
"""
Meta-update.
Here the policy is updated for good average performance across tasks.
:return:
"""
# update policy (if we are not pre-training, have enough data in the vae buffer, and are not at iteration 0)
if self.iter_idx >= self.args.pretrain_len and self.iter_idx > 0:
# bootstrap next value prediction
with torch.no_grad():
next_value = self.get_value(obs=obs,
latent_sample=latent_sample,
latent_mean=latent_mean,
latent_logvar=latent_logvar)
# compute returns for current rollouts
self.policy_storage.compute_returns(
next_value,
self.args.policy_use_gae,
self.args.policy_gamma,
self.args.policy_tau,
use_proper_time_limits=self.args.use_proper_time_limits)
# update agent (this will also call the VAE update!)
policy_train_stats = self.policy.update(
args=self.args,
policy_storage=self.policy_storage,
encoder=self.vae.encoder,
rlloss_through_encoder=self.args.rlloss_through_encoder,
compute_vae_loss=self.vae.compute_vae_loss)
else:
policy_train_stats = 0, 0, 0, 0
# pre-train the VAE
if self.iter_idx < self.args.pretrain_len:
self.vae.compute_vae_loss(update=True)
return policy_train_stats, None
def log(self, run_stats, train_stats, start_time):
train_stats, meta_train_stats = train_stats
# --- visualise behaviour of policy ---
if self.iter_idx % self.args.vis_interval == 0:
obs_rms = self.envs.venv.obs_rms if self.args.norm_obs_for_policy else None
ret_rms = self.envs.venv.ret_rms if self.args.norm_rew_for_policy else None
utl_eval.visualise_behaviour(
args=self.args,
policy=self.policy,
image_folder=self.logger.full_output_folder,
iter_idx=self.iter_idx,
obs_rms=obs_rms,
ret_rms=ret_rms,
encoder=self.vae.encoder,
reward_decoder=self.vae.reward_decoder,
state_decoder=self.vae.state_decoder,
task_decoder=self.vae.task_decoder,
compute_rew_reconstruction_loss=self.vae.
compute_rew_reconstruction_loss,
compute_state_reconstruction_loss=self.vae.
compute_state_reconstruction_loss,
compute_task_reconstruction_loss=self.vae.
compute_task_reconstruction_loss,
compute_kl_loss=self.vae.compute_kl_loss,
)
# --- evaluate policy ----
if self.iter_idx % self.args.eval_interval == 0:
obs_rms = self.envs.venv.obs_rms if self.args.norm_obs_for_policy else None
ret_rms = self.envs.venv.ret_rms if self.args.norm_rew_for_policy else None
returns_per_episode = utl_eval.evaluate(args=self.args,
policy=self.policy,
obs_rms=obs_rms,
ret_rms=ret_rms,
encoder=self.vae.encoder,
iter_idx=self.iter_idx)
# log the return avg/std across tasks (=processes)
returns_avg = returns_per_episode.mean(dim=0)
returns_std = returns_per_episode.std(dim=0)
for k in range(len(returns_avg)):
self.logger.add('return_avg_per_iter/episode_{}'.format(k + 1),
returns_avg[k], self.iter_idx)
self.logger.add(
'return_avg_per_frame/episode_{}'.format(k + 1),
returns_avg[k], self.frames)
self.logger.add('return_std_per_iter/episode_{}'.format(k + 1),
returns_std[k], self.iter_idx)
self.logger.add(
'return_std_per_frame/episode_{}'.format(k + 1),
returns_std[k], self.frames)
print(
"Updates {}, num timesteps {}, FPS {}, {} \n Mean return (train): {:.5f} \n"
.format(self.iter_idx, self.frames,
int(self.frames / (time.time() - start_time)),
self.vae.rollout_storage.prev_obs.shape,
returns_avg[-1].item()))
# --- save models ---
if self.iter_idx % self.args.save_interval == 0:
save_path = os.path.join(self.logger.full_output_folder, 'models')
if not os.path.exists(save_path):
os.mkdir(save_path)
torch.save(
self.policy.actor_critic,
os.path.join(save_path, "policy{0}.pt".format(self.iter_idx)))
torch.save(
self.vae.encoder,
os.path.join(save_path, "encoder{0}.pt".format(self.iter_idx)))
if self.vae.state_decoder is not None:
torch.save(
self.vae.state_decoder,
os.path.join(save_path,
"state_decoder{0}.pt".format(self.iter_idx)))
if self.vae.reward_decoder is not None:
torch.save(
self.vae.reward_decoder,
os.path.join(save_path,
"reward_decoder{0}.pt".format(self.iter_idx)))
if self.vae.task_decoder is not None:
torch.save(
self.vae.task_decoder,
os.path.join(save_path,
"task_decoder{0}.pt".format(self.iter_idx)))
# save normalisation params of envs
if self.args.norm_rew_for_policy:
# save rolling mean and std
rew_rms = self.envs.venv.ret_rms
utl.save_obj(rew_rms, save_path,
"env_rew_rms{0}.pkl".format(self.iter_idx))
if self.args.norm_obs_for_policy:
obs_rms = self.envs.venv.obs_rms
utl.save_obj(obs_rms, save_path,
"env_obs_rms{0}.pkl".format(self.iter_idx))
# --- log some other things ---
if self.iter_idx % self.args.log_interval == 0:
self.logger.add('policy_losses/value_loss', train_stats[0],
self.iter_idx)
self.logger.add('policy_losses/action_loss', train_stats[1],
self.iter_idx)
self.logger.add('policy_losses/dist_entropy', train_stats[2],
self.iter_idx)
self.logger.add('policy_losses/sum', train_stats[3], self.iter_idx)
self.logger.add('policy/action', run_stats[0][0].float().mean(),
self.iter_idx)
if hasattr(self.policy.actor_critic, 'logstd'):
self.logger.add('policy/action_logstd',
self.policy.actor_critic.dist.logstd.mean(),
self.iter_idx)
self.logger.add('policy/action_logprob', run_stats[1].mean(),
self.iter_idx)
self.logger.add('policy/value', run_stats[2].mean(), self.iter_idx)
self.logger.add('encoder/latent_mean',
torch.cat(self.policy_storage.latent_mean).mean(),
self.iter_idx)
self.logger.add(
'encoder/latent_logvar',
torch.cat(self.policy_storage.latent_logvar).mean(),
self.iter_idx)
# log the average weights and gradients of all models (where applicable)
for [model, name
] in [[self.policy.actor_critic, 'policy'],
[self.vae.encoder, 'encoder'],
[self.vae.reward_decoder, 'reward_decoder'],
[self.vae.state_decoder, 'state_transition_decoder'],
[self.vae.task_decoder, 'task_decoder']]:
if model is not None:
param_list = list(model.parameters())
param_mean = np.mean([
param_list[i].data.cpu().numpy().mean()
for i in range(len(param_list))
])
self.logger.add('weights/{}'.format(name), param_mean,
self.iter_idx)
if name == 'policy':
self.logger.add('weights/policy_std',
param_list[0].data.mean(),
self.iter_idx)
if param_list[0].grad is not None:
param_grad_mean = np.mean([
param_list[i].grad.cpu().numpy().mean()
for i in range(len(param_list))
])
self.logger.add('gradients/{}'.format(name),
param_grad_mean, self.iter_idx)
def load_and_render(self, load_iter):
#save_path = os.path.join('/ext/varibad_github/v2/varibad/logs/logs_HalfCheetahJoint-v0/varibad_73__15:05_17:14:07', 'models')
#save_path = os.path.join('/ext/varibad_github/v2/varibad/logs/hfield', 'models')
save_path = os.path.join(
'/ext/varibad_github/v2/varibad/logs/logs_HalfCheetahBlocks-v0/varibad_73__15:05_20:20:25',
'models')
self.policy.actor_critic = torch.load(
os.path.join(save_path, "policy{0}.pt".format(load_iter)))
self.vae.encoder = torch.load(
os.path.join(save_path, "encoder{0}.pt").format(load_iter))
args = self.args
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
num_processes = 1
num_episodes = 100
num_steps = 1999
#import pdb; pdb.set_trace()
# initialise environments
envs = make_vec_envs(
env_name=args.env_name,
seed=args.seed,
num_processes=num_processes, # 1
gamma=args.policy_gamma,
log_dir=args.agent_log_dir,
device=device,
allow_early_resets=False,
episodes_per_task=self.args.max_rollouts_per_task,
obs_rms=None,
ret_rms=None,
)
# reset latent state to prior
latent_sample, latent_mean, latent_logvar, hidden_state = self.vae.encoder.prior(
num_processes)
for episode_idx in range(num_episodes):
(prev_obs_raw, prev_obs_normalised) = envs.reset()
prev_obs_raw = prev_obs_raw.to(device)
prev_obs_normalised = prev_obs_normalised.to(device)
for step_idx in range(num_steps):
with torch.no_grad():
_, action, _ = utl.select_action(
args=self.args,
policy=self.policy,
obs=prev_obs_normalised
if self.args.norm_obs_for_policy else prev_obs_raw,
latent_sample=latent_sample,
latent_mean=latent_mean,
latent_logvar=latent_logvar,
deterministic=True)
# observe reward and next obs
(next_obs_raw, next_obs_normalised), (
rew_raw,
rew_normalised), done, infos = utl.env_step(envs, action)
# render
envs.venv.venv.envs[0].env.env.env.env.render()
# update the hidden state
latent_sample, latent_mean, latent_logvar, hidden_state = utl.update_encoding(
encoder=self.vae.encoder,
next_obs=next_obs_raw,
action=action,
reward=rew_raw,
done=None,
hidden_state=hidden_state)
prev_obs_normalised = next_obs_normalised
prev_obs_raw = next_obs_raw
if done[0]:
break
def initialise_policy(self):
# variables for task encoder (used for oracle)
state_dim = self.envs.observation_space.shape[0]
# TODO: this isn't ideal, find a nicer way to get the task dimension!
if 'BeliefOracle' in self.args.env_name:
task_dim = gym.make(self.args.env_name).observation_space.shape[0] - \
gym.make(self.args.env_name.replace('BeliefOracle', '')).observation_space.shape[0]
latent_dim = self.args.latent_dim
state_embedding_size = self.args.state_embedding_size
use_task_encoder = True
elif 'Oracle' in self.args.env_name:
task_dim = gym.make(self.args.env_name).observation_space.shape[0] - \
gym.make(self.args.env_name.replace('Oracle', '')).observation_space.shape[0]
latent_dim = self.args.latent_dim
state_embedding_size = self.args.state_embedding_size
use_task_encoder = True
else:
task_dim = latent_dim = state_embedding_size = 0
use_task_encoder = False
# initialise rollout storage for the policy
self.policy_storage = OnlineStorage(
self.args,
self.args.policy_num_steps,
self.args.num_processes,
self.args.obs_dim,
self.args.act_space,
hidden_size=0,
latent_dim=self.args.latent_dim,
normalise_observations=self.args.norm_obs_for_policy,
normalise_rewards=self.args.norm_rew_for_policy,
)
if hasattr(self.envs.action_space, 'low'):
action_low = self.envs.action_space.low
action_high = self.envs.action_space.high
else:
action_low = action_high = None
# initialise policy network
policy_net = Policy(
# general
state_dim=int(self.args.condition_policy_on_state) * state_dim,
action_space=self.envs.action_space,
init_std=self.args.policy_init_std,
hidden_layers=self.args.policy_layers,
activation_function=self.args.policy_activation_function,
use_task_encoder=use_task_encoder,
# task encoding things (for oracle)
task_dim=task_dim,
latent_dim=latent_dim,
state_embed_dim=state_embedding_size,
#
normalise_actions=self.args.normalise_actions,
action_low=action_low,
action_high=action_high,
).to(device)
# initialise policy
if self.args.policy == 'a2c':
# initialise policy trainer (A2C)
self.policy = A2C(
policy_net,
self.args.policy_value_loss_coef,
self.args.policy_entropy_coef,
lr=self.args.lr_policy,
eps=self.args.policy_eps,
alpha=self.args.a2c_alpha,
)
elif self.args.policy == 'ppo':
# initialise policy network
self.policy = PPO(
policy_net,
self.args.policy_value_loss_coef,
self.args.policy_entropy_coef,
lr=self.args.lr_policy,
eps=self.args.policy_eps,
ppo_epoch=self.args.ppo_num_epochs,
num_mini_batch=self.args.ppo_num_minibatch,
use_huber_loss=self.args.ppo_use_huberloss,
use_clipped_value_loss=self.args.ppo_use_clipped_value_loss,
clip_param=self.args.ppo_clip_param,
)
else:
raise NotImplementedError
class Learner:
"""
Learner (no meta-learning), can be used to train Oracle policies.
"""
def __init__(self, args):
self.args = args
# make sure everything has the same seed
utl.seed(self.args.seed, self.args.deterministic_execution)
# initialise tensorboard logger
self.logger = TBLogger(self.args, self.args.exp_label)
# initialise environments
self.envs = make_vec_envs(
env_name=args.env_name,
seed=args.seed,
num_processes=args.num_processes,
gamma=args.policy_gamma,
log_dir=args.agent_log_dir,
device=device,
allow_early_resets=False,
episodes_per_task=self.args.max_rollouts_per_task,
obs_rms=None,
ret_rms=None,
)
# calculate what the maximum length of the trajectories is
args.max_trajectory_len = self.envs._max_episode_steps
args.max_trajectory_len *= self.args.max_rollouts_per_task
# calculate number of meta updates
self.args.num_updates = int(
args.num_frames) // args.policy_num_steps // args.num_processes
# get action / observation dimensions
if isinstance(self.envs.action_space, gym.spaces.discrete.Discrete):
self.args.action_dim = 1
else:
self.args.action_dim = self.envs.action_space.shape[0]
self.args.obs_dim = self.envs.observation_space.shape[0]
self.args.num_states = self.envs.num_states if str.startswith(
self.args.env_name, 'Grid') else None
self.args.act_space = self.envs.action_space
self.initialise_policy()
# count number of frames and updates
self.frames = 0
self.iter_idx = 0
def initialise_policy(self):
# variables for task encoder (used for oracle)
state_dim = self.envs.observation_space.shape[0]
# TODO: this isn't ideal, find a nicer way to get the task dimension!
if 'BeliefOracle' in self.args.env_name:
task_dim = gym.make(self.args.env_name).observation_space.shape[0] - \
gym.make(self.args.env_name.replace('BeliefOracle', '')).observation_space.shape[0]
latent_dim = self.args.latent_dim
state_embedding_size = self.args.state_embedding_size
use_task_encoder = True
elif 'Oracle' in self.args.env_name:
task_dim = gym.make(self.args.env_name).observation_space.shape[0] - \
gym.make(self.args.env_name.replace('Oracle', '')).observation_space.shape[0]
latent_dim = self.args.latent_dim
state_embedding_size = self.args.state_embedding_size
use_task_encoder = True
else:
task_dim = latent_dim = state_embedding_size = 0
use_task_encoder = False
# initialise rollout storage for the policy
self.policy_storage = OnlineStorage(
self.args,
self.args.policy_num_steps,
self.args.num_processes,
self.args.obs_dim,
self.args.act_space,
hidden_size=0,
latent_dim=self.args.latent_dim,
normalise_observations=self.args.norm_obs_for_policy,
normalise_rewards=self.args.norm_rew_for_policy,
)
if hasattr(self.envs.action_space, 'low'):
action_low = self.envs.action_space.low
action_high = self.envs.action_space.high
else:
action_low = action_high = None
# initialise policy network
policy_net = Policy(
# general
state_dim=int(self.args.condition_policy_on_state) * state_dim,
action_space=self.envs.action_space,
init_std=self.args.policy_init_std,
hidden_layers=self.args.policy_layers,
activation_function=self.args.policy_activation_function,
use_task_encoder=use_task_encoder,
# task encoding things (for oracle)
task_dim=task_dim,
latent_dim=latent_dim,
state_embed_dim=state_embedding_size,
#
normalise_actions=self.args.normalise_actions,
action_low=action_low,
action_high=action_high,
).to(device)
# initialise policy
if self.args.policy == 'a2c':
# initialise policy trainer (A2C)
self.policy = A2C(
policy_net,
self.args.policy_value_loss_coef,
self.args.policy_entropy_coef,
lr=self.args.lr_policy,
eps=self.args.policy_eps,
alpha=self.args.a2c_alpha,
)
elif self.args.policy == 'ppo':
# initialise policy network
self.policy = PPO(
policy_net,
self.args.policy_value_loss_coef,
self.args.policy_entropy_coef,
lr=self.args.lr_policy,
eps=self.args.policy_eps,
ppo_epoch=self.args.ppo_num_epochs,
num_mini_batch=self.args.ppo_num_minibatch,
use_huber_loss=self.args.ppo_use_huberloss,
use_clipped_value_loss=self.args.ppo_use_clipped_value_loss,
clip_param=self.args.ppo_clip_param,
)
else:
raise NotImplementedError
def train(self):
"""
Given some stream of environments and a logger (tensorboard),
(meta-)trains the policy.
"""
start_time = time.time()
# reset environments
(prev_obs_raw, prev_obs_normalised) = self.envs.reset()
prev_obs_raw = prev_obs_raw.to(device)
prev_obs_normalised = prev_obs_normalised.to(device)
# insert initial observation / embeddings to rollout storage
self.policy_storage.prev_obs_raw[0].copy_(prev_obs_raw)
self.policy_storage.prev_obs_normalised[0].copy_(prev_obs_normalised)
self.policy_storage.to(device)
for self.iter_idx in range(self.args.num_updates):
# check if we flushed the policy storage
assert len(self.policy_storage.latent_mean) == 0
# rollouts policies for a few steps
for step in range(self.args.policy_num_steps):
# sample actions from policy
with torch.no_grad():
value, action, action_log_prob = utl.select_action(
policy=self.policy,
args=self.args,
obs=prev_obs_normalised
if self.args.norm_obs_for_policy else prev_obs_raw,
deterministic=False)
# observe reward and next obs
(next_obs_raw, next_obs_normalised), (
rew_raw, rew_normalised), done, infos = utl.env_step(
self.envs, action)
action = action.float()
# create mask for episode ends
masks_done = torch.FloatTensor([[0.0] if done_ else [1.0]
for done_ in done]).to(device)
# bad_mask is true if episode ended because time limit was reached
bad_masks = torch.FloatTensor(
[[0.0] if 'bad_transition' in info.keys() else [1.0]
for info in infos]).to(device)
# add the obs before reset to the policy storage
self.policy_storage.next_obs_raw[step] = next_obs_raw.clone()
self.policy_storage.next_obs_normalised[
step] = next_obs_normalised.clone()
# reset environments that are done
done_indices = np.argwhere(done.flatten()).flatten()
if len(done_indices) == self.args.num_processes:
[next_obs_raw, next_obs_normalised] = self.envs.reset()
if not self.args.sample_embeddings:
latent_sample = latent_sample
else:
for i in done_indices:
[next_obs_raw[i],
next_obs_normalised[i]] = self.envs.reset(index=i)
if not self.args.sample_embeddings:
latent_sample[i] = latent_sample[i]
# add experience to policy buffer
self.policy_storage.insert(
obs_raw=next_obs_raw.clone(),
obs_normalised=next_obs_normalised.clone(),
actions=action.clone(),
action_log_probs=action_log_prob.clone(),
rewards_raw=rew_raw.clone(),
rewards_normalised=rew_normalised.clone(),
value_preds=value.clone(),
masks=masks_done.clone(),
bad_masks=bad_masks.clone(),
done=torch.from_numpy(np.array(
done, dtype=float)).unsqueeze(1).clone(),
)
prev_obs_normalised = next_obs_normalised
prev_obs_raw = next_obs_raw
self.frames += self.args.num_processes
# --- UPDATE ---
train_stats = self.update(prev_obs_normalised if self.args.
norm_obs_for_policy else prev_obs_raw)
# log
run_stats = [action, action_log_prob, value]
if train_stats is not None:
self.log(run_stats, train_stats, start_time)
# clean up after update
self.policy_storage.after_update()
def get_value(self, obs):
obs = utl.get_augmented_obs(args=self.args, obs=obs)
return self.policy.actor_critic.get_value(obs).detach()
def update(self, obs):
"""
Meta-update.
Here the policy is updated for good average performance across tasks.
:return: policy_train_stats which are: value_loss_epoch, action_loss_epoch, dist_entropy_epoch, loss_epoch
"""
# bootstrap next value prediction
with torch.no_grad():
next_value = self.get_value(obs)
# compute returns for current rollouts
self.policy_storage.compute_returns(
next_value,
self.args.policy_use_gae,
self.args.policy_gamma,
self.args.policy_tau,
use_proper_time_limits=self.args.use_proper_time_limits)
policy_train_stats = self.policy.update(
args=self.args, policy_storage=self.policy_storage)
return policy_train_stats, None
def log(self, run_stats, train_stats, start):
"""
Evaluate policy, save model, write to tensorboard logger.
"""
train_stats, meta_train_stats = train_stats
# --- visualise behaviour of policy ---
if self.iter_idx % self.args.vis_interval == 0:
obs_rms = self.envs.venv.obs_rms if self.args.norm_obs_for_policy else None
ret_rms = self.envs.venv.ret_rms if self.args.norm_rew_for_policy else None
utl_eval.visualise_behaviour(
args=self.args,
policy=self.policy,
image_folder=self.logger.full_output_folder,
iter_idx=self.iter_idx,
obs_rms=obs_rms,
ret_rms=ret_rms,
)
# --- evaluate policy ----
if self.iter_idx % self.args.eval_interval == 0:
obs_rms = self.envs.venv.obs_rms if self.args.norm_obs_for_policy else None
ret_rms = self.envs.venv.ret_rms if self.args.norm_rew_for_policy else None
returns_per_episode = utl_eval.evaluate(args=self.args,
policy=self.policy,
obs_rms=obs_rms,
ret_rms=ret_rms,
iter_idx=self.iter_idx)
# log the average return across tasks (=processes)
returns_avg = returns_per_episode.mean(dim=0)
returns_std = returns_per_episode.std(dim=0)
for k in range(len(returns_avg)):
self.logger.add('return_avg_per_iter/episode_{}'.format(k + 1),
returns_avg[k], self.iter_idx)
self.logger.add(
'return_avg_per_frame/episode_{}'.format(k + 1),
returns_avg[k], self.frames)
self.logger.add('return_std_per_iter/episode_{}'.format(k + 1),
returns_std[k], self.iter_idx)
self.logger.add(
'return_std_per_frame/episode_{}'.format(k + 1),
returns_std[k], self.frames)
print(
"Updates {}, num timesteps {}, FPS {} \n Mean return (train): {:.5f} \n"
.format(self.iter_idx, self.frames,
int(self.frames / (time.time() - start)),
returns_avg[-1].item()))
# save model
if self.iter_idx % self.args.save_interval == 0:
save_path = os.path.join(self.logger.full_output_folder, 'models')
if not os.path.exists(save_path):
os.mkdir(save_path)
torch.save(
self.policy.actor_critic,
os.path.join(save_path, "policy{0}.pt".format(self.iter_idx)))
# save normalisation params of envs
if self.args.norm_rew_for_policy:
# save rolling mean and std
rew_rms = self.envs.venv.ret_rms
utl.save_obj(rew_rms, save_path,
"env_rew_rms{0}.pkl".format(self.iter_idx))
if self.args.norm_obs_for_policy:
obs_rms = self.envs.venv.obs_rms
utl.save_obj(obs_rms, save_path,
"env_obs_rms{0}.pkl".format(self.iter_idx))
# --- log some other things ---
if self.iter_idx % self.args.log_interval == 0:
self.logger.add('policy_losses/value_loss', train_stats[0],
self.iter_idx)
self.logger.add('policy_losses/action_loss', train_stats[1],
self.iter_idx)
self.logger.add('policy_losses/dist_entropy', train_stats[2],
self.iter_idx)
self.logger.add('policy_losses/sum', train_stats[3], self.iter_idx)
# writer.add_scalar('policy/action', action.mean(), j)
self.logger.add('policy/action', run_stats[0][0].float().mean(),
self.iter_idx)
if hasattr(self.policy.actor_critic, 'logstd'):
self.logger.add('policy/action_logstd',
self.policy.actor_critic.dist.logstd.mean(),
self.iter_idx)
self.logger.add('policy/action_logprob', run_stats[1].mean(),
self.iter_idx)
self.logger.add('policy/value', run_stats[2].mean(), self.iter_idx)
param_list = list(self.policy.actor_critic.parameters())
param_mean = np.mean(
[param_list[i].data.mean() for i in range(len(param_list))])
param_grad_mean = np.mean(
[param_list[i].grad.mean() for i in range(len(param_list))])
self.logger.add('weights/policy', param_mean, self.iter_idx)
self.logger.add('weights/policy_std', param_list[0].data.mean(),
self.iter_idx)
self.logger.add('gradients/policy', param_grad_mean, self.iter_idx)
self.logger.add('gradients/policy_std', param_list[0].grad.mean(),
self.iter_idx)