def __init__(self, args):
self.args = args
utl.seed(self.args.seed, self.args.deterministic_execution)
# calculate number of updates and keep count of frames/iterations
self.num_updates = int(
args.num_frames) // args.policy_num_steps // args.num_processes
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,
device=device,
episodes_per_task=self.args.max_rollouts_per_task,
normalise_rew=args.norm_rew_for_policy,
ret_rms=None,
)
# calculate what the maximum length of the trajectories is
self.args.max_trajectory_len = self.envs._max_episode_steps
self.args.max_trajectory_len *= self.args.max_rollouts_per_task
# get policy input dimensions
self.args.state_dim = self.envs.observation_space.shape[0]
self.args.task_dim = self.envs.task_dim
self.args.belief_dim = self.envs.belief_dim
self.args.num_states = self.envs.num_states
# get policy output (action) dimensions
self.args.action_space = self.envs.action_space
if isinstance(self.envs.action_space, gym.spaces.discrete.Discrete):
self.args.action_dim = 1
elif isinstance(self.envs.action_space,
gym.spaces.multi_discrete.MultiDiscrete):
self.args.action_dim = self.envs.action_space.nvec[0]
else:
self.args.action_dim = self.envs.action_space.shape[0]
# initialise VAE and policy
self.vae = VaribadVAE(self.args, self.logger, lambda: self.iter_idx)
self.policy_storage = self.initialise_policy_storage()
self.policy = self.initialise_policy()
def __init__(self, args):
"""
Seeds everything.
Initialises: logger, environments, policy (+storage +optimiser).
"""
self.args = args
# make sure everything has the same seed
utl.seed(self.args.seed)
# initialize tensorboard logger
if self.args.log_tensorboard:
self.tb_logger = TBLogger(self.args)
# initialise environment
self.env = make_env(self.args.env_name,
self.args.max_rollouts_per_task,
seed=self.args.seed,
n_tasks=self.args.num_tasks)
# unwrapped env to get some info about the environment
unwrapped_env = self.env.unwrapped
# split to train/eval tasks
shuffled_tasks = np.random.permutation(
unwrapped_env.get_all_task_idx())
self.train_tasks = shuffled_tasks[:self.args.num_train_tasks]
if self.args.num_eval_tasks > 0:
self.eval_tasks = shuffled_tasks[-self.args.num_eval_tasks:]
else:
self.eval_tasks = []
# calculate what the maximum length of the trajectories is
args.max_trajectory_len = unwrapped_env._max_episode_steps
args.max_trajectory_len *= self.args.max_rollouts_per_task
self.args.max_trajectory_len = args.max_trajectory_len
# get action / observation dimensions
if isinstance(self.env.action_space, gym.spaces.discrete.Discrete):
self.args.action_dim = 1
else:
self.args.action_dim = self.env.action_space.shape[0]
self.args.obs_dim = self.env.observation_space.shape[0]
self.args.num_states = unwrapped_env.num_states if hasattr(
unwrapped_env, 'num_states') else None
self.args.act_space = self.env.action_space
# initialize policy
self.initialize_policy()
# initialize buffer for RL updates
self.policy_storage = MultiTaskPolicyStorage(
max_replay_buffer_size=int(self.args.policy_buffer_size),
obs_dim=self._get_augmented_obs_dim(),
action_space=self.env.action_space,
tasks=self.train_tasks,
trajectory_len=args.max_trajectory_len,
)
self.current_experience_storage = None
self.args.belief_reward = False # initialize arg to not use belief rewards
def __init__(self, args):
"""
Seeds everything.
Initialises: logger, environments, policy (+storage +optimiser).
"""
self.args = args
# make sure everything has the same seed
utl.seed(self.args.seed)
# initialize tensorboard logger
if self.args.log_tensorboard:
self.tb_logger = TBLogger(self.args)
self.args, env = off_utl.expand_args(self.args, include_act_space=True)
if self.args.act_space.__class__.__name__ == "Discrete":
self.args.policy = 'dqn'
else:
self.args.policy = 'sac'
# load buffers with data
if 'load_data' not in self.args or self.args.load_data:
goals, augmented_obs_dim = self.load_buffer(
env) # env is input just for possible relabelling option
self.args.augmented_obs_dim = augmented_obs_dim
self.goals = goals
# initialize policy
self.initialize_policy()
# load vae for inference in evaluation
self.load_vae()
# create environment for evaluation
self.env = make_env(
args.env_name,
args.max_rollouts_per_task,
presampled_tasks=args.presampled_tasks,
seed=args.seed,
)
# n_tasks=self.args.num_eval_tasks)
if self.args.env_name == 'GridNavi-v2':
self.env.unwrapped.goals = [
tuple(goal.astype(int)) for goal in self.goals
]
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()
class OfflineMetaLearner:
"""
Off-line Meta-Learner class, a.k.a no interaction with env.
"""
def __init__(self, args):
"""
Seeds everything.
Initialises: logger, environments, policy (+storage +optimiser).
"""
self.args = args
# make sure everything has the same seed
utl.seed(self.args.seed)
# initialize tensorboard logger
if self.args.log_tensorboard:
self.tb_logger = TBLogger(self.args)
self.args, env = off_utl.expand_args(self.args, include_act_space=True)
if self.args.act_space.__class__.__name__ == "Discrete":
self.args.policy = 'dqn'
else:
self.args.policy = 'sac'
# load buffers with data
if 'load_data' not in self.args or self.args.load_data:
goals, augmented_obs_dim = self.load_buffer(
env) # env is input just for possible relabelling option
self.args.augmented_obs_dim = augmented_obs_dim
self.goals = goals
# initialize policy
self.initialize_policy()
# load vae for inference in evaluation
self.load_vae()
# create environment for evaluation
self.env = make_env(
args.env_name,
args.max_rollouts_per_task,
presampled_tasks=args.presampled_tasks,
seed=args.seed,
)
# n_tasks=self.args.num_eval_tasks)
if self.args.env_name == 'GridNavi-v2':
self.env.unwrapped.goals = [
tuple(goal.astype(int)) for goal in self.goals
]
def initialize_policy(self):
if self.args.policy == 'dqn':
q_network = FlattenMlp(input_size=self.args.augmented_obs_dim,
output_size=self.args.act_space.n,
hidden_sizes=self.args.dqn_layers).to(
ptu.device)
self.agent = DQN(
q_network,
# optimiser_vae=self.optimizer_vae,
lr=self.args.policy_lr,
gamma=self.args.gamma,
tau=self.args.soft_target_tau,
).to(ptu.device)
else:
# assert self.args.act_space.__class__.__name__ == "Box", (
# "Can't train SAC with discrete action space!")
q1_network = FlattenMlp(
input_size=self.args.augmented_obs_dim + self.args.action_dim,
output_size=1,
hidden_sizes=self.args.dqn_layers).to(ptu.device)
q2_network = FlattenMlp(
input_size=self.args.augmented_obs_dim + self.args.action_dim,
output_size=1,
hidden_sizes=self.args.dqn_layers).to(ptu.device)
policy = TanhGaussianPolicy(
obs_dim=self.args.augmented_obs_dim,
action_dim=self.args.action_dim,
hidden_sizes=self.args.policy_layers).to(ptu.device)
self.agent = SAC(
policy,
q1_network,
q2_network,
actor_lr=self.args.actor_lr,
critic_lr=self.args.critic_lr,
gamma=self.args.gamma,
tau=self.args.soft_target_tau,
use_cql=self.args.use_cql if 'use_cql' in self.args else False,
alpha_cql=self.args.alpha_cql
if 'alpha_cql' in self.args else None,
entropy_alpha=self.args.entropy_alpha,
automatic_entropy_tuning=self.args.automatic_entropy_tuning,
alpha_lr=self.args.alpha_lr,
clip_grad_value=self.args.clip_grad_value,
).to(ptu.device)
def load_vae(self):
self.vae = VAE(self.args)
vae_models_path = os.path.join(self.args.vae_dir, self.args.env_name,
self.args.vae_model_name, 'models')
off_utl.load_trained_vae(self.vae, vae_models_path)
def load_buffer(self, env):
if self.args.hindsight_relabelling: # without arr_type loading -- GPU will explode
dataset, goals = off_utl.load_dataset(
data_dir=self.args.relabelled_data_dir,
args=self.args,
num_tasks=self.args.num_train_tasks,
allow_dense_data_loading=False,
arr_type='numpy')
dataset = off_utl.batch_to_trajectories(dataset, self.args)
dataset, goals = off_utl.mix_task_rollouts(
dataset, env, goals, self.args) # reward relabelling
dataset = off_utl.trajectories_to_batch(dataset)
else:
dataset, goals = off_utl.load_dataset(
data_dir=self.args.relabelled_data_dir,
args=self.args,
num_tasks=self.args.num_train_tasks,
allow_dense_data_loading=False,
arr_type='numpy')
augmented_obs_dim = dataset[0][0].shape[1]
self.storage = MultiTaskPolicyStorage(
max_replay_buffer_size=max([d[0].shape[0] for d in dataset]),
obs_dim=dataset[0][0].shape[1],
action_space=self.args.act_space,
tasks=range(len(goals)),
trajectory_len=self.args.trajectory_len)
for task, set in enumerate(dataset):
self.storage.add_samples(task,
observations=set[0],
actions=set[1],
rewards=set[2],
next_observations=set[3],
terminals=set[4])
return goals, augmented_obs_dim
def train(self):
self._start_training()
for iter_ in range(self.args.num_iters):
self.training_mode(True)
indices = np.random.choice(len(self.goals), self.args.meta_batch)
train_stats = self.update(indices)
self.training_mode(False)
self.log(iter_ + 1, train_stats)
def update(self, tasks):
rl_losses_agg = {}
for update in range(self.args.rl_updates_per_iter):
# sample random RL batch
obs, actions, rewards, next_obs, terms = self.sample_rl_batch(
tasks, self.args.batch_size)
# flatten out task dimension
t, b, _ = obs.size()
obs = obs.view(t * b, -1)
actions = actions.view(t * b, -1)
rewards = rewards.view(t * b, -1)
next_obs = next_obs.view(t * b, -1)
terms = terms.view(t * b, -1)
# RL update
rl_losses = self.agent.update(obs,
actions,
rewards,
next_obs,
terms,
action_space=self.env.action_space)
for k, v in rl_losses.items():
if update == 0: # first iterate - create list
rl_losses_agg[k] = [v]
else: # append values
rl_losses_agg[k].append(v)
# take mean
for k in rl_losses_agg:
rl_losses_agg[k] = np.mean(rl_losses_agg[k])
self._n_rl_update_steps_total += self.args.rl_updates_per_iter
return rl_losses_agg
def evaluate(self):
num_episodes = self.args.max_rollouts_per_task
num_steps_per_episode = self.env.unwrapped._max_episode_steps
num_tasks = self.args.num_eval_tasks
obs_size = self.env.unwrapped.observation_space.shape[0]
returns_per_episode = np.zeros((num_tasks, num_episodes))
success_rate = np.zeros(num_tasks)
rewards = np.zeros((num_tasks, self.args.trajectory_len))
reward_preds = np.zeros((num_tasks, self.args.trajectory_len))
observations = np.zeros(
(num_tasks, self.args.trajectory_len + 1, obs_size))
if self.args.policy == 'sac':
log_probs = np.zeros((num_tasks, self.args.trajectory_len))
# This part is very specific for the Semi-Circle env
# if self.args.env_name == 'PointRobotSparse-v0':
# reward_belief = np.zeros((num_tasks, self.args.trajectory_len))
#
# low_x, high_x, low_y, high_y = -2., 2., -1., 2.
# resolution = 0.1
# grid_x = np.arange(low_x, high_x + resolution, resolution)
# grid_y = np.arange(low_y, high_y + resolution, resolution)
# centers_x = (grid_x[:-1] + grid_x[1:]) / 2
# centers_y = (grid_y[:-1] + grid_y[1:]) / 2
# yv, xv = np.meshgrid(centers_y, centers_x, sparse=False, indexing='ij')
# centers = np.vstack([xv.ravel(), yv.ravel()]).T
# n_grid_points = centers.shape[0]
# reward_belief_discretized = np.zeros((num_tasks, self.args.trajectory_len, centers.shape[0]))
for task_loop_i, task in enumerate(
self.env.unwrapped.get_all_eval_task_idx()):
obs = ptu.from_numpy(self.env.reset(task))
obs = obs.reshape(-1, obs.shape[-1])
step = 0
# get prior parameters
with torch.no_grad():
task_sample, task_mean, task_logvar, hidden_state = self.vae.encoder.prior(
batch_size=1)
observations[task_loop_i,
step, :] = ptu.get_numpy(obs[0, :obs_size])
for episode_idx in range(num_episodes):
running_reward = 0.
for step_idx in range(num_steps_per_episode):
# add distribution parameters to observation - policy is conditioned on posterior
augmented_obs = self.get_augmented_obs(
obs, task_mean, task_logvar)
if self.args.policy == 'dqn':
action, value = self.agent.act(obs=augmented_obs,
deterministic=True)
else:
action, _, _, log_prob = self.agent.act(
obs=augmented_obs,
deterministic=self.args.eval_deterministic,
return_log_prob=True)
# observe reward and next obs
next_obs, reward, done, info = utl.env_step(
self.env, action.squeeze(dim=0))
running_reward += reward.item()
# done_rollout = False if ptu.get_numpy(done[0][0]) == 0. else True
# update encoding
task_sample, task_mean, task_logvar, hidden_state = self.update_encoding(
obs=next_obs,
action=action,
reward=reward,
done=done,
hidden_state=hidden_state)
rewards[task_loop_i, step] = reward.item()
reward_preds[task_loop_i, step] = ptu.get_numpy(
self.vae.reward_decoder(task_sample, next_obs, obs,
action)[0, 0])
# This part is very specific for the Semi-Circle env
# if self.args.env_name == 'PointRobotSparse-v0':
# reward_belief[task, step] = ptu.get_numpy(
# self.vae.compute_belief_reward(task_mean, task_logvar, obs, next_obs, action)[0])
#
# reward_belief_discretized[task, step, :] = ptu.get_numpy(
# self.vae.compute_belief_reward(task_mean.repeat(n_grid_points, 1),
# task_logvar.repeat(n_grid_points, 1),
# None,
# torch.cat((ptu.FloatTensor(centers),
# ptu.zeros(centers.shape[0], 1)), dim=-1).unsqueeze(0),
# None)[:, 0])
observations[task_loop_i, step + 1, :] = ptu.get_numpy(
next_obs[0, :obs_size])
if self.args.policy != 'dqn':
log_probs[task_loop_i,
step] = ptu.get_numpy(log_prob[0])
if "is_goal_state" in dir(
self.env.unwrapped
) and self.env.unwrapped.is_goal_state():
success_rate[task_loop_i] = 1.
# set: obs <- next_obs
obs = next_obs.clone()
step += 1
returns_per_episode[task_loop_i, episode_idx] = running_reward
if self.args.policy == 'dqn':
return returns_per_episode, success_rate, observations, rewards, reward_preds
# This part is very specific for the Semi-Circle env
# elif self.args.env_name == 'PointRobotSparse-v0':
# return returns_per_episode, success_rate, log_probs, observations, \
# rewards, reward_preds, reward_belief, reward_belief_discretized, centers
else:
return returns_per_episode, success_rate, log_probs, observations, rewards, reward_preds
def log(self, iteration, train_stats):
# --- save model ---
if iteration % self.args.save_interval == 0:
save_path = os.path.join(self.tb_logger.full_output_folder,
'models')
if not os.path.exists(save_path):
os.mkdir(save_path)
torch.save(
self.agent.state_dict(),
os.path.join(save_path, "agent{0}.pt".format(iteration)))
if iteration % self.args.log_interval == 0:
if self.args.policy == 'dqn':
returns, success_rate, observations, rewards, reward_preds = self.evaluate(
)
# This part is super specific for the Semi-Circle env
# elif self.args.env_name == 'PointRobotSparse-v0':
# returns, success_rate, log_probs, observations, \
# rewards, reward_preds, reward_belief, reward_belief_discretized, points = self.evaluate()
else:
returns, success_rate, log_probs, observations, rewards, reward_preds = self.evaluate(
)
if self.args.log_tensorboard:
tasks_to_vis = np.random.choice(self.args.num_eval_tasks, 5)
for i, task in enumerate(tasks_to_vis):
self.env.reset(task)
if PLOT_VIS:
self.tb_logger.writer.add_figure(
'policy_vis/task_{}'.format(i),
utl_eval.plot_rollouts(observations[task, :],
self.env),
self._n_rl_update_steps_total)
self.tb_logger.writer.add_figure(
'reward_prediction_train/task_{}'.format(i),
utl_eval.plot_rew_pred_vs_rew(rewards[task, :],
reward_preds[task, :]),
self._n_rl_update_steps_total)
# self.tb_logger.writer.add_figure('reward_prediction_train/task_{}'.format(i),
# utl_eval.plot_rew_pred_vs_reward_belief_vs_rew(rewards[task, :],
# reward_preds[task, :],
# reward_belief[task, :]),
# self._n_rl_update_steps_total)
# if self.args.env_name == 'PointRobotSparse-v0': # This part is super specific for the Semi-Circle env
# for t in range(0, int(self.args.trajectory_len/4), 3):
# self.tb_logger.writer.add_figure('discrete_belief_reward_pred_task_{}/timestep_{}'.format(i, t),
# utl_eval.plot_discretized_belief_halfcircle(reward_belief_discretized[task, t, :],
# points, self.env,
# observations[task, :t+1]),
# self._n_rl_update_steps_total)
if self.args.max_rollouts_per_task > 1:
for episode_idx in range(self.args.max_rollouts_per_task):
self.tb_logger.writer.add_scalar(
'returns_multi_episode/episode_{}'.format(
episode_idx + 1),
np.mean(returns[:, episode_idx]),
self._n_rl_update_steps_total)
self.tb_logger.writer.add_scalar(
'returns_multi_episode/sum',
np.mean(np.sum(returns, axis=-1)),
self._n_rl_update_steps_total)
self.tb_logger.writer.add_scalar(
'returns_multi_episode/success_rate',
np.mean(success_rate), self._n_rl_update_steps_total)
else:
self.tb_logger.writer.add_scalar(
'returns/returns_mean', np.mean(returns),
self._n_rl_update_steps_total)
self.tb_logger.writer.add_scalar(
'returns/returns_std', np.std(returns),
self._n_rl_update_steps_total)
self.tb_logger.writer.add_scalar(
'returns/success_rate', np.mean(success_rate),
self._n_rl_update_steps_total)
if self.args.policy == 'dqn':
self.tb_logger.writer.add_scalar(
'rl_losses/qf_loss_vs_n_updates',
train_stats['qf_loss'], self._n_rl_update_steps_total)
self.tb_logger.writer.add_scalar(
'weights/q_network',
list(self.agent.qf.parameters())[0].mean(),
self._n_rl_update_steps_total)
if list(self.agent.qf.parameters())[0].grad is not None:
param_list = list(self.agent.qf.parameters())
self.tb_logger.writer.add_scalar(
'gradients/q_network',
sum([
param_list[i].grad.mean()
for i in range(len(param_list))
]), self._n_rl_update_steps_total)
self.tb_logger.writer.add_scalar(
'weights/q_target',
list(self.agent.target_qf.parameters())[0].mean(),
self._n_rl_update_steps_total)
if list(self.agent.target_qf.parameters()
)[0].grad is not None:
param_list = list(self.agent.target_qf.parameters())
self.tb_logger.writer.add_scalar(
'gradients/q_target',
sum([
param_list[i].grad.mean()
for i in range(len(param_list))
]), self._n_rl_update_steps_total)
else:
self.tb_logger.writer.add_scalar(
'policy/log_prob', np.mean(log_probs),
self._n_rl_update_steps_total)
self.tb_logger.writer.add_scalar(
'rl_losses/qf1_loss', train_stats['qf1_loss'],
self._n_rl_update_steps_total)
self.tb_logger.writer.add_scalar(
'rl_losses/qf2_loss', train_stats['qf2_loss'],
self._n_rl_update_steps_total)
self.tb_logger.writer.add_scalar(
'rl_losses/policy_loss', train_stats['policy_loss'],
self._n_rl_update_steps_total)
self.tb_logger.writer.add_scalar(
'rl_losses/alpha_entropy_loss',
train_stats['alpha_entropy_loss'],
self._n_rl_update_steps_total)
# weights and gradients
self.tb_logger.writer.add_scalar(
'weights/q1_network',
list(self.agent.qf1.parameters())[0].mean(),
self._n_rl_update_steps_total)
if list(self.agent.qf1.parameters())[0].grad is not None:
param_list = list(self.agent.qf1.parameters())
self.tb_logger.writer.add_scalar(
'gradients/q1_network',
sum([
param_list[i].grad.mean()
for i in range(len(param_list))
]), self._n_rl_update_steps_total)
self.tb_logger.writer.add_scalar(
'weights/q1_target',
list(self.agent.qf1_target.parameters())[0].mean(),
self._n_rl_update_steps_total)
if list(self.agent.qf1_target.parameters()
)[0].grad is not None:
param_list = list(self.agent.qf1_target.parameters())
self.tb_logger.writer.add_scalar(
'gradients/q1_target',
sum([
param_list[i].grad.mean()
for i in range(len(param_list))
]), self._n_rl_update_steps_total)
self.tb_logger.writer.add_scalar(
'weights/q2_network',
list(self.agent.qf2.parameters())[0].mean(),
self._n_rl_update_steps_total)
if list(self.agent.qf2.parameters())[0].grad is not None:
param_list = list(self.agent.qf2.parameters())
self.tb_logger.writer.add_scalar(
'gradients/q2_network',
sum([
param_list[i].grad.mean()
for i in range(len(param_list))
]), self._n_rl_update_steps_total)
self.tb_logger.writer.add_scalar(
'weights/q2_target',
list(self.agent.qf2_target.parameters())[0].mean(),
self._n_rl_update_steps_total)
if list(self.agent.qf2_target.parameters()
)[0].grad is not None:
param_list = list(self.agent.qf2_target.parameters())
self.tb_logger.writer.add_scalar(
'gradients/q2_target',
sum([
param_list[i].grad.mean()
for i in range(len(param_list))
]), self._n_rl_update_steps_total)
self.tb_logger.writer.add_scalar(
'weights/policy',
list(self.agent.policy.parameters())[0].mean(),
self._n_rl_update_steps_total)
if list(self.agent.policy.parameters()
)[0].grad is not None:
param_list = list(self.agent.policy.parameters())
self.tb_logger.writer.add_scalar(
'gradients/policy',
sum([
param_list[i].grad.mean()
for i in range(len(param_list))
]), self._n_rl_update_steps_total)
for k, v in [
('num_rl_updates', self._n_rl_update_steps_total),
('time_elapsed', time.time() - self._start_time),
('iteration', iteration),
]:
self.tb_logger.writer.add_scalar(k, v,
self._n_rl_update_steps_total)
self.tb_logger.finish_iteration(iteration)
print(
"Iteration -- {}, Success rate -- {:.3f}, Avg. return -- {:.3f}, Elapsed time {:5d}[s]"
.format(iteration, np.mean(success_rate),
np.mean(np.sum(returns, axis=-1)),
int(time.time() - self._start_time)))
def sample_rl_batch(self, tasks, batch_size):
''' sample batch of unordered rl training data from a list/array of tasks '''
# this batch consists of transitions sampled randomly from replay buffer
batches = [
ptu.np_to_pytorch_batch(self.storage.random_batch(
task, batch_size)) for task in tasks
]
unpacked = [utl.unpack_batch(batch) for batch in batches]
# group elements together
unpacked = [[x[i] for x in unpacked] for i in range(len(unpacked[0]))]
unpacked = [torch.cat(x, dim=0) for x in unpacked]
return unpacked
def _start_training(self):
self._n_rl_update_steps_total = 0
self._start_time = time.time()
def training_mode(self, mode):
self.agent.train(mode)
def update_encoding(self, obs, action, reward, done, hidden_state):
# reset hidden state of the recurrent net when the task is done
hidden_state = self.vae.encoder.reset_hidden(hidden_state, done)
with torch.no_grad(): # size should be (batch, dim)
task_sample, task_mean, task_logvar, hidden_state = self.vae.encoder(
actions=action.float(),
states=obs,
rewards=reward,
hidden_state=hidden_state,
return_prior=False)
return task_sample, task_mean, task_logvar, hidden_state
@staticmethod
def get_augmented_obs(obs, mean, logvar):
mean = mean.reshape((-1, mean.shape[-1]))
logvar = logvar.reshape((-1, logvar.shape[-1]))
return torch.cat((obs, mean, logvar), dim=-1)
def load_model(self, agent_path, device='cpu'):
self.agent.load_state_dict(torch.load(agent_path, map_location=device))
self.load_vae()
self.training_mode(False)
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
class Learner:
"""
Learner (no meta-learning), can be used to train avg/oracle/belief-oracle policies.
"""
def __init__(self, args):
self.args = args
utl.seed(self.args.seed, self.args.deterministic_execution)
# calculate number of updates and keep count of frames/iterations
self.num_updates = int(
args.num_frames) // args.policy_num_steps // args.num_processes
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,
device=device,
episodes_per_task=self.args.max_rollouts_per_task,
normalise_rew=args.norm_rew_for_policy,
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
# get policy input dimensions
self.args.state_dim = self.envs.observation_space.shape[0]
self.args.task_dim = self.envs.task_dim
self.args.belief_dim = self.envs.belief_dim
self.args.num_states = self.envs.num_states
# get policy output (action) dimensions
self.args.action_space = self.envs.action_space
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]
# initialise policy
self.policy_storage = self.initialise_policy_storage()
self.policy = self.initialise_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):
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(
args=self.args,
#
pass_state_to_policy=self.args.pass_state_to_policy,
pass_latent_to_policy=
False, # use metalearner.py if you want to use the VAE
pass_belief_to_policy=self.args.pass_belief_to_policy,
pass_task_to_policy=self.args.pass_task_to_policy,
dim_state=self.args.state_dim,
dim_latent=0,
dim_belief=self.args.belief_dim,
dim_task=self.args.task_dim,
#
hidden_layers=self.args.policy_layers,
activation_function=self.args.policy_activation_function,
policy_initialisation=self.args.policy_initialisation,
#
action_space=self.envs.action_space,
init_std=self.args.policy_init_std,
norm_actions_of_policy=self.args.norm_actions_of_policy,
action_low=action_low,
action_high=action_high,
).to(device)
# initialise policy trainer
if self.args.policy == 'a2c':
policy = A2C(
self.args,
policy_net,
self.args.policy_value_loss_coef,
self.args.policy_entropy_coef,
policy_optimiser=self.args.policy_optimiser,
policy_anneal_lr=self.args.policy_anneal_lr,
train_steps=self.num_updates,
lr=self.args.lr_policy,
eps=self.args.policy_eps,
)
elif self.args.policy == 'ppo':
policy = PPO(
self.args,
policy_net,
self.args.policy_value_loss_coef,
self.args.policy_entropy_coef,
policy_optimiser=self.args.policy_optimiser,
policy_anneal_lr=self.args.policy_anneal_lr,
train_steps=self.num_updates,
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
return policy
def train(self):
""" Main training loop """
start_time = time.time()
# reset environments
state, belief, task = utl.reset_env(self.envs, self.args)
# insert initial observation / embeddings to rollout storage
self.policy_storage.prev_state[0].copy_(state)
# log once before training
with torch.no_grad():
self.log(None, None, start_time)
for self.iter_idx in range(self.num_updates):
# 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,
state=state,
belief=belief,
task=task,
deterministic=False)
# observe reward and next obs
[state, belief,
task], (rew_raw, rew_normalised), done, infos = utl.env_step(
self.envs, action, self.args)
# 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)
# reset environments that are done
done_indices = np.argwhere(done.flatten()).flatten()
if len(done_indices) > 0:
state, belief, task = utl.reset_env(self.envs,
self.args,
indices=done_indices,
state=state)
# add experience to policy buffer
self.policy_storage.insert(
state=state,
belief=belief,
task=task,
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=torch.from_numpy(np.array(done,
dtype=float)).unsqueeze(1),
)
self.frames += self.args.num_processes
# --- UPDATE ---
train_stats = self.update(state=state, belief=belief, task=task)
# log
run_stats = [action, action_log_prob, value]
if train_stats is not None:
with torch.no_grad():
self.log(run_stats, train_stats, start_time)
# clean up after update
self.policy_storage.after_update()
def get_value(self, state, belief, task):
return self.policy.actor_critic.get_value(state=state,
belief=belief,
task=task,
latent=None).detach()
def update(self, state, belief, task):
"""
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(state=state, belief=belief, task=task)
# 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(
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.
"""
# --- visualise behaviour of policy ---
if self.iter_idx % self.args.vis_interval == 0:
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,
ret_rms=ret_rms,
)
# --- evaluate policy ----
if self.iter_idx % self.args.eval_interval == 0:
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,
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)
idx_labels = ['']
if self.args.save_intermediate_models:
idx_labels.append(int(self.iter_idx))
for idx_label in idx_labels:
torch.save(self.policy.actor_critic,
os.path.join(save_path, f"policy{idx_label}.pt"))
# save normalisation params of envs
if self.args.norm_rew_for_policy:
rew_rms = self.envs.venv.ret_rms
utl.save_obj(rew_rms, save_path, f"env_rew_rms{idx_label}")
# TODO: grab from policy and save?
# if self.args.norm_obs_for_policy:
# obs_rms = self.envs.venv.obs_rms
# utl.save_obj(obs_rms, save_path, f"env_obs_rms{idx_label}")
# --- log some other things ---
if (self.iter_idx % self.args.log_interval == 0) and (train_stats
is not None):
train_stats, _ = train_stats
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.cpu().numpy().mean()
for i in range(len(param_list))
])
param_grad_mean = np.mean([
param_list[i].grad.cpu().numpy().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.cpu().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.cpu().numpy().mean(),
self.iter_idx)
def __init__(self, args):
self.args = args
utl.seed(self.args.seed, self.args.deterministic_execution)
# calculate number of updates and keep count of frames/iterations
self.num_updates = int(
args.num_frames) // args.policy_num_steps // args.num_processes
self.frames = 0
self.iter_idx = -1
# 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,
device=device,
episodes_per_task=self.args.max_rollouts_per_task,
normalise_rew=args.norm_rew_for_policy,
ret_rms=None,
tasks=None)
if self.args.single_task_mode:
# get the current tasks (which will be num_process many different tasks)
self.train_tasks = self.envs.get_task()
# set the tasks to the first task (i.e. just a random task)
self.train_tasks[1:] = self.train_tasks[0]
# make it a list
self.train_tasks = [t for t in self.train_tasks]
# re-initialise environments with those tasks
self.envs = make_vec_envs(
env_name=args.env_name,
seed=args.seed,
num_processes=args.num_processes,
gamma=args.policy_gamma,
device=device,
episodes_per_task=self.args.max_rollouts_per_task,
normalise_rew=args.norm_rew_for_policy,
ret_rms=None,
tasks=self.train_tasks,
)
# save the training tasks so we can evaluate on the same envs later
utl.save_obj(self.train_tasks, self.logger.full_output_folder,
"train_tasks")
else:
self.train_tasks = 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
# get policy input dimensions
self.args.state_dim = self.envs.observation_space.shape[0]
self.args.task_dim = self.envs.task_dim
self.args.belief_dim = self.envs.belief_dim
self.args.num_states = self.envs.num_states
# get policy output (action) dimensions
self.args.action_space = self.envs.action_space
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]
# initialise policy
self.policy_storage = self.initialise_policy_storage()
self.policy = self.initialise_policy()
def __init__(self, args):
"""
Seeds everything.
Initialises: logger, environments, policy (+storage +optimiser).
"""
self.args = args
# make sure everything has the same seed
utl.seed(self.args.seed)
# initialise environment
self.env = make_env(
self.args.env_name,
self.args.max_rollouts_per_task,
seed=self.args.seed,
n_tasks=1,
modify_init_state_dist=self.args.modify_init_state_dist
if 'modify_init_state_dist' in self.args else False,
on_circle_init_state=self.args.on_circle_init_state
if 'on_circle_init_state' in self.args else True)
# saving buffer with task in name folder
if hasattr(self.args, 'save_buffer') and self.args.save_buffer:
env_dir = os.path.join(self.args.main_save_dir,
'{}'.format(self.args.env_name))
goal = self.env.unwrapped._goal
self.output_dir = os.path.join(
env_dir, self.args.save_dir,
'seed_{}_'.format(self.args.seed) +
off_utl.create_goal_path_ext_from_goal(goal))
if self.args.save_models or self.args.save_buffer:
os.makedirs(self.output_dir, exist_ok=True)
config_utl.save_config_file(args, self.output_dir)
# initialize tensorboard logger
if self.args.log_tensorboard:
self.tb_logger = TBLogger(self.args)
# if not self.args.log_tensorboard:
# self.save_config_json_file()
# unwrapped env to get some info about the environment
unwrapped_env = self.env.unwrapped
# calculate what the maximum length of the trajectories is
args.max_trajectory_len = unwrapped_env._max_episode_steps
args.max_trajectory_len *= self.args.max_rollouts_per_task
self.args.max_trajectory_len = args.max_trajectory_len
# get action / observation dimensions
if isinstance(self.env.action_space, gym.spaces.discrete.Discrete):
self.args.action_dim = 1
else:
self.args.action_dim = self.env.action_space.shape[0]
self.args.obs_dim = self.env.observation_space.shape[0]
self.args.num_states = unwrapped_env.num_states if hasattr(
unwrapped_env, 'num_states') else None
self.args.act_space = self.env.action_space
# simulate env step to get reward types
_, _, _, info = unwrapped_env.step(unwrapped_env.action_space.sample())
reward_types = [
reward_type for reward_type in list(info.keys())
if reward_type.startswith('reward')
]
# support dense rewards training (if exists)
self.args.dense_train_sparse_test = self.args.dense_train_sparse_test \
if 'dense_train_sparse_test' in self.args else False
# initialize policy
self.initialize_policy()
# initialize buffer for RL updates
self.policy_storage = MultiTaskPolicyStorage(
max_replay_buffer_size=int(self.args.policy_buffer_size),
obs_dim=self.args.obs_dim,
action_space=self.env.action_space,
tasks=[0],
trajectory_len=args.max_trajectory_len,
num_reward_arrays=len(reward_types)
if reward_types and self.args.dense_train_sparse_test else 1,
reward_types=reward_types,
)
self.args.belief_reward = False # initialize arg to not use belief rewards
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)
class Learner:
"""
Learner class.
"""
def __init__(self, args):
"""
Seeds everything.
Initialises: logger, environments, policy (+storage +optimiser).
"""
self.args = args
# make sure everything has the same seed
utl.seed(self.args.seed)
# initialise environment
self.env = make_env(self.args.env_name,
self.args.max_rollouts_per_task,
seed=self.args.seed,
n_tasks=1,
modify_init_state_dist=self.args.modify_init_state_dist
if 'modify_init_state_dist' in self.args else False,
on_circle_init_state=self.args.on_circle_init_state
if 'on_circle_init_state' in self.args else True)
# saving buffer with task in name folder
if hasattr(self.args, 'save_buffer') and self.args.save_buffer:
env_dir = os.path.join(self.args.main_save_dir,
'{}'.format(self.args.env_name))
goal = self.env.unwrapped._goal
self.output_dir = os.path.join(env_dir, self.args.save_dir, 'seed_{}_'.format(self.args.seed) +
off_utl.create_goal_path_ext_from_goal(goal))
if self.args.save_models or self.args.save_buffer:
os.makedirs(self.output_dir, exist_ok=True)
config_utl.save_config_file(args, self.output_dir)
# initialize tensorboard logger
if self.args.log_tensorboard:
self.tb_logger = TBLogger(self.args)
# if not self.args.log_tensorboard:
# self.save_config_json_file()
# unwrapped env to get some info about the environment
unwrapped_env = self.env.unwrapped
# calculate what the maximum length of the trajectories is
args.max_trajectory_len = unwrapped_env._max_episode_steps
args.max_trajectory_len *= self.args.max_rollouts_per_task
self.args.max_trajectory_len = args.max_trajectory_len
# get action / observation dimensions
if isinstance(self.env.action_space, gym.spaces.discrete.Discrete):
self.args.action_dim = 1
else:
self.args.action_dim = self.env.action_space.shape[0]
self.args.obs_dim = self.env.observation_space.shape[0]
self.args.num_states = unwrapped_env.num_states if hasattr(unwrapped_env, 'num_states') else None
self.args.act_space = self.env.action_space
# simulate env step to get reward types
_, _, _, info = unwrapped_env.step(unwrapped_env.action_space.sample())
reward_types = [reward_type for reward_type in list(info.keys()) if reward_type.startswith('reward')]
# support dense rewards training (if exists)
self.args.dense_train_sparse_test = self.args.dense_train_sparse_test \
if 'dense_train_sparse_test' in self.args else False
# initialize policy
self.initialize_policy()
# initialize buffer for RL updates
self.policy_storage = MultiTaskPolicyStorage(
max_replay_buffer_size=int(self.args.policy_buffer_size),
obs_dim=self.args.obs_dim,
action_space=self.env.action_space,
tasks=[0],
trajectory_len=args.max_trajectory_len,
num_reward_arrays=len(reward_types) if reward_types and self.args.dense_train_sparse_test else 1,
reward_types=reward_types,
)
self.args.belief_reward = False # initialize arg to not use belief rewards
def initialize_policy(self):
if self.args.policy == 'dqn':
assert self.args.act_space.__class__.__name__ == "Discrete", (
"Can't train DQN with continuous action space!")
q_network = FlattenMlp(input_size=self.args.obs_dim,
output_size=self.args.act_space.n,
hidden_sizes=self.args.dqn_layers)
self.agent = DQN(
q_network,
# optimiser_vae=self.optimizer_vae,
lr=self.args.policy_lr,
gamma=self.args.gamma,
eps_init=self.args.dqn_epsilon_init,
eps_final=self.args.dqn_epsilon_final,
exploration_iters=self.args.dqn_exploration_iters,
tau=self.args.soft_target_tau,
).to(ptu.device)
# elif self.args.policy == 'ddqn':
# assert self.args.act_space.__class__.__name__ == "Discrete", (
# "Can't train DDQN with continuous action space!")
# q_network = FlattenMlp(input_size=self.args.obs_dim,
# output_size=self.args.act_space.n,
# hidden_sizes=self.args.dqn_layers)
# self.agent = DoubleDQN(
# q_network,
# # optimiser_vae=self.optimizer_vae,
# lr=self.args.policy_lr,
# eps_optim=self.args.dqn_eps,
# alpha_optim=self.args.dqn_alpha,
# gamma=self.args.gamma,
# eps_init=self.args.dqn_epsilon_init,
# eps_final=self.args.dqn_epsilon_final,
# exploration_iters=self.args.dqn_exploration_iters,
# tau=self.args.soft_target_tau,
# ).to(ptu.device)
elif self.args.policy == 'sac':
assert self.args.act_space.__class__.__name__ == "Box", (
"Can't train SAC with discrete action space!")
q1_network = FlattenMlp(input_size=self.args.obs_dim + self.args.action_dim,
output_size=1,
hidden_sizes=self.args.dqn_layers)
q2_network = FlattenMlp(input_size=self.args.obs_dim + self.args.action_dim,
output_size=1,
hidden_sizes=self.args.dqn_layers)
policy = TanhGaussianPolicy(obs_dim=self.args.obs_dim,
action_dim=self.args.action_dim,
hidden_sizes=self.args.policy_layers)
self.agent = SAC(
policy,
q1_network,
q2_network,
actor_lr=self.args.actor_lr,
critic_lr=self.args.critic_lr,
gamma=self.args.gamma,
tau=self.args.soft_target_tau,
entropy_alpha=self.args.entropy_alpha,
automatic_entropy_tuning=self.args.automatic_entropy_tuning,
alpha_lr=self.args.alpha_lr
).to(ptu.device)
else:
raise NotImplementedError
def train(self):
"""
meta-training loop
"""
self._start_training()
self.task_idx = 0
for iter_ in range(self.args.num_iters):
self.training_mode(True)
if iter_ == 0:
print('Collecting initial pool of data..')
self.env.reset_task(idx=self.task_idx)
self.collect_rollouts(num_rollouts=self.args.num_init_rollouts_pool, random_actions=True)
print('Done!')
# collect data from subset of train tasks
self.env.reset_task(idx=self.task_idx)
self.collect_rollouts(num_rollouts=self.args.num_rollouts_per_iter)
# update
train_stats = self.update([self.task_idx])
self.training_mode(False)
if self.args.policy == 'dqn':
self.agent.set_exploration_parameter(iter_ + 1)
# evaluate and log
if (iter_ + 1) % self.args.log_interval == 0:
self.log(iter_ + 1, train_stats)
def update(self, tasks):
'''
RL updates
:param tasks: list/array of task indices. perform update based on the tasks
:return:
'''
# --- RL TRAINING ---
rl_losses_agg = {}
for update in range(self.args.rl_updates_per_iter):
# sample random RL batch
obs, actions, rewards, next_obs, terms = self.sample_rl_batch(tasks, self.args.batch_size)
# flatten out task dimension
t, b, _ = obs.size()
obs = obs.view(t * b, -1)
actions = actions.view(t * b, -1)
rewards = rewards.view(t * b, -1)
next_obs = next_obs.view(t * b, -1)
terms = terms.view(t * b, -1)
# RL update
rl_losses = self.agent.update(obs, actions, rewards, next_obs, terms)
for k, v in rl_losses.items():
if update == 0: # first iterate - create list
rl_losses_agg[k] = [v]
else: # append values
rl_losses_agg[k].append(v)
# take mean
for k in rl_losses_agg:
rl_losses_agg[k] = np.mean(rl_losses_agg[k])
self._n_rl_update_steps_total += self.args.rl_updates_per_iter
return rl_losses_agg
def evaluate(self, tasks):
num_episodes = self.args.max_rollouts_per_task
num_steps_per_episode = self.env.unwrapped._max_episode_steps
returns_per_episode = np.zeros((len(tasks), num_episodes))
success_rate = np.zeros(len(tasks))
if self.args.policy == 'dqn':
values = np.zeros((len(tasks), self.args.max_trajectory_len))
else:
obs_size = self.env.unwrapped.observation_space.shape[0]
observations = np.zeros((len(tasks), self.args.max_trajectory_len + 1, obs_size))
log_probs = np.zeros((len(tasks), self.args.max_trajectory_len))
for task_idx, task in enumerate(tasks):
obs = ptu.from_numpy(self.env.reset(task))
obs = obs.reshape(-1, obs.shape[-1])
step = 0
if self.args.policy == 'sac':
observations[task_idx, step, :] = ptu.get_numpy(obs[0, :obs_size])
for episode_idx in range(num_episodes):
running_reward = 0.
for step_idx in range(num_steps_per_episode):
# add distribution parameters to observation - policy is conditioned on posterior
if self.args.policy == 'dqn':
action, value = self.agent.act(obs=obs, deterministic=True)
else:
action, _, _, log_prob = self.agent.act(obs=obs,
deterministic=self.args.eval_deterministic,
return_log_prob=True)
# observe reward and next obs
next_obs, reward, done, info = utl.env_step(self.env, action.squeeze(dim=0))
running_reward += reward.item()
if self.args.policy == 'dqn':
values[task_idx, step] = value.item()
else:
observations[task_idx, step + 1, :] = ptu.get_numpy(next_obs[0, :obs_size])
log_probs[task_idx, step] = ptu.get_numpy(log_prob[0])
if "is_goal_state" in dir(self.env.unwrapped) and self.env.unwrapped.is_goal_state():
success_rate[task_idx] = 1.
# set: obs <- next_obs
obs = next_obs.clone()
step += 1
returns_per_episode[task_idx, episode_idx] = running_reward
if self.args.policy == 'dqn':
return returns_per_episode, success_rate, values
else:
return returns_per_episode, success_rate, log_probs, observations
def log(self, iteration, train_stats):
# --- save models ---
if iteration % self.args.save_interval == 0:
if self.args.save_models:
if self.args.log_tensorboard:
save_path = os.path.join(self.tb_logger.full_output_folder, 'models')
else:
save_path = os.path.join(self.output_dir, 'models')
if not os.path.exists(save_path):
os.mkdir(save_path)
torch.save(self.agent.state_dict(), os.path.join(save_path, "agent{0}.pt".format(iteration)))
if hasattr(self.args, 'save_buffer') and self.args.save_buffer:
self.save_buffer()
# evaluate to get more stats
if self.args.policy == 'dqn':
# get stats on train tasks
returns_train, success_rate_train, values = self.evaluate([0])
else:
# get stats on train tasks
returns_train, success_rate_train, log_probs, observations = self.evaluate([0])
if self.args.log_tensorboard:
if self.args.policy != 'dqn':
# self.env.reset(0)
# self.tb_logger.writer.add_figure('policy_vis_train/task_0',
# utl_eval.plot_rollouts(observations[0, :], self.env),
# self._n_env_steps_total)
# obs, _, _, _, _ = self.sample_rl_batch(tasks=[0],
# batch_size=self.policy_storage.task_buffers[0].size())
# self.tb_logger.writer.add_figure('state_space_coverage/task_0',
# utl_eval.plot_visited_states(ptu.get_numpy(obs[0][:, :2]), self.env),
# self._n_env_steps_total)
pass
# some metrics
self.tb_logger.writer.add_scalar('metrics/successes_in_buffer',
self._successes_in_buffer / self._n_env_steps_total,
self._n_env_steps_total)
if self.args.max_rollouts_per_task > 1:
for episode_idx in range(self.args.max_rollouts_per_task):
self.tb_logger.writer.add_scalar('returns_multi_episode/episode_{}'.
format(episode_idx + 1),
np.mean(returns_train[:, episode_idx]),
self._n_env_steps_total)
self.tb_logger.writer.add_scalar('returns_multi_episode/sum',
np.mean(np.sum(returns_train, axis=-1)),
self._n_env_steps_total)
self.tb_logger.writer.add_scalar('returns_multi_episode/success_rate',
np.mean(success_rate_train),
self._n_env_steps_total)
else:
self.tb_logger.writer.add_scalar('returns/returns_mean_train', np.mean(returns_train),
self._n_env_steps_total)
self.tb_logger.writer.add_scalar('returns/returns_std_train', np.std(returns_train),
self._n_env_steps_total)
self.tb_logger.writer.add_scalar('returns/success_rate_train', np.mean(success_rate_train),
self._n_env_steps_total)
# policy
if self.args.policy == 'dqn':
self.tb_logger.writer.add_scalar('policy/value_init', np.mean(values[:, 0]), self._n_env_steps_total)
self.tb_logger.writer.add_scalar('policy/value_halfway', np.mean(values[:, int(values.shape[-1]/2)]), self._n_env_steps_total)
self.tb_logger.writer.add_scalar('policy/value_final', np.mean(values[:, -1]), self._n_env_steps_total)
self.tb_logger.writer.add_scalar('policy/exploration_epsilon', self.agent.eps, self._n_env_steps_total)
# RL losses
self.tb_logger.writer.add_scalar('rl_losses/qf_loss_vs_n_updates', train_stats['qf_loss'],
self._n_rl_update_steps_total)
self.tb_logger.writer.add_scalar('rl_losses/qf_loss_vs_n_env_steps', train_stats['qf_loss'],
self._n_env_steps_total)
else:
self.tb_logger.writer.add_scalar('policy/log_prob', np.mean(log_probs), self._n_env_steps_total)
self.tb_logger.writer.add_scalar('rl_losses/qf1_loss', train_stats['qf1_loss'],
self._n_env_steps_total)
self.tb_logger.writer.add_scalar('rl_losses/qf2_loss', train_stats['qf2_loss'],
self._n_env_steps_total)
self.tb_logger.writer.add_scalar('rl_losses/policy_loss', train_stats['policy_loss'],
self._n_env_steps_total)
self.tb_logger.writer.add_scalar('rl_losses/alpha_entropy_loss', train_stats['alpha_entropy_loss'],
self._n_env_steps_total)
# weights and gradients
if self.args.policy == 'dqn':
self.tb_logger.writer.add_scalar('weights/q_network',
list(self.agent.qf.parameters())[0].mean(),
self._n_env_steps_total)
if list(self.agent.qf.parameters())[0].grad is not None:
param_list = list(self.agent.qf.parameters())
self.tb_logger.writer.add_scalar('gradients/q_network',
sum([param_list[i].grad.mean() for i in range(len(param_list))]),
self._n_env_steps_total)
self.tb_logger.writer.add_scalar('weights/q_target',
list(self.agent.target_qf.parameters())[0].mean(),
self._n_env_steps_total)
if list(self.agent.target_qf.parameters())[0].grad is not None:
param_list = list(self.agent.target_qf.parameters())
self.tb_logger.writer.add_scalar('gradients/q_target',
sum([param_list[i].grad.mean() for i in range(len(param_list))]),
self._n_env_steps_total)
else:
self.tb_logger.writer.add_scalar('weights/q1_network',
list(self.agent.qf1.parameters())[0].mean(),
self._n_env_steps_total)
if list(self.agent.qf1.parameters())[0].grad is not None:
param_list = list(self.agent.qf1.parameters())
self.tb_logger.writer.add_scalar('gradients/q1_network',
sum([param_list[i].grad.mean() for i in range(len(param_list))]),
self._n_env_steps_total)
self.tb_logger.writer.add_scalar('weights/q1_target',
list(self.agent.qf1_target.parameters())[0].mean(),
self._n_env_steps_total)
if list(self.agent.qf1_target.parameters())[0].grad is not None:
param_list = list(self.agent.qf1_target.parameters())
self.tb_logger.writer.add_scalar('gradients/q1_target',
sum([param_list[i].grad.mean() for i in range(len(param_list))]),
self._n_env_steps_total)
self.tb_logger.writer.add_scalar('weights/q2_network',
list(self.agent.qf2.parameters())[0].mean(),
self._n_env_steps_total)
if list(self.agent.qf2.parameters())[0].grad is not None:
param_list = list(self.agent.qf2.parameters())
self.tb_logger.writer.add_scalar('gradients/q2_network',
sum([param_list[i].grad.mean() for i in range(len(param_list))]),
self._n_env_steps_total)
self.tb_logger.writer.add_scalar('weights/q2_target',
list(self.agent.qf2_target.parameters())[0].mean(),
self._n_env_steps_total)
if list(self.agent.qf2_target.parameters())[0].grad is not None:
param_list = list(self.agent.qf2_target.parameters())
self.tb_logger.writer.add_scalar('gradients/q2_target',
sum([param_list[i].grad.mean() for i in range(len(param_list))]),
self._n_env_steps_total)
self.tb_logger.writer.add_scalar('weights/policy',
list(self.agent.policy.parameters())[0].mean(),
self._n_env_steps_total)
if list(self.agent.policy.parameters())[0].grad is not None:
param_list = list(self.agent.policy.parameters())
self.tb_logger.writer.add_scalar('gradients/policy',
sum([param_list[i].grad.mean() for i in range(len(param_list))]),
self._n_env_steps_total)
self.tb_logger.finish_iteration(iteration)
print("Iteration -- {}, Success rate -- {:.3f}, Avg. return -- {:.3f}, Elapsed time {:5d}[s]"
.format(iteration, np.mean(success_rate_train), np.mean(np.sum(returns_train, axis=-1)),
int(time.time() - self._start_time)))
# output to user
# print("Iteration -- {:3d}, Num. RL updates -- {:6d}, Elapsed time {:5d}[s]".
# format(iteration,
# self._n_rl_update_steps_total,
# int(time.time() - self._start_time)))
def training_mode(self, mode):
self.agent.train(mode)
def collect_rollouts(self, num_rollouts, random_actions=False):
'''
:param num_rollouts:
:param random_actions: whether to use policy to sample actions, or randomly sample action space
:return:
'''
for rollout in range(num_rollouts):
obs = ptu.from_numpy(self.env.reset(self.task_idx))
obs = obs.reshape(-1, obs.shape[-1])
done_rollout = False
while not done_rollout:
if random_actions:
if self.args.policy == 'dqn':
action = ptu.FloatTensor([[[self.env.action_space.sample()]]]).long() # Sample random action
else:
action = ptu.FloatTensor([self.env.action_space.sample()]) # Sample random action
else:
if self.args.policy == 'dqn':
action, _ = self.agent.act(obs=obs) # DQN
else:
action, _, _, _ = self.agent.act(obs=obs) # SAC
# observe reward and next obs
next_obs, reward, done, info = utl.env_step(self.env, action.squeeze(dim=0))
done_rollout = False if ptu.get_numpy(done[0][0]) == 0. else True
# add data to policy buffer - (s+, a, r, s'+, term)
term = self.env.unwrapped.is_goal_state() if "is_goal_state" in dir(self.env.unwrapped) else False
if self.args.dense_train_sparse_test:
rew_to_buffer = {rew_type: rew for rew_type, rew in info.items()
if rew_type.startswith('reward')}
else:
rew_to_buffer = ptu.get_numpy(reward.squeeze(dim=0))
self.policy_storage.add_sample(task=self.task_idx,
observation=ptu.get_numpy(obs.squeeze(dim=0)),
action=ptu.get_numpy(action.squeeze(dim=0)),
reward=rew_to_buffer,
terminal=np.array([term], dtype=float),
next_observation=ptu.get_numpy(next_obs.squeeze(dim=0)))
# set: obs <- next_obs
obs = next_obs.clone()
# update statistics
self._n_env_steps_total += 1
if "is_goal_state" in dir(self.env.unwrapped) and self.env.unwrapped.is_goal_state(): # count successes
self._successes_in_buffer += 1
self._n_rollouts_total += 1
def sample_rl_batch(self, tasks, batch_size):
''' sample batch of unordered rl training data from a list/array of tasks '''
# this batch consists of transitions sampled randomly from replay buffer
batches = [ptu.np_to_pytorch_batch(
self.policy_storage.random_batch(task, batch_size)) for task in tasks]
unpacked = [utl.unpack_batch(batch) for batch in batches]
# group elements together
unpacked = [[x[i] for x in unpacked] for i in range(len(unpacked[0]))]
unpacked = [torch.cat(x, dim=0) for x in unpacked]
return unpacked
def _start_training(self):
self._n_env_steps_total = 0
self._n_rl_update_steps_total = 0
self._n_vae_update_steps_total = 0
self._n_rollouts_total = 0
self._successes_in_buffer = 0
self._start_time = time.time()
def load_model(self, device='cpu', **kwargs):
if "agent_path" in kwargs:
self.agent.load_state_dict(torch.load(kwargs["agent_path"], map_location=device))
self.training_mode(False)
def save_buffer(self):
size = self.policy_storage.task_buffers[0].size()
np.save(os.path.join(self.output_dir, 'obs'), self.policy_storage.task_buffers[0]._observations[:size])
np.save(os.path.join(self.output_dir, 'actions'), self.policy_storage.task_buffers[0]._actions[:size])
if self.args.dense_train_sparse_test:
for reward_type, reward_arr in self.policy_storage.task_buffers[0]._rewards.items():
np.save(os.path.join(self.output_dir, reward_type), reward_arr[:size])
else:
np.save(os.path.join(self.output_dir, 'rewards'), self.policy_storage.task_buffers[0]._rewards[:size])
np.save(os.path.join(self.output_dir, 'next_obs'), self.policy_storage.task_buffers[0]._next_obs[:size])
np.save(os.path.join(self.output_dir, 'terminals'), self.policy_storage.task_buffers[0]._terminals[:size])
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)
# calculate number of updates and keep count of frames/iterations
self.num_updates = int(
args.num_frames) // args.policy_num_steps // args.num_processes
self.frames = 0
self.iter_idx = -1
# 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,
device=device,
episodes_per_task=self.args.max_rollouts_per_task,
normalise_rew=args.norm_rew_for_policy,
ret_rms=None,
tasks=None)
if self.args.single_task_mode:
# get the current tasks (which will be num_process many different tasks)
self.train_tasks = self.envs.get_task()
# set the tasks to the first task (i.e. just a random task)
self.train_tasks[1:] = self.train_tasks[0]
# make it a list
self.train_tasks = [t for t in self.train_tasks]
# re-initialise environments with those tasks
self.envs = make_vec_envs(
env_name=args.env_name,
seed=args.seed,
num_processes=args.num_processes,
gamma=args.policy_gamma,
device=device,
episodes_per_task=self.args.max_rollouts_per_task,
normalise_rew=args.norm_rew_for_policy,
ret_rms=None,
tasks=self.train_tasks)
# save the training tasks so we can evaluate on the same envs later
utl.save_obj(self.train_tasks, self.logger.full_output_folder,
"train_tasks")
else:
self.train_tasks = None
# calculate what the maximum length of the trajectories is
self.args.max_trajectory_len = self.envs._max_episode_steps
self.args.max_trajectory_len *= self.args.max_rollouts_per_task
# get policy input dimensions
self.args.state_dim = self.envs.observation_space.shape[0]
self.args.task_dim = self.envs.task_dim
self.args.belief_dim = self.envs.belief_dim
self.args.num_states = self.envs.num_states
# get policy output (action) dimensions
self.args.action_space = self.envs.action_space
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]
# initialise VAE and policy
self.vae = VaribadVAE(self.args, self.logger, lambda: self.iter_idx)
self.policy_storage = self.initialise_policy_storage()
self.policy = self.initialise_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=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(self):
# initialise policy network
policy_net = Policy(
args=self.args,
#
pass_state_to_policy=self.args.pass_state_to_policy,
pass_latent_to_policy=self.args.pass_latent_to_policy,
pass_belief_to_policy=self.args.pass_belief_to_policy,
pass_task_to_policy=self.args.pass_task_to_policy,
dim_state=self.args.state_dim,
dim_latent=self.args.latent_dim * 2,
dim_belief=self.args.belief_dim,
dim_task=self.args.task_dim,
#
hidden_layers=self.args.policy_layers,
activation_function=self.args.policy_activation_function,
policy_initialisation=self.args.policy_initialisation,
#
action_space=self.envs.action_space,
init_std=self.args.policy_init_std,
).to(device)
# initialise policy trainer
if self.args.policy == 'a2c':
policy = A2C(
self.args,
policy_net,
self.args.policy_value_loss_coef,
self.args.policy_entropy_coef,
policy_optimiser=self.args.policy_optimiser,
policy_anneal_lr=self.args.policy_anneal_lr,
train_steps=self.num_updates,
optimiser_vae=self.vae.optimiser_vae,
lr=self.args.lr_policy,
eps=self.args.policy_eps,
)
elif self.args.policy == 'ppo':
policy = PPO(
self.args,
policy_net,
self.args.policy_value_loss_coef,
self.args.policy_entropy_coef,
policy_optimiser=self.args.policy_optimiser,
policy_anneal_lr=self.args.policy_anneal_lr,
train_steps=self.num_updates,
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,
optimiser_vae=self.vae.optimiser_vae,
)
else:
raise NotImplementedError
return policy
def train(self):
""" Main Meta-Training loop """
start_time = time.time()
# reset environments
prev_state, belief, task = utl.reset_env(self.envs, self.args)
# insert initial observation / embeddings to rollout storage
self.policy_storage.prev_state[0].copy_(prev_state)
# log once before training
with torch.no_grad():
self.log(None, None, start_time)
for self.iter_idx in range(self.num_updates):
# First, re-compute the hidden states given the current rollouts (since the VAE might've changed)
with torch.no_grad():
latent_sample, latent_mean, latent_logvar, hidden_state = self.encode_running_trajectory(
)
# add this initial hidden state to the policy storage
assert len(self.policy_storage.latent_mean
) == 0 # make sure we emptied buffers
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 = utl.select_action(
args=self.args,
policy=self.policy,
state=prev_state,
belief=belief,
task=task,
deterministic=False,
latent_sample=latent_sample,
latent_mean=latent_mean,
latent_logvar=latent_logvar,
)
# take step in the environment
[next_state, belief,
task], (rew_raw, rew_normalised), done, infos = utl.env_step(
self.envs, action, self.args)
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)
with torch.no_grad():
# 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_state,
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_kl_term):
self.vae.rollout_storage.insert(
prev_state.clone(),
action.detach().clone(), next_state.clone(),
rew_raw.clone(), done.clone(),
task.clone() if task is not None else None)
# add the obs before reset to the policy storage
self.policy_storage.next_state[step] = next_state.clone()
# reset environments that are done
done_indices = np.argwhere(done.cpu().flatten()).flatten()
if len(done_indices) > 0:
next_state, belief, task = utl.reset_env(
self.envs,
self.args,
indices=done_indices,
state=next_state)
# TODO: deal with resampling for posterior sampling algorithm
# latent_sample = latent_sample
# latent_sample[i] = latent_sample[i]
# add experience to policy buffer
self.policy_storage.insert(
state=next_state,
belief=belief,
task=task,
actions=action,
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),
latent_sample=latent_sample,
latent_mean=latent_mean,
latent_logvar=latent_logvar,
)
prev_state = next_state
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 > self.iter_idx:
for p in range(self.args.num_vae_updates_per_pretrain):
self.vae.compute_vae_loss(
update=True,
pretrain_index=self.iter_idx *
self.args.num_vae_updates_per_pretrain + p)
# otherwise do the normal update (policy + vae)
else:
train_stats = self.update(state=prev_state,
belief=belief,
task=task,
latent_sample=latent_sample,
latent_mean=latent_mean,
latent_logvar=latent_logvar)
# log
run_stats = [
action, self.policy_storage.action_log_probs, value
]
with torch.no_grad():
self.log(run_stats, train_stats, start_time)
# clean up after update
self.policy_storage.after_update()
self.envs.close()
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))
])).to(device)
latent_mean = (torch.stack([
all_latent_means[lens[i]][i] for i in range(len(lens))
])).to(device)
latent_logvar = (torch.stack([
all_latent_logvars[lens[i]][i] for i in range(len(lens))
])).to(device)
hidden_state = (torch.stack([
all_hidden_states[lens[i]][i] for i in range(len(lens))
])).to(device)
return latent_sample, latent_mean, latent_logvar, hidden_state
def get_value(self, state, belief, task, latent_sample, latent_mean,
latent_logvar):
latent = utl.get_latent_for_policy(self.args,
latent_sample=latent_sample,
latent_mean=latent_mean,
latent_logvar=latent_logvar)
return self.policy.actor_critic.get_value(state=state,
belief=belief,
task=task,
latent=latent).detach()
def update(self, state, belief, task, 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(state=state,
belief=belief,
task=task,
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(
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
def log(self, run_stats, train_stats, start_time):
# --- visualise behaviour of policy ---
if (self.iter_idx + 1) % self.args.vis_interval == 0:
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,
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,
tasks=self.train_tasks,
)
# --- evaluate policy ----
if (self.iter_idx + 1) % self.args.eval_interval == 0:
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,
ret_rms=ret_rms,
encoder=self.vae.encoder,
iter_idx=self.iter_idx,
tasks=self.train_tasks,
)
# 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(f"Updates {self.iter_idx}, "
f"Frames {self.frames}, "
f"FPS {int(self.frames / (time.time() - start_time))}, "
f"\n Mean return (train): {returns_avg[-1].item()} \n")
# --- save models ---
if (self.iter_idx + 1) % 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)
idx_labels = ['']
if self.args.save_intermediate_models:
idx_labels.append(int(self.iter_idx))
for idx_label in idx_labels:
torch.save(self.policy.actor_critic,
os.path.join(save_path, f"policy{idx_label}.pt"))
torch.save(self.vae.encoder,
os.path.join(save_path, f"encoder{idx_label}.pt"))
if self.vae.state_decoder is not None:
torch.save(
self.vae.state_decoder,
os.path.join(save_path,
f"state_decoder{idx_label}.pt"))
if self.vae.reward_decoder is not None:
torch.save(
self.vae.reward_decoder,
os.path.join(save_path,
f"reward_decoder{idx_label}.pt"))
if self.vae.task_decoder is not None:
torch.save(
self.vae.task_decoder,
os.path.join(save_path, f"task_decoder{idx_label}.pt"))
# save normalisation params of envs
if self.args.norm_rew_for_policy:
rew_rms = self.envs.venv.ret_rms
utl.save_obj(rew_rms, save_path, f"env_rew_rms{idx_label}")
# TODO: grab from policy and save?
# if self.args.norm_obs_for_policy:
# obs_rms = self.envs.venv.obs_rms
# utl.save_obj(obs_rms, save_path, f"env_obs_rms{idx_label}")
# --- log some other things ---
if ((self.iter_idx + 1) % self.args.log_interval
== 0) and (train_stats is not None):
self.logger.add('environment/state_max',
self.policy_storage.prev_state.max(),
self.iter_idx)
self.logger.add('environment/state_min',
self.policy_storage.prev_state.min(),
self.iter_idx)
self.logger.add('environment/rew_max',
self.policy_storage.rewards_raw.max(),
self.iter_idx)
self.logger.add('environment/rew_min',
self.policy_storage.rewards_raw.min(),
self.iter_idx)
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)