def setup_replay(args: argparse.Namespace, env: Env) -> ExperienceReplay:
D = ExperienceReplay(
args.experience_size,
env.observation_size,
env.action_size,
args.device
)
# Initialise dataset D with random seed episodes
for _ in range(1, args.seed_episodes + 1):
observation, done = env.reset(), False
while not done:
action = env.sample_random_action()
next_observation, _, done, info = env.step(action)
D.append(observation, action, info["reward_dist"], info["reward_coll"], done)
observation = next_observation
return D
args.action_repeat, args.bit_depth)
if args.load_experience:
D = torch.load(os.path.join(results_dir, 'experience.pth'))
metrics['steps'], metrics['episodes'] = [D.steps] * D.episodes, list(
range(1, D.episodes + 1))
else:
D = ExperienceReplay(args.experience_size, args.symbolic_env,
env.observation_size, env.action_size, args.bit_depth,
args.device)
# Initialise dataset D with S random seed episodes
for s in range(1, args.seed_episodes + 1):
observation, done, t = env.reset(), False, 0
while not done:
action = env.sample_random_action()
next_observation, reward, done = env.step(action)
D.append(observation, action, reward, done)
observation = next_observation
t += 1
metrics['steps'].append(t * args.action_repeat + (
0 if len(metrics['steps']) == 0 else metrics['steps'][-1]))
metrics['episodes'].append(s)
print("Initializing model parameters!")
# Initialise model parameters randomly
transition_model = TransitionModel(
args.belief_size, args.state_size, env.action_size, args.hidden_size,
args.embedding_size, args.activation_function).to(device=args.device)
observation_model = ObservationModel(
args.symbolic_env, env.observation_size, args.belief_size, args.state_size,
args.embedding_size, args.activation_function).to(device=args.device)
reward_model = RewardModel(args.belief_size, args.state_size, args.hidden_size,
class Initializer():
def __init__(self):
self.parms = Parameters()
self.results_dir = os.path.join(self.parms.results_path)
self.dataset_path = os.path.join(self.parms.results_path, 'dataset/')
os.makedirs(self.dataset_path, exist_ok=True)
self.metrics = {
'steps': [],
'episodes': [],
'train_rewards': [],
'predicted_rewards': [],
'test_episodes': [],
'test_rewards': [],
'observation_loss': [],
'reward_loss': [],
'kl_loss': [],
'regularizer_loss': []
}
os.makedirs(self.results_dir, exist_ok=True)
## Setting cuda options
if torch.cuda.is_available() and self.parms.use_cuda:
self.parms.device = torch.device('cuda')
torch.cuda.set_device(self.parms.gpu_id)
print("Using gpu: ", torch.cuda.current_device())
else:
self.parms.device = torch.device('cpu')
self.use_cuda = False
print("Work on: ", self.parms.device)
# Initilize buffer experience replay
self.env = ControlSuiteEnv(self.parms.env_name, self.parms.seed,
self.parms.max_episode_length,
self.parms.bit_depth)
self.D = ExperienceReplay(self.parms.ex_replay_buff_size,
self.env.observation_size,
self.env.action_size, self.parms.bit_depth,
self.parms.device)
if self.parms.seed > 0:
self.set_seed()
self.trainer = Trainer(self.parms, self.D, self.metrics,
self.results_dir, self.env)
self.init_exp_rep()
# Start Training
print("Total training episodes: ", self.parms.training_episodes,
" Buffer sampling: ", self.parms.collect_interval)
self.trainer.train_models()
print("END.")
def set_seed(self):
print("Setting seed")
os.environ['PYTHONHASHSEED'] = str(self.parms.seed)
random.seed(self.parms.seed)
np.random.seed(self.parms.seed)
torch.manual_seed(self.parms.seed)
if self.parms.use_cuda:
torch.cuda.manual_seed(self.parms.seed)
#torch.backends.cudnn.enabled=False # This makes the training slower
#torch.backends.cudnn.deterministic=True # This makes the training slower
# Init buffer experience replay
def init_exp_rep(self):
print("Starting initialization buffer.")
for s in tqdm(range(1, self.parms.num_init_episodes + 1)):
observation, done, t = self.env.reset(), False, 0
while not done:
action = self.env.sample_random_action()
next_observation, reward, done = self.env.step(action)
self.D.append(observation, action, reward, done)
observation = next_observation
t += 1
self.metrics['steps'].append(t * self.env.action_repeat +
(0 if len(self.metrics['steps']) ==
0 else self.metrics['steps'][-1]))
self.metrics['episodes'].append(s)
class Plan(object):
def __init__(self):
self.results_dir = os.path.join(
'results',
'{}_seed_{}_{}_action_scale_{}_no_explore_{}_pool_len_{}_optimisation_iters_{}_top_planning-horizon'
.format(args.env, args.seed, args.algo, args.action_scale,
args.pool_len, args.optimisation_iters,
args.top_planning_horizon))
args.results_dir = self.results_dir
args.MultiGPU = True if torch.cuda.device_count(
) > 1 and args.MultiGPU else False
self.__basic_setting()
self.__init_sample() # Sampleing The Init Data
# Initialise model parameters randomly
self.transition_model = TransitionModel(
args.belief_size, args.state_size, self.env.action_size,
args.hidden_size, args.embedding_size,
args.dense_activation_function).to(device=args.device)
self.observation_model = ObservationModel(
args.symbolic_env, self.env.observation_size, args.belief_size,
args.state_size, args.embedding_size,
args.cnn_activation_function).to(device=args.device)
self.reward_model = RewardModel(
args.belief_size, args.state_size, args.hidden_size,
args.dense_activation_function).to(device=args.device)
self.encoder = Encoder(
args.symbolic_env, self.env.observation_size, args.embedding_size,
args.cnn_activation_function).to(device=args.device)
print("We Have {} GPUS".format(torch.cuda.device_count())
) if args.MultiGPU else print("We use CPU")
self.transition_model = nn.DataParallel(
self.transition_model.to(device=args.device)
) if args.MultiGPU else self.transition_model
self.observation_model = nn.DataParallel(
self.observation_model.to(device=args.device)
) if args.MultiGPU else self.observation_model
self.reward_model = nn.DataParallel(
self.reward_model.to(
device=args.device)) if args.MultiGPU else self.reward_model
# encoder = nn.DataParallel(encoder.cuda())
# actor_model = nn.DataParallel(actor_model.cuda())
# value_model = nn.DataParallel(value_model.cuda())
# share the global parameters in multiprocessing
self.encoder.share_memory()
self.observation_model.share_memory()
self.reward_model.share_memory()
# Set all_model/global_actor_optimizer/global_value_optimizer
self.param_list = list(self.transition_model.parameters()) + list(
self.observation_model.parameters()) + list(
self.reward_model.parameters()) + list(
self.encoder.parameters())
self.model_optimizer = optim.Adam(
self.param_list,
lr=0
if args.learning_rate_schedule != 0 else args.model_learning_rate,
eps=args.adam_epsilon)
def update_belief_and_act(self,
args,
env,
belief,
posterior_state,
action,
observation,
explore=False):
# Infer belief over current state q(s_t|o≤t,a<t) from the history
# print("action size: ",action.size()) torch.Size([1, 6])
belief, _, _, _, posterior_state, _, _ = self.upper_transition_model(
posterior_state, action.unsqueeze(dim=0), belief,
self.encoder(observation).unsqueeze(dim=0), None)
if hasattr(env, "envs"):
belief, posterior_state = list(
map(lambda x: x.view(-1, args.test_episodes, x.shape[2]),
[x for x in [belief, posterior_state]]))
belief, posterior_state = belief.squeeze(
dim=0), posterior_state.squeeze(
dim=0) # Remove time dimension from belief/state
action = self.algorithms.get_action(belief, posterior_state, explore)
if explore:
action = torch.clamp(
Normal(action, args.action_noise).rsample(), -1, 1
) # Add gaussian exploration noise on top of the sampled action
# action = action + args.action_noise * torch.randn_like(action) # Add exploration noise ε ~ p(ε) to the action
next_observation, reward, done = env.step(
action.cpu() if isinstance(env, EnvBatcher) else action[0].cpu(
)) # Perform environment step (action repeats handled internally)
return belief, posterior_state, action, next_observation, reward, done
def run(self):
if args.algo == "dreamer":
print("DREAMER")
from algorithms.dreamer import Algorithms
self.algorithms = Algorithms(self.env.action_size,
self.transition_model, self.encoder,
self.reward_model,
self.observation_model)
elif args.algo == "p2p":
print("planing to plan")
from algorithms.plan_to_plan import Algorithms
self.algorithms = Algorithms(self.env.action_size,
self.transition_model, self.encoder,
self.reward_model,
self.observation_model)
elif args.algo == "actor_pool_1":
print("async sub actor")
from algorithms.actor_pool_1 import Algorithms_actor
self.algorithms = Algorithms_actor(self.env.action_size,
self.transition_model,
self.encoder, self.reward_model,
self.observation_model)
elif args.algo == "aap":
from algorithms.asynchronous_actor_planet import Algorithms
self.algorithms = Algorithms(self.env.action_size,
self.transition_model, self.encoder,
self.reward_model,
self.observation_model)
else:
print("planet")
from algorithms.planet import Algorithms
# args.MultiGPU = False
self.algorithms = Algorithms(self.env.action_size,
self.transition_model,
self.reward_model)
if args.test: self.test_only()
self.global_prior = Normal(
torch.zeros(args.batch_size, args.state_size, device=args.device),
torch.ones(args.batch_size, args.state_size,
device=args.device)) # Global prior N(0, I)
self.free_nats = torch.full(
(1, ), args.free_nats,
device=args.device) # Allowed deviation in KL divergence
# Training (and testing)
# args.episodes = 1
for episode in tqdm(range(self.metrics['episodes'][-1] + 1,
args.episodes + 1),
total=args.episodes,
initial=self.metrics['episodes'][-1] + 1):
losses = self.train()
# self.algorithms.save_loss_data(self.metrics['episodes']) # Update and plot loss metrics
self.save_loss_data(tuple(
zip(*losses))) # Update and plot loss metrics
self.data_collection(episode=episode) # Data collection
# args.test_interval = 1
if episode % args.test_interval == 0:
self.test(episode=episode) # Test model
self.save_model_data(episode=episode) # save model
self.env.close() # Close training environment
def train_env_model(self, beliefs, prior_states, prior_means,
prior_std_devs, posterior_states, posterior_means,
posterior_std_devs, observations, actions, rewards,
nonterminals):
# Calculate observation likelihood, reward likelihood and KL losses (for t = 0 only for latent overshooting); sum over final dims, average over batch and time (original implementation, though paper seems to miss 1/T scaling?)
if args.worldmodel_LogProbLoss:
observation_dist = Normal(
bottle(self.observation_model, (beliefs, posterior_states)), 1)
observation_loss = -observation_dist.log_prob(
observations[1:]).sum(
dim=2 if args.symbolic_env else (2, 3, 4)).mean(dim=(0, 1))
else:
observation_loss = F.mse_loss(
bottle(self.observation_model, (beliefs, posterior_states)),
observations[1:],
reduction='none').sum(
dim=2 if args.symbolic_env else (2, 3, 4)).mean(dim=(0, 1))
if args.worldmodel_LogProbLoss:
reward_dist = Normal(
bottle(self.reward_model, (beliefs, posterior_states)), 1)
reward_loss = -reward_dist.log_prob(rewards[:-1]).mean(dim=(0, 1))
else:
reward_loss = F.mse_loss(bottle(self.reward_model,
(beliefs, posterior_states)),
rewards[:-1],
reduction='none').mean(dim=(0, 1))
# transition loss
div = kl_divergence(Normal(posterior_means, posterior_std_devs),
Normal(prior_means, prior_std_devs)).sum(dim=2)
kl_loss = torch.max(div, self.free_nats).mean(
dim=(0, 1)
) # Note that normalisation by overshooting distance and weighting by overshooting distance cancel out
if args.global_kl_beta != 0:
kl_loss += args.global_kl_beta * kl_divergence(
Normal(posterior_means, posterior_std_devs),
self.global_prior).sum(dim=2).mean(dim=(0, 1))
# Calculate latent overshooting objective for t > 0
if args.overshooting_kl_beta != 0:
overshooting_vars = [
] # Collect variables for overshooting to process in batch
for t in range(1, args.chunk_size - 1):
d = min(t + args.overshooting_distance,
args.chunk_size - 1) # Overshooting distance
t_, d_ = t - 1, d - 1 # Use t_ and d_ to deal with different time indexing for latent states
seq_pad = (
0, 0, 0, 0, 0, t - d + args.overshooting_distance
) # Calculate sequence padding so overshooting terms can be calculated in one batch
# Store (0) actions, (1) nonterminals, (2) rewards, (3) beliefs, (4) prior states, (5) posterior means, (6) posterior standard deviations and (7) sequence masks
overshooting_vars.append(
(F.pad(actions[t:d],
seq_pad), F.pad(nonterminals[t:d], seq_pad),
F.pad(rewards[t:d],
seq_pad[2:]), beliefs[t_], prior_states[t_],
F.pad(posterior_means[t_ + 1:d_ + 1].detach(), seq_pad),
F.pad(posterior_std_devs[t_ + 1:d_ + 1].detach(),
seq_pad,
value=1),
F.pad(
torch.ones(d - t,
args.batch_size,
args.state_size,
device=args.device), seq_pad))
) # Posterior standard deviations must be padded with > 0 to prevent infinite KL divergences
overshooting_vars = tuple(zip(*overshooting_vars))
# Update belief/state using prior from previous belief/state and previous action (over entire sequence at once)
beliefs, prior_states, prior_means, prior_std_devs = self.upper_transition_model(
torch.cat(overshooting_vars[4], dim=0),
torch.cat(overshooting_vars[0], dim=1),
torch.cat(overshooting_vars[3], dim=0), None,
torch.cat(overshooting_vars[1], dim=1))
seq_mask = torch.cat(overshooting_vars[7], dim=1)
# Calculate overshooting KL loss with sequence mask
kl_loss += (
1 / args.overshooting_distance
) * args.overshooting_kl_beta * torch.max((kl_divergence(
Normal(torch.cat(overshooting_vars[5], dim=1),
torch.cat(overshooting_vars[6], dim=1)),
Normal(prior_means, prior_std_devs)
) * seq_mask).sum(dim=2), self.free_nats).mean(dim=(0, 1)) * (
args.chunk_size
- 1
) # Update KL loss (compensating for extra average over each overshooting/open loop sequence)
# Calculate overshooting reward prediction loss with sequence mask
if args.overshooting_reward_scale != 0:
reward_loss += (
1 / args.overshooting_distance
) * args.overshooting_reward_scale * F.mse_loss(
bottle(self.reward_model,
(beliefs, prior_states)) * seq_mask[:, :, 0],
torch.cat(overshooting_vars[2], dim=1),
reduction='none'
).mean(dim=(0, 1)) * (
args.chunk_size - 1
) # Update reward loss (compensating for extra average over each overshooting/open loop sequence)
# Apply linearly ramping learning rate schedule
if args.learning_rate_schedule != 0:
for group in self.model_optimizer.param_groups:
group['lr'] = min(
group['lr'] + args.model_learning_rate /
args.model_learning_rate_schedule,
args.model_learning_rate)
model_loss = observation_loss + reward_loss + kl_loss
# Update model parameters
self.model_optimizer.zero_grad()
model_loss.backward()
nn.utils.clip_grad_norm_(self.param_list,
args.grad_clip_norm,
norm_type=2)
self.model_optimizer.step()
return observation_loss, reward_loss, kl_loss
def train(self):
# Model fitting
losses = []
print("training loop")
# args.collect_interval = 1
for s in tqdm(range(args.collect_interval)):
# Draw sequence chunks {(o_t, a_t, r_t+1, terminal_t+1)} ~ D uniformly at random from the dataset (including terminal flags)
observations, actions, rewards, nonterminals = self.D.sample(
args.batch_size,
args.chunk_size) # Transitions start at time t = 0
# Create initial belief and state for time t = 0
init_belief, init_state = torch.zeros(
args.batch_size, args.belief_size,
device=args.device), torch.zeros(args.batch_size,
args.state_size,
device=args.device)
# Update belief/state using posterior from previous belief/state, previous action and current observation (over entire sequence at once)
obs = bottle(self.encoder, (observations[1:], ))
beliefs, prior_states, prior_means, prior_std_devs, posterior_states, posterior_means, posterior_std_devs = self.upper_transition_model(
prev_state=init_state,
actions=actions[:-1],
prev_belief=init_belief,
obs=obs,
nonterminals=nonterminals[:-1])
# Calculate observation likelihood, reward likelihood and KL losses (for t = 0 only for latent overshooting); sum over final dims, average over batch and time (original implementation, though paper seems to miss 1/T scaling?)
observation_loss, reward_loss, kl_loss = self.train_env_model(
beliefs, prior_states, prior_means, prior_std_devs,
posterior_states, posterior_means, posterior_std_devs,
observations, actions, rewards, nonterminals)
# Dreamer implementation: actor loss calculation and optimization
with torch.no_grad():
actor_states = posterior_states.detach().to(
device=args.device).share_memory_()
actor_beliefs = beliefs.detach().to(
device=args.device).share_memory_()
# if not os.path.exists(os.path.join(os.getcwd(), 'tensor_data/' + args.results_dir)): os.mkdir(os.path.join(os.getcwd(), 'tensor_data/' + args.results_dir))
torch.save(
actor_states,
os.path.join(os.getcwd(),
args.results_dir + '/actor_states.pt'))
torch.save(
actor_beliefs,
os.path.join(os.getcwd(),
args.results_dir + '/actor_beliefs.pt'))
# [self.actor_pipes[i][0].send(1) for i, w in enumerate(self.workers_actor)] # Parent_pipe send data using i'th pipes
# [self.actor_pipes[i][0].recv() for i, _ in enumerate(self.actor_pool)] # waitting the children finish
self.algorithms.train_algorithm(actor_states, actor_beliefs)
losses.append(
[observation_loss.item(),
reward_loss.item(),
kl_loss.item()])
# if self.algorithms.train_algorithm(actor_states, actor_beliefs) is not None:
# merge_actor_loss, merge_value_loss = self.algorithms.train_algorithm(actor_states, actor_beliefs)
# losses.append([observation_loss.item(), reward_loss.item(), kl_loss.item(), merge_actor_loss.item(), merge_value_loss.item()])
# else:
# losses.append([observation_loss.item(), reward_loss.item(), kl_loss.item()])
return losses
def data_collection(self, episode):
print("Data collection")
with torch.no_grad():
observation, total_reward = self.env.reset(), 0
belief, posterior_state, action = torch.zeros(
1, args.belief_size, device=args.device), torch.zeros(
1, args.state_size,
device=args.device), torch.zeros(1,
self.env.action_size,
device=args.device)
pbar = tqdm(range(args.max_episode_length // args.action_repeat))
for t in pbar:
# print("step",t)
belief, posterior_state, action, next_observation, reward, done = self.update_belief_and_act(
args, self.env, belief, posterior_state, action,
observation.to(device=args.device))
self.D.append(observation, action.cpu(), reward, done)
total_reward += reward
observation = next_observation
if args.render: self.env.render()
if done:
pbar.close()
break
# Update and plot train reward metrics
self.metrics['steps'].append(t + self.metrics['steps'][-1])
self.metrics['episodes'].append(episode)
self.metrics['train_rewards'].append(total_reward)
Save_Txt(self.metrics['episodes'][-1],
self.metrics['train_rewards'][-1], 'train_rewards',
args.results_dir)
# lineplot(metrics['episodes'][-len(metrics['train_rewards']):], metrics['train_rewards'], 'train_rewards', results_dir)
def test(self, episode):
print("Test model")
# Set models to eval mode
self.transition_model.eval()
self.observation_model.eval()
self.reward_model.eval()
self.encoder.eval()
self.algorithms.train_to_eval()
# self.actor_model_g.eval()
# self.value_model_g.eval()
# Initialise parallelised test environments
test_envs = EnvBatcher(
Env, (args.env, args.symbolic_env, args.seed,
args.max_episode_length, args.action_repeat, args.bit_depth),
{}, args.test_episodes)
with torch.no_grad():
observation, total_rewards, video_frames = test_envs.reset(
), np.zeros((args.test_episodes, )), []
belief, posterior_state, action = torch.zeros(
args.test_episodes, args.belief_size,
device=args.device), torch.zeros(
args.test_episodes, args.state_size,
device=args.device), torch.zeros(args.test_episodes,
self.env.action_size,
device=args.device)
pbar = tqdm(range(args.max_episode_length // args.action_repeat))
for t in pbar:
belief, posterior_state, action, next_observation, reward, done = self.update_belief_and_act(
args, test_envs, belief, posterior_state, action,
observation.to(device=args.device))
total_rewards += reward.numpy()
if not args.symbolic_env: # Collect real vs. predicted frames for video
video_frames.append(
make_grid(torch.cat([
observation,
self.observation_model(belief,
posterior_state).cpu()
],
dim=3) + 0.5,
nrow=5).numpy()) # Decentre
observation = next_observation
if done.sum().item() == args.test_episodes:
pbar.close()
break
# Update and plot reward metrics (and write video if applicable) and save metrics
self.metrics['test_episodes'].append(episode)
self.metrics['test_rewards'].append(total_rewards.tolist())
Save_Txt(self.metrics['test_episodes'][-1],
self.metrics['test_rewards'][-1], 'test_rewards',
args.results_dir)
# Save_Txt(np.asarray(metrics['steps'])[np.asarray(metrics['test_episodes']) - 1], metrics['test_rewards'],'test_rewards_steps', results_dir, xaxis='step')
# lineplot(metrics['test_episodes'], metrics['test_rewards'], 'test_rewards', results_dir)
# lineplot(np.asarray(metrics['steps'])[np.asarray(metrics['test_episodes']) - 1], metrics['test_rewards'], 'test_rewards_steps', results_dir, xaxis='step')
if not args.symbolic_env:
episode_str = str(episode).zfill(len(str(args.episodes)))
write_video(video_frames, 'test_episode_%s' % episode_str,
args.results_dir) # Lossy compression
save_image(
torch.as_tensor(video_frames[-1]),
os.path.join(args.results_dir,
'test_episode_%s.png' % episode_str))
torch.save(self.metrics, os.path.join(args.results_dir, 'metrics.pth'))
# Set models to train mode
self.transition_model.train()
self.observation_model.train()
self.reward_model.train()
self.encoder.train()
# self.actor_model_g.train()
# self.value_model_g.train()
self.algorithms.eval_to_train()
# Close test environments
test_envs.close()
def test_only(self):
# Set models to eval mode
self.transition_model.eval()
self.reward_model.eval()
self.encoder.eval()
with torch.no_grad():
total_reward = 0
for _ in tqdm(range(args.test_episodes)):
observation = self.env.reset()
belief, posterior_state, action = torch.zeros(
1, args.belief_size, device=args.device), torch.zeros(
1, args.state_size,
device=args.device), torch.zeros(1,
self.env.action_size,
device=args.device)
pbar = tqdm(
range(args.max_episode_length // args.action_repeat))
for t in pbar:
belief, posterior_state, action, observation, reward, done = self.update_belief_and_act(
args, self.env, belief, posterior_state, action,
observation.to(evice=args.device))
total_reward += reward
if args.render: self.env.render()
if done:
pbar.close()
break
print('Average Reward:', total_reward / args.test_episodes)
self.env.close()
quit()
def __basic_setting(self):
args.overshooting_distance = min(
args.chunk_size, args.overshooting_distance
) # Overshooting distance cannot be greater than chunk size
print(' ' * 26 + 'Options')
for k, v in vars(args).items():
print(' ' * 26 + k + ': ' + str(v))
print("torch.cuda.device_count() {}".format(torch.cuda.device_count()))
os.makedirs(args.results_dir, exist_ok=True)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
# Set Cuda
if torch.cuda.is_available() and not args.disable_cuda:
print("using CUDA")
args.device = torch.device('cuda')
torch.cuda.manual_seed(args.seed)
else:
print("using CPU")
args.device = torch.device('cpu')
self.summary_name = args.results_dir + "/{}_{}_log"
self.writer = SummaryWriter(self.summary_name.format(
args.env, args.id))
self.env = Env(args.env, args.symbolic_env, args.seed,
args.max_episode_length, args.action_repeat,
args.bit_depth)
self.metrics = {
'steps': [],
'episodes': [],
'train_rewards': [],
'test_episodes': [],
'test_rewards': [],
'observation_loss': [],
'reward_loss': [],
'kl_loss': [],
'merge_actor_loss': [],
'merge_value_loss': []
}
def __init_sample(self):
if args.experience_replay is not '' and os.path.exists(
args.experience_replay):
self.D = torch.load(args.experience_replay)
self.metrics['steps'], self.metrics['episodes'] = [
self.D.steps
] * self.D.episodes, list(range(1, self.D.episodes + 1))
elif not args.test:
self.D = ExperienceReplay(args.experience_size, args.symbolic_env,
self.env.observation_size,
self.env.action_size, args.bit_depth,
args.device)
# Initialise dataset D with S random seed episodes
print(
"Start Multi Sample Processing -------------------------------"
)
start_time = time.time()
data_lists = [
Manager().list() for i in range(1, args.seed_episodes + 1)
] # Set Global Lists
pipes = [Pipe() for i in range(1, args.seed_episodes + 1)
] # Set Multi Pipe
workers_init_sample = [
Worker_init_Sample(child_conn=child, id=i + 1)
for i, [parent, child] in enumerate(pipes)
]
for i, w in enumerate(workers_init_sample):
w.start() # Start Single Process
pipes[i][0].send(
data_lists[i]) # Parent_pipe send data using i'th pipes
[w.join() for w in workers_init_sample] # wait sub_process done
for i, [parent, child] in enumerate(pipes):
# datas = parent.recv()
for data in list(parent.recv()):
if isinstance(data, tuple):
assert len(data) == 4
self.D.append(data[0], data[1], data[2], data[3])
elif isinstance(data, int):
t = data
self.metrics['steps'].append(t * args.action_repeat + (
0 if len(self.metrics['steps']) ==
0 else self.metrics['steps'][-1]))
self.metrics['episodes'].append(i + 1)
else:
print(
"The Recvive Data Have Some Problems, Need To Fix")
end_time = time.time()
print("the process times {} s".format(end_time - start_time))
print(
"End Multi Sample Processing -------------------------------")
def upper_transition_model(self, prev_state, actions, prev_belief, obs,
nonterminals):
actions = torch.transpose(actions, 0, 1) if args.MultiGPU else actions
nonterminals = torch.transpose(nonterminals, 0, 1).to(
device=args.device
) if args.MultiGPU and nonterminals is not None else nonterminals
obs = torch.transpose(obs, 0, 1).to(
device=args.device) if args.MultiGPU and obs is not None else obs
temp_val = self.transition_model(prev_state.to(device=args.device),
actions.to(device=args.device),
prev_belief.to(device=args.device),
obs, nonterminals)
return list(
map(
lambda x: torch.cat(x.chunk(torch.cuda.device_count(), 0), 1)
if x.shape[1] != prev_state.shape[0] else x,
[x for x in temp_val]))
def save_loss_data(self, losses):
self.metrics['observation_loss'].append(losses[0])
self.metrics['reward_loss'].append(losses[1])
self.metrics['kl_loss'].append(losses[2])
self.metrics['merge_actor_loss'].append(
losses[3]) if losses.__len__() > 3 else None
self.metrics['merge_value_loss'].append(
losses[4]) if losses.__len__() > 3 else None
Save_Txt(self.metrics['episodes'][-1],
self.metrics['observation_loss'][-1], 'observation_loss',
args.results_dir)
Save_Txt(self.metrics['episodes'][-1], self.metrics['reward_loss'][-1],
'reward_loss', args.results_dir)
Save_Txt(self.metrics['episodes'][-1], self.metrics['kl_loss'][-1],
'kl_loss', args.results_dir)
Save_Txt(self.metrics['episodes'][-1],
self.metrics['merge_actor_loss'][-1], 'merge_actor_loss',
args.results_dir) if losses.__len__() > 3 else None
Save_Txt(self.metrics['episodes'][-1],
self.metrics['merge_value_loss'][-1], 'merge_value_loss',
args.results_dir) if losses.__len__() > 3 else None
# lineplot(metrics['episodes'][-len(metrics['observation_loss']):], metrics['observation_loss'], 'observation_loss', results_dir)
# lineplot(metrics['episodes'][-len(metrics['reward_loss']):], metrics['reward_loss'], 'reward_loss', results_dir)
# lineplot(metrics['episodes'][-len(metrics['kl_loss']):], metrics['kl_loss'], 'kl_loss', results_dir)
# lineplot(metrics['episodes'][-len(metrics['actor_loss']):], metrics['actor_loss'], 'actor_loss', results_dir)
# lineplot(metrics['episodes'][-len(metrics['value_loss']):], metrics['value_loss'], 'value_loss', results_dir)
def save_model_data(self, episode):
# writer.add_scalar("train_reward", metrics['train_rewards'][-1], metrics['steps'][-1])
# writer.add_scalar("train/episode_reward", metrics['train_rewards'][-1], metrics['steps'][-1]*args.action_repeat)
# writer.add_scalar("observation_loss", metrics['observation_loss'][0][-1], metrics['steps'][-1])
# writer.add_scalar("reward_loss", metrics['reward_loss'][0][-1], metrics['steps'][-1])
# writer.add_scalar("kl_loss", metrics['kl_loss'][0][-1], metrics['steps'][-1])
# writer.add_scalar("actor_loss", metrics['actor_loss'][0][-1], metrics['steps'][-1])
# writer.add_scalar("value_loss", metrics['value_loss'][0][-1], metrics['steps'][-1])
# print("episodes: {}, total_steps: {}, train_reward: {} ".format(metrics['episodes'][-1], metrics['steps'][-1], metrics['train_rewards'][-1]))
# Checkpoint models
if episode % args.checkpoint_interval == 0:
# torch.save({'transition_model': transition_model.state_dict(),
# 'observation_model': observation_model.state_dict(),
# 'reward_model': reward_model.state_dict(),
# 'encoder': encoder.state_dict(),
# 'actor_model': actor_model_g.state_dict(),
# 'value_model': value_model_g.state_dict(),
# 'model_optimizer': model_optimizer.state_dict(),
# 'actor_optimizer': actor_optimizer_g.state_dict(),
# 'value_optimizer': value_optimizer_g.state_dict()
# }, os.path.join(results_dir, 'models_%d.pth' % episode))
if args.checkpoint_experience:
torch.save(
self.D, os.path.join(args.results_dir, 'experience.pth')
) # Warning: will fail with MemoryError with large memory sizes
class Dreamer(Agent):
# The agent has its own replay buffer, update, act
def __init__(self, args):
"""
All paras are passed by args
:param args: a dict that includes parameters
"""
super().__init__()
self.args = args
# Initialise model parameters randomly
self.transition_model = TransitionModel(
args.belief_size, args.state_size, args.action_size,
args.hidden_size, args.embedding_size,
args.dense_act).to(device=args.device)
self.observation_model = ObservationModel(
args.symbolic,
args.observation_size,
args.belief_size,
args.state_size,
args.embedding_size,
activation_function=(args.dense_act if args.symbolic else
args.cnn_act)).to(device=args.device)
self.reward_model = RewardModel(args.belief_size, args.state_size,
args.hidden_size,
args.dense_act).to(device=args.device)
self.encoder = Encoder(args.symbolic, args.observation_size,
args.embedding_size,
args.cnn_act).to(device=args.device)
self.actor_model = ActorModel(
args.action_size,
args.belief_size,
args.state_size,
args.hidden_size,
activation_function=args.dense_act,
fix_speed=args.fix_speed,
throttle_base=args.throttle_base).to(device=args.device)
self.value_model = ValueModel(args.belief_size, args.state_size,
args.hidden_size,
args.dense_act).to(device=args.device)
self.value_model2 = ValueModel(args.belief_size, args.state_size,
args.hidden_size,
args.dense_act).to(device=args.device)
self.pcont_model = PCONTModel(args.belief_size, args.state_size,
args.hidden_size,
args.dense_act).to(device=args.device)
self.target_value_model = deepcopy(self.value_model)
self.target_value_model2 = deepcopy(self.value_model2)
for p in self.target_value_model.parameters():
p.requires_grad = False
for p in self.target_value_model2.parameters():
p.requires_grad = False
# setup the paras to update
self.world_param = list(self.transition_model.parameters())\
+ list(self.observation_model.parameters())\
+ list(self.reward_model.parameters())\
+ list(self.encoder.parameters())
if args.pcont:
self.world_param += list(self.pcont_model.parameters())
# setup optimizer
self.world_optimizer = optim.Adam(self.world_param, lr=args.world_lr)
self.actor_optimizer = optim.Adam(self.actor_model.parameters(),
lr=args.actor_lr)
self.value_optimizer = optim.Adam(list(self.value_model.parameters()) +
list(self.value_model2.parameters()),
lr=args.value_lr)
# setup the free_nat to
self.free_nats = torch.full(
(1, ), args.free_nats, dtype=torch.float32,
device=args.device) # Allowed deviation in KL divergence
# TODO: change it to the new replay buffer, in buffer.py
self.D = ExperienceReplay(args.experience_size, args.symbolic,
args.observation_size, args.action_size,
args.bit_depth, args.device)
if self.args.auto_temp:
# setup for learning of alpha term (temp of the entropy term)
self.log_temp = torch.zeros(1,
requires_grad=True,
device=args.device)
self.target_entropy = -np.prod(
args.action_size if not args.fix_speed else self.args.
action_size - 1).item() # heuristic value from SAC paper
self.temp_optimizer = optim.Adam(
[self.log_temp], lr=args.value_lr) # use the same value_lr
# TODO: print out the param used in Dreamer
# var_counts = tuple(count_vars(module) for module in [self., self.ac.q1, self.ac.q2])
# print('\nNumber of parameters: \t pi: %d, \t q1: %d, \t q2: %d\n' % var_counts)
# def process_im(self, image, image_size=None, rgb=None):
# # Resize, put channel first, convert it to a tensor, centre it to [-0.5, 0.5] and add batch dimenstion.
#
# def preprocess_observation_(observation, bit_depth):
# # Preprocesses an observation inplace (from float32 Tensor [0, 255] to [-0.5, 0.5])
# observation.div_(2 ** (8 - bit_depth)).floor_().div_(2 ** bit_depth).sub_(
# 0.5) # Quantise to given bit depth and centre
# observation.add_(torch.rand_like(observation).div_(
# 2 ** bit_depth)) # Dequantise (to approx. match likelihood of PDF of continuous images vs. PMF of discrete images)
#
# image = image[40:, :, :] # clip the above 40 rows
# image = torch.tensor(cv2.resize(image, (40, 40), interpolation=cv2.INTER_LINEAR).transpose(2, 0, 1),
# dtype=torch.float32) # Resize and put channel first
#
# preprocess_observation_(image, self.args.bit_depth)
# return image.unsqueeze(dim=0)
def process_im(self, images, image_size=None, rgb=None):
images = cv2.resize(images, (40, 40))
images = np.dot(images, [0.299, 0.587, 0.114])
obs = torch.tensor(images,
dtype=torch.float32).div_(255.).sub_(0.5).unsqueeze(
dim=0) # shape [1, 40, 40], range:[-0.5,0.5]
return obs.unsqueeze(dim=0) # add batch dimension
def append_buffer(self, new_traj):
# append new collected trajectory, not implement the data augmentation
# shape of new_traj: [(o, a, r, d) * steps]
for state in new_traj:
observation, action, reward, done = state
self.D.append(observation, action.cpu(), reward, done)
def _compute_loss_world(self, state, data):
# unpackage data
beliefs, prior_states, prior_means, prior_std_devs, posterior_states, posterior_means, posterior_std_devs = state
observations, rewards, nonterminals = data
# observation_loss = F.mse_loss(
# bottle(self.observation_model, (beliefs, posterior_states)),
# observations[1:],
# reduction='none').sum(dim=2 if self.args.symbolic else (2, 3, 4)).mean(dim=(0, 1))
#
# reward_loss = F.mse_loss(
# bottle(self.reward_model, (beliefs, posterior_states)),
# rewards[1:],
# reduction='none').mean(dim=(0,1))
observation_loss = F.mse_loss(
bottle(self.observation_model, (beliefs, posterior_states)),
observations,
reduction='none').sum(
dim=2 if self.args.symbolic else (2, 3, 4)).mean(dim=(0, 1))
reward_loss = F.mse_loss(bottle(self.reward_model,
(beliefs, posterior_states)),
rewards,
reduction='none').mean(dim=(0, 1)) # TODO: 5
# transition loss
kl_loss = torch.max(
kl_divergence(
Independent(Normal(posterior_means, posterior_std_devs), 1),
Independent(Normal(prior_means, prior_std_devs), 1)),
self.free_nats).mean(dim=(0, 1))
# print("check the reward", bottle(pcont_model, (beliefs, posterior_states)).shape, nonterminals[:-1].shape)
if self.args.pcont:
pcont_loss = F.binary_cross_entropy(
bottle(self.pcont_model, (beliefs, posterior_states)),
nonterminals)
# pcont_pred = torch.distributions.Bernoulli(logits=bottle(self.pcont_model, (beliefs, posterior_states)))
# pcont_loss = -pcont_pred.log_prob(nonterminals[1:]).mean(dim=(0, 1))
return observation_loss, self.args.reward_scale * reward_loss, kl_loss, (
self.args.pcont_scale * pcont_loss if self.args.pcont else 0)
def _compute_loss_actor(self,
imag_beliefs,
imag_states,
imag_ac_logps=None):
# reward and value prediction of imagined trajectories
imag_rewards = bottle(self.reward_model, (imag_beliefs, imag_states))
imag_values = bottle(self.value_model, (imag_beliefs, imag_states))
imag_values2 = bottle(self.value_model2, (imag_beliefs, imag_states))
imag_values = torch.min(imag_values, imag_values2)
with torch.no_grad():
if self.args.pcont:
pcont = bottle(self.pcont_model, (imag_beliefs, imag_states))
else:
pcont = self.args.discount * torch.ones_like(imag_rewards)
pcont = pcont.detach()
if imag_ac_logps is not None:
imag_values[
1:] -= self.args.temp * imag_ac_logps # add entropy here
returns = cal_returns(imag_rewards[:-1],
imag_values[:-1],
imag_values[-1],
pcont[:-1],
lambda_=self.args.disclam)
discount = torch.cumprod(
torch.cat([torch.ones_like(pcont[:1]), pcont[:-2]], 0), 0)
discount = discount.detach()
assert list(discount.size()) == list(returns.size())
actor_loss = -torch.mean(discount * returns)
return actor_loss
def _compute_loss_critic(self,
imag_beliefs,
imag_states,
imag_ac_logps=None):
with torch.no_grad():
# calculate the target with the target nn
target_imag_values = bottle(self.target_value_model,
(imag_beliefs, imag_states))
target_imag_values2 = bottle(self.target_value_model2,
(imag_beliefs, imag_states))
target_imag_values = torch.min(target_imag_values,
target_imag_values2)
imag_rewards = bottle(self.reward_model,
(imag_beliefs, imag_states))
if self.args.pcont:
pcont = bottle(self.pcont_model, (imag_beliefs, imag_states))
else:
pcont = self.args.discount * torch.ones_like(imag_rewards)
# print("check pcont", pcont)
if imag_ac_logps is not None:
target_imag_values[1:] -= self.args.temp * imag_ac_logps
returns = cal_returns(imag_rewards[:-1],
target_imag_values[:-1],
target_imag_values[-1],
pcont[:-1],
lambda_=self.args.disclam)
target_return = returns.detach()
value_pred = bottle(self.value_model, (imag_beliefs, imag_states))[:-1]
value_pred2 = bottle(self.value_model2,
(imag_beliefs, imag_states))[:-1]
value_loss = F.mse_loss(value_pred, target_return,
reduction="none").mean(dim=(0, 1))
value_loss2 = F.mse_loss(value_pred2, target_return,
reduction="none").mean(dim=(0, 1))
value_loss += value_loss2
return value_loss
def _latent_imagination(self,
beliefs,
posterior_states,
with_logprob=False):
# Rollout to generate imagined trajectories
chunk_size, batch_size, _ = list(
posterior_states.size()) # flatten the tensor
flatten_size = chunk_size * batch_size
posterior_states = posterior_states.detach().reshape(flatten_size, -1)
beliefs = beliefs.detach().reshape(flatten_size, -1)
imag_beliefs, imag_states, imag_ac_logps = [beliefs
], [posterior_states], []
for i in range(self.args.planning_horizon):
imag_action, imag_ac_logp = self.actor_model(
imag_beliefs[-1].detach(),
imag_states[-1].detach(),
deterministic=False,
with_logprob=with_logprob,
)
imag_action = imag_action.unsqueeze(dim=0) # add time dim
# print(imag_states[-1].shape, imag_action.shape, imag_beliefs[-1].shape)
imag_belief, imag_state, _, _ = self.transition_model(
imag_states[-1], imag_action, imag_beliefs[-1])
imag_beliefs.append(imag_belief.squeeze(dim=0))
imag_states.append(imag_state.squeeze(dim=0))
if with_logprob:
imag_ac_logps.append(imag_ac_logp.squeeze(dim=0))
imag_beliefs = torch.stack(imag_beliefs, dim=0).to(
self.args.device
) # shape [horizon+1, (chuck-1)*batch, belief_size]
imag_states = torch.stack(imag_states, dim=0).to(self.args.device)
if with_logprob:
imag_ac_logps = torch.stack(imag_ac_logps, dim=0).to(
self.args.device) # shape [horizon, (chuck-1)*batch]
return imag_beliefs, imag_states, imag_ac_logps if with_logprob else None
def update_parameters(self, gradient_steps):
loss_info = [] # used to record loss
for s in tqdm(range(gradient_steps)):
# get state and belief of samples
observations, actions, rewards, nonterminals = self.D.sample(
self.args.batch_size, self.args.chunk_size)
# print("check sampled rewrads", rewards)
init_belief = torch.zeros(self.args.batch_size,
self.args.belief_size,
device=self.args.device)
init_state = torch.zeros(self.args.batch_size,
self.args.state_size,
device=self.args.device)
# Update belief/state using posterior from previous belief/state, previous action and current observation (over entire sequence at once)
# beliefs, prior_states, prior_means, prior_std_devs, posterior_states, posterior_means, posterior_std_devs = self.transition_model(
# init_state,
# actions[:-1],
# init_belief,
# bottle(self.encoder, (observations[1:], )),
# nonterminals[:-1])
beliefs, prior_states, prior_means, prior_std_devs, posterior_states, posterior_means, posterior_std_devs = self.transition_model(
init_state, actions, init_belief,
bottle(self.encoder, (observations, )),
nonterminals) # TODO: 4
# update paras of world model
world_model_loss = self._compute_loss_world(
state=(beliefs, prior_states, prior_means, prior_std_devs,
posterior_states, posterior_means, posterior_std_devs),
data=(observations, rewards, nonterminals))
observation_loss, reward_loss, kl_loss, pcont_loss = world_model_loss
self.world_optimizer.zero_grad()
(observation_loss + reward_loss + kl_loss + pcont_loss).backward()
nn.utils.clip_grad_norm_(self.world_param,
self.args.grad_clip_norm,
norm_type=2)
self.world_optimizer.step()
# freeze params to save memory
for p in self.world_param:
p.requires_grad = False
for p in self.value_model.parameters():
p.requires_grad = False
for p in self.value_model2.parameters():
p.requires_gard = False
# latent imagination
imag_beliefs, imag_states, imag_ac_logps = self._latent_imagination(
beliefs, posterior_states, with_logprob=self.args.with_logprob)
# update temp
if self.args.auto_temp:
temp_loss = -(
self.log_temp *
(imag_ac_logps[0] + self.target_entropy).detach()).mean()
self.temp_optimizer.zero_grad()
temp_loss.backward()
self.temp_optimizer.step()
self.args.temp = self.log_temp.exp()
# update actor
actor_loss = self._compute_loss_actor(imag_beliefs,
imag_states,
imag_ac_logps=imag_ac_logps)
self.actor_optimizer.zero_grad()
actor_loss.backward()
nn.utils.clip_grad_norm_(self.actor_model.parameters(),
self.args.grad_clip_norm,
norm_type=2)
self.actor_optimizer.step()
for p in self.world_param:
p.requires_grad = True
for p in self.value_model.parameters():
p.requires_grad = True
for p in self.value_model2.parameters():
p.requires_grad = True
# update critic
imag_beliefs = imag_beliefs.detach()
imag_states = imag_states.detach()
critic_loss = self._compute_loss_critic(
imag_beliefs, imag_states, imag_ac_logps=imag_ac_logps)
self.value_optimizer.zero_grad()
critic_loss.backward()
nn.utils.clip_grad_norm_(self.value_model.parameters(),
self.args.grad_clip_norm,
norm_type=2)
nn.utils.clip_grad_norm_(self.value_model2.parameters(),
self.args.grad_clip_norm,
norm_type=2)
self.value_optimizer.step()
loss_info.append([
observation_loss.item(),
reward_loss.item(),
kl_loss.item(),
pcont_loss.item() if self.args.pcont else 0,
actor_loss.item(),
critic_loss.item()
])
# finally, update target value function every #gradient_steps
with torch.no_grad():
self.target_value_model.load_state_dict(
self.value_model.state_dict())
with torch.no_grad():
self.target_value_model2.load_state_dict(
self.value_model2.state_dict())
return loss_info
def infer_state(self, observation, action, belief=None, state=None):
""" Infer belief over current state q(s_t|o≤t,a<t) from the history,
return updated belief and posterior_state at time t
returned shape: belief/state [belief/state_dim] (remove the time_dim)
"""
# observation is obs.to(device), action.shape=[act_dim] (will add time dim inside this fn), belief.shape
belief, _, _, _, posterior_state, _, _ = self.transition_model(
state, action.unsqueeze(dim=0), belief,
self.encoder(observation).unsqueeze(
dim=0)) # Action and observation need extra time dimension
belief, posterior_state = belief.squeeze(
dim=0), posterior_state.squeeze(
dim=0) # Remove time dimension from belief/state
return belief, posterior_state
def select_action(self, state, deterministic=False):
# get action with the inputs get from fn: infer_state; return a numpy with shape [batch, act_size]
belief, posterior_state = state
action, _ = self.actor_model(belief,
posterior_state,
deterministic=deterministic,
with_logprob=False)
if not deterministic and not self.args.with_logprob:
print("e")
action = Normal(action, self.args.expl_amount).rsample()
# clip the angle
action[:, 0].clamp_(min=self.args.angle_min,
max=self.args.angle_max)
# clip the throttle
if self.args.fix_speed:
action[:, 1] = self.args.throttle_base
else:
action[:, 1].clamp_(min=self.args.throttle_min,
max=self.args.throttle_max)
print("action", action)
# return action.cup().numpy()
return action # this is a Tonsor.cuda
def import_parameters(self, params):
# only import or export the parameters used when local rollout
self.encoder.load_state_dict(params["encoder"])
self.actor_model.load_state_dict(params["policy"])
self.transition_model.load_state_dict(params["transition"])
def export_parameters(self):
""" return the model paras used for local rollout """
params = {
"encoder": self.encoder.cpu().state_dict(),
"policy": self.actor_model.cpu().state_dict(),
"transition": self.transition_model.cpu().state_dict()
}
self.encoder.to(self.args.device)
self.actor_model.to(self.args.device)
self.transition_model.to(self.args.device)
return params
args.observation_size, args.action_size = env.observation_size, env.action_size
# Initialise agent
agent = Dreamer(args)
D = ExperienceReplay(args.experience_size, args.symbolic, env.observation_size,
env.action_size, args.bit_depth, args.device)
# Initialise dataset D with S random seed episodes
for s in range(1, args.seed_episodes + 1):
observation, done, t = env.reset(), False, 0
while not done:
action = env.sample_random_action()
next_observation, reward, done = env.step(action)
D.append(next_observation, action.cpu(), reward,
done) # here use the next_observation
observation = next_observation
t += 1
metrics['env_steps'].append(t * args.action_repeat + (
0 if len(metrics['env_steps']) == 0 else metrics['env_steps'][-1]))
metrics['episodes'].append(s)
print("(random)episodes: {}, total_env_steps: {} ".format(
metrics['episodes'][-1], metrics['env_steps'][-1]))
print("--- Finish random data collection --- ")
if args.models and os.path.exists(args.models):
model_dicts = torch.load(args.models)
agent.transition_model.load_state_dict(model_dicts['transition_model'])
agent.observation_model.load_state_dict(model_dicts['observation_model'])
agent.reward_model.load_state_dict(model_dicts['reward_model1'])
def train(args: argparse.Namespace,
env: Env,
D: ExperienceReplay,
models: Tuple[nn.Module, nn.Module, nn.Module, nn.Module],
optimizer: Tuple[optim.Optimizer, optim.Optimizer],
param_list: List[nn.parameter.Parameter],
planner: nn.Module):
# auxilliary tensors
global_prior = Normal(
torch.zeros(args.batch_size, args.state_size, device=args.device),
torch.ones(args.batch_size, args.state_size, device=args.device)
) # Global prior N(0, I)
# Allowed deviation in KL divergence
free_nats = torch.full((1, ), args.free_nats, dtype=torch.float32, device=args.device)
summary_writter = SummaryWriter(args.tensorboard_dir)
# unpack models
transition_model, observation_model, reward_model, encoder = models
transition_optimizer, reward_optimizer = optimizer
for idx_episode in trange(args.episodes, leave=False, desc="Episode"):
for idx_train in trange(args.collect_interval, leave=False, desc="Training"):
# Draw sequence chunks {(o[t], a[t], r[t+1], z[t+1])} ~ D uniformly at random from the dataset
# The first two dimensions of the tensors are L (chunk size) and n (batch size)
# We want to use o[t+1] to correct the error of the transition model,
# so we need to convert the sequence to {(o[t+1], a[t], r[t+1], z[t+1])}
observations, actions, rewards_dist, rewards_coll, nonterminals = D.sample(args.batch_size, args.chunk_size)
# Create initial belief and state for time t = 0
init_belief = torch.zeros(args.batch_size, args.belief_size, device=args.device)
init_state = torch.zeros(args.batch_size, args.state_size, device=args.device)
# Transition model forward
# deterministic: h[t+1] = f(h[t], a[t])
# prior: s[t+1] ~ Prob(s|h[t+1])
# posterior: s[t+1] ~ Prob(s|h[t+1], o[t+1])
beliefs, prior_states, prior_means, prior_std_devs, posterior_states, posterior_means, posterior_std_devs = transition_model(
init_state,
actions[:-1],
init_belief,
bottle(encoder, (observations[1:], )),
nonterminals[:-1]
)
# observation loss
predictions = bottle(observation_model, (beliefs, posterior_states))
visual_loss = F.mse_loss(
predictions[:, :, :3*64*64],
observations[1:, :, :3*64*64]
).mean()
symbol_loss = F.mse_loss(
predictions[:, :, 3*64*64:],
observations[1:, :, 3*64*64:]
).mean()
observation_loss = visual_loss + symbol_loss
# KL divergence loss. Minimize the difference between posterior and prior
kl_loss = torch.max(
kl_divergence(
Normal(posterior_means, posterior_std_devs),
Normal(prior_means, prior_std_devs)
).sum(dim=2),
free_nats
).mean(dim=(0, 1)) # Note that normalisation by overshooting distance and weighting by overshooting distance cancel out
if args.global_kl_beta != 0:
kl_loss += args.global_kl_beta * kl_divergence(
Normal(posterior_means, posterior_std_devs),
global_prior
).sum(dim=2).mean(dim=(0, 1))
# overshooting loss
if args.overshooting_kl_beta != 0:
overshooting_vars = [] # Collect variables for overshooting to process in batch
for t in range(1, args.chunk_size - 1):
d = min(t + args.overshooting_distance, args.chunk_size - 1) # Overshooting distance
# Use t_ and d_ to deal with different time indexing for latent states
t_, d_ = t - 1, d - 1
# Calculate sequence padding so overshooting terms can be calculated in one batch
seq_pad = (0, 0, 0, 0, 0, t - d + args.overshooting_distance)
# Store
# * a[t:d],
# * z[t+1:d+1]
# * r[t+1:d+1]
# * h[t]
# * s[t] prior
# * E[s[t:d]] posterior
# * Var[s[t:d]] posterior
# * mask:
# the last few sequences do not have enough length,
# so we pad it with 0 to the same length as previous sequence for batch operation,
# and use mask to indicate invalid variables.
overshooting_vars.append(
(F.pad(actions[t:d], seq_pad),
F.pad(nonterminals[t:d], seq_pad),
F.pad(rewards_dist[t:d], seq_pad[2:]),
beliefs[t_],
prior_states[t_],
F.pad(posterior_means[t_ + 1:d_ + 1].detach(), seq_pad),
F.pad(posterior_std_devs[t_ + 1:d_ + 1].detach(), seq_pad, value=1),
F.pad(torch.ones(d - t, args.batch_size, args.state_size, device=args.device), seq_pad)
)
) # Posterior standard deviations must be padded with > 0 to prevent infinite KL divergences
overshooting_vars = tuple(zip(*overshooting_vars))
# Update belief/state using prior from previous belief/state and previous action (over entire sequence at once)
beliefs, prior_states, prior_means, prior_std_devs = transition_model(
torch.cat(overshooting_vars[4], dim=0),
torch.cat(overshooting_vars[0], dim=1),
torch.cat(overshooting_vars[3], dim=0),
None,
torch.cat(overshooting_vars[1], dim=1)
)
seq_mask = torch.cat(overshooting_vars[7], dim=1)
# Calculate overshooting KL loss with sequence mask
kl_loss += (1 / args.overshooting_distance) * args.overshooting_kl_beta * torch.max(
(kl_divergence(
Normal(torch.cat(overshooting_vars[5], dim=1), torch.cat(overshooting_vars[6], dim=1)),
Normal(prior_means, prior_std_devs)
) * seq_mask).sum(dim=2),
free_nats
).mean(dim=(0, 1)) * (args.chunk_size - 1) # Update KL loss (compensating for extra average over each overshooting/open loop sequence)
# TODO: add learning rate schedule
# Update model parameters
transition_optimizer.zero_grad()
loss = observation_loss * 200 + kl_loss
loss.backward()
nn.utils.clip_grad_norm_(param_list, args.grad_clip_norm, norm_type=2)
transition_optimizer.step()
# reward loss
rewards_dist_predict, rewards_coll_predict = bottle(reward_model.raw, (beliefs.detach(), posterior_states.detach()))
reward_loss = F.mse_loss(
rewards_dist_predict,
rewards_dist[:-1],
reduction='mean'
) + F.binary_cross_entropy(
rewards_coll_predict,
rewards_coll[:-1],
reduction='mean'
)
reward_optimizer.zero_grad()
reward_loss.backward()
reward_optimizer.step()
# add tensorboard log
global_step = idx_train + idx_episode * args.collect_interval
summary_writter.add_scalar("observation_loss", observation_loss, global_step)
summary_writter.add_scalar("reward_loss", reward_loss, global_step)
summary_writter.add_scalar("kl_loss", kl_loss, global_step)
for idx_collect in trange(1, leave=False, desc="Collecting"):
experience = collect_experience(args, env, models, planner, True, desc="Collecting experience {}".format(idx_collect))
T = len(experience["observation"])
for idx_step in range(T):
D.append(experience["observation"][idx_step],
experience["action"][idx_step],
experience["reward_dist"][idx_step],
experience["reward_coll"][idx_step],
experience["done"][idx_step])
# Checkpoint models
if (idx_episode + 1) % args.checkpoint_interval == 0:
record_path = os.path.join(args.checkpoint_dir, "checkpoint")
checkpoint_path = os.path.join(args.checkpoint_dir, 'models_%d.pth' % (idx_episode+1))
torch.save(
{
'transition_model': transition_model.state_dict(),
'observation_model': observation_model.state_dict(),
'reward_model': reward_model.state_dict(),
'encoder': encoder.state_dict(),
'transition_optimizer': transition_optimizer.state_dict(),
'reward_optimizer': reward_optimizer.state_dict()
},
checkpoint_path)
with open(record_path, "w") as f:
f.write('models_%d.pth' % (idx_episode+1))
planner.save(os.path.join(args.torchscript_dir, "mpc_planner.pth"))
transition_model.save(os.path.join(args.torchscript_dir, "transition_model.pth"))
reward_model.save(os.path.join(args.torchscript_dir, "reward_model.pth"))
observation_model.save(os.path.join(args.torchscript_dir, "observation_decoder.pth"))
encoder.save(os.path.join(args.torchscript_dir, "observation_encoder.pth"))
summary_writter.close()