in train_curiosity_exploration.py [0:0]
def main():
torch.set_num_threads(1)
device = torch.device("cuda:0" if args.cuda else "cpu")
ndevices = torch.cuda.device_count()
# Setup loggers
tbwriter = SummaryWriter(log_dir=args.log_dir)
logging.basicConfig(filename=f"{args.log_dir}/train_log.txt", level=logging.DEBUG)
logging.getLogger().addHandler(logging.StreamHandler(sys.stdout))
logging.getLogger().setLevel(logging.INFO)
if "habitat" in args.env_name:
devices = [int(dev) for dev in os.environ["CUDA_VISIBLE_DEVICES"].split(",")]
# Devices need to be indexed between 0 to N-1
devices = [dev for dev in range(len(devices))]
if len(devices) > 2:
devices = devices[1:]
envs = make_vec_envs_habitat(
args.habitat_config_file, device, devices, seed=args.seed
)
else:
train_log_dir = os.path.join(args.log_dir, "train_monitor")
try:
os.makedirs(train_log_dir)
except OSError:
pass
envs = make_vec_envs_avd(
args.env_name,
args.seed,
args.num_processes,
train_log_dir,
device,
True,
num_frame_stack=1,
split="train",
nRef=args.num_pose_refs,
)
args.feat_shape_sim = (512,)
args.feat_shape_pose = (512 * 9,)
args.obs_shape = envs.observation_space.spaces["im"].shape
# =================== Create models ====================
if args.encoder_type == "rgb":
encoder = RGBEncoder(fix_cnn=args.fix_cnn)
elif args.encoder_type == "rgb+map":
encoder = MapRGBEncoder(fix_cnn=args.fix_cnn)
else:
raise ValueError(f"encoder_type {args.encoder_type} not defined!")
action_config = (
{"nactions": envs.action_space.n, "embedding_size": args.action_embedding_size}
if args.use_action_embedding
else None
)
collision_config = (
{"collision_dim": 2, "embedding_size": args.collision_embedding_size}
if args.use_collision_embedding
else None
)
actor_critic = Policy(
envs.action_space,
base_kwargs={
"feat_dim": args.feat_shape_sim[0],
"recurrent": True,
"hidden_size": args.feat_shape_sim[0],
"action_config": action_config,
"collision_config": collision_config,
},
)
icm_phi = Phi() if args.icm_embedding_type == "imagenet" else None
icm_fd = ForwardDynamics(envs.action_space.n)
# =================== Load models ====================
save_path = os.path.join(args.save_dir, "checkpoints")
checkpoint_path = os.path.join(save_path, "ckpt.latest.pth")
if os.path.isfile(checkpoint_path):
logging.info("Resuming from old model!")
loaded_states = torch.load(checkpoint_path)
encoder_state, actor_critic_state, icm_fd_state, j_start = loaded_states
encoder.load_state_dict(encoder_state)
actor_critic.load_state_dict(actor_critic_state)
icm_fd.load_state_dict(icm_fd_state)
elif args.pretrained_il_model != "":
logging.info("Initializing with pre-trained model!")
encoder_state, actor_critic_state, _ = torch.load(args.pretrained_il_model)
encoder.load_state_dict(encoder_state)
actor_critic.load_state_dict(actor_critic_state)
j_start = -1
else:
j_start = -1
encoder.to(device)
actor_critic.to(device)
if args.icm_embedding_type == "imagenet":
icm_phi.to(device)
icm_fd.to(device)
encoder.eval()
actor_critic.eval()
if args.icm_embedding_type == "imagenet":
icm_phi.eval() # Do not train/the feature model for ICM
icm_fd.eval()
# =================== Define ICM training algorithm ====================
icm_optimizer = optim.Adam(icm_fd.parameters(), lr=args.lr)
# Maintain a running mean of the variance of returns after every
# num-rl-steps
if args.normalize_icm_rewards:
args.returns_rms = RunningMeanStd()
# =================== Define RL training algorithm ====================
rl_algo_config = {}
rl_algo_config["lr"] = args.lr
rl_algo_config["eps"] = args.eps
rl_algo_config["encoder_type"] = args.encoder_type
rl_algo_config["max_grad_norm"] = args.max_grad_norm
rl_algo_config["clip_param"] = args.clip_param
rl_algo_config["ppo_epoch"] = args.ppo_epoch
rl_algo_config["entropy_coef"] = args.entropy_coef
rl_algo_config["num_mini_batch"] = args.num_mini_batch
rl_algo_config["value_loss_coef"] = args.value_loss_coef
rl_algo_config["use_clipped_value_loss"] = False
rl_algo_config["nactions"] = envs.action_space.n
rl_algo_config["encoder"] = encoder
rl_algo_config["actor_critic"] = actor_critic
rl_algo_config["use_action_embedding"] = args.use_action_embedding
rl_algo_config["use_collision_embedding"] = args.use_collision_embedding
rl_agent = PPO(rl_algo_config)
# =================== Define stats buffer ====================
train_metrics_tracker = defaultdict(lambda: deque(maxlen=10))
# =================== Define rollouts ====================
rollouts_policy = RolloutStoragePPO(
args.num_rl_steps,
args.num_processes,
args.obs_shape,
envs.action_space,
args.feat_shape_sim[0],
encoder_type=args.encoder_type,
)
rollouts_policy.to(device)
def get_obs(obs):
obs_im = process_image(obs["im"])
if args.encoder_type == "rgb+map":
obs_lm = process_image(obs["coarse_occupancy"])
obs_sm = process_image(obs["fine_occupancy"])
else:
obs_lm = None
obs_sm = None
return obs_im, obs_sm, obs_lm
start = time.time()
NPROC = args.num_processes
for j in range(j_start + 1, num_updates):
# =================== Start a new episode ====================
obs = envs.reset()
# Reset ICM data buffer
all_icm_feats = []
all_icm_acts = []
# Set icm models to evaluate mode for data gathering
if args.icm_embedding_type == "imagenet":
icm_phi.eval()
icm_fd.eval()
# Processing environment inputs
obs_im, obs_sm, obs_lm = get_obs(obs) # (num_processes, 3, 84, 84)
obs_collns = obs["collisions"].long() # (num_processes, 1)
# Initialize the memory of rollouts for policy
rollouts_policy.reset()
rollouts_policy.obs_im[0].copy_(obs_im)
if args.encoder_type == "rgb+map":
rollouts_policy.obs_sm[0].copy_(obs_sm)
rollouts_policy.obs_lm[0].copy_(obs_lm)
rollouts_policy.collisions[0].copy_(obs_collns)
# Episode statistics
episode_expl_rewards = np.zeros((NPROC, 1))
episode_collisions = np.zeros((NPROC, 1))
episode_collisions += obs_collns.cpu().numpy()
# Metrics
osr_tracker = [0.0 for _ in range(NPROC)]
objects_tracker = [0.0 for _ in range(NPROC)]
area_tracker = [0.0 for _ in range(NPROC)]
novelty_tracker = [0.0 for _ in range(NPROC)]
smooth_coverage_tracker = [0.0 for _ in range(NPROC)]
per_proc_area = [0.0 for _ in range(NPROC)]
# Other states
prev_action = torch.zeros(NPROC, 1).to(device)
prev_collision = obs_collns
action_onehot = torch.zeros(NPROC, envs.action_space.n).to(
device
) # (N, n_actions)
# ================= Update over a full batch of episodes =================
# num_steps must be total number of steps in each episode
for step in range(args.num_steps):
pstep = rollouts_policy.step
with torch.no_grad():
encoder_inputs = [rollouts_policy.obs_im[pstep]]
if args.encoder_type == "rgb+map":
encoder_inputs.append(rollouts_policy.obs_sm[pstep])
encoder_inputs.append(rollouts_policy.obs_lm[pstep])
obs_feats = encoder(*encoder_inputs)
policy_inputs = {"features": obs_feats}
if args.use_action_embedding:
policy_inputs["actions"] = prev_action.long()
if args.use_collision_embedding:
policy_inputs["collisions"] = prev_collision.long()
policy_outputs = actor_critic.act(
policy_inputs,
rollouts_policy.recurrent_hidden_states[pstep],
rollouts_policy.masks[pstep],
)
(
value,
action,
action_log_probs,
recurrent_hidden_states,
) = policy_outputs
# Gather curiosity experience. By default, the features are deatached
# from the forward dynamics loss.
if args.icm_embedding_type == "imagenet":
with torch.no_grad():
icm_feats = icm_phi(obs_im)
else:
icm_feats = recurrent_hidden_states
all_icm_feats.append(icm_feats)
all_icm_acts.append(action)
# Act, get reward and next obs
obs, reward, done, infos = envs.step(action)
# Processing environment inputs
obs_im, obs_sm, obs_lm = get_obs(obs) # (num_processes, 3, 84, 84)
obs_collns = obs["collisions"] # (num_processes, 1)
# Always set masks to 1 (since this loop happens within one episode)
masks = torch.FloatTensor([[1.0] for _ in range(NPROC)]).to(device)
# Compute curiosity rewards for the previous action (not the current)
reward_exploration = torch.zeros(NPROC, 1)
if step >= 1:
phi_st = all_icm_feats[-2]
phi_st1 = all_icm_feats[-1]
action_onehot.zero_()
act = all_icm_acts[-2]
action_onehot.scatter_(1, act, 1)
with torch.no_grad():
phi_st1_hat = icm_fd(phi_st, action_onehot)
reward_exploration = (
F.mse_loss(phi_st1_hat, phi_st1, reduction="none")
.sum(dim=1)
.unsqueeze(1)
.detach()
) # (N, 1)
# Since this reward corresponds to the previous action, update it
# accordingly in the rollouts buffer.
rollouts_policy.update_prev_rewards(
reward_exploration * args.reward_scale
)
reward_exploration = reward_exploration.cpu()
for proc in range(NPROC):
seen_area = float(infos[proc]["seen_area"])
objects_visited = infos[proc].get("num_objects_visited", 0.0)
oracle_success = float(infos[proc].get("oracle_pose_success", 0.0))
novelty_reward = infos[proc].get("count_based_reward", 0.0)
smooth_coverage_reward = infos[proc].get("coverage_novelty_reward", 0.0)
area_tracker[proc] = seen_area
objects_tracker[proc] = objects_visited
osr_tracker[proc] = oracle_success
per_proc_area[proc] = seen_area
novelty_tracker[proc] += novelty_reward
smooth_coverage_tracker[proc] += smooth_coverage_reward
# Instrinsic reward is updated separately (delayed by 1 time step)
overall_reward = reward * (1 - args.reward_scale)
# Update statistics
episode_expl_rewards += reward_exploration.numpy() * args.reward_scale
# Update rollouts_policy
rollouts_policy.insert(
obs_im,
obs_sm,
obs_lm,
recurrent_hidden_states,
action,
action_log_probs,
value,
overall_reward,
masks,
obs_collns,
)
# Update prev values
prev_collision = obs_collns
prev_action = action
episode_collisions += obs_collns.cpu().numpy()
# Update RL policy
if (step + 1) % args.num_rl_steps == 0:
# Update value function for last step
with torch.no_grad():
encoder_inputs = [rollouts_policy.obs_im[-1]]
if args.encoder_type == "rgb+map":
encoder_inputs.append(rollouts_policy.obs_sm[-1])
encoder_inputs.append(rollouts_policy.obs_lm[-1])
obs_feats = encoder(*encoder_inputs)
policy_inputs = {"features": obs_feats}
if args.use_action_embedding:
policy_inputs["actions"] = prev_action.long()
if args.use_collision_embedding:
policy_inputs["collisions"] = prev_collision.long()
next_value = actor_critic.get_value(
policy_inputs,
rollouts_policy.recurrent_hidden_states[-1],
rollouts_policy.masks[-1],
).detach()
# Normalize the rewards if applicable
if args.normalize_icm_rewards:
current_returns = 0.0
for rew in torch.flip(rollouts_policy.rewards, dims=[0]):
current_returns = current_returns * args.gamma + rew
current_returns = current_returns.squeeze(1).cpu().numpy()
args.returns_rms.update(current_returns)
rollouts_policy.rewards /= args.returns_rms.var.item()
# Compute returns
rollouts_policy.compute_returns(
next_value, args.use_gae, args.gamma, args.tau,
)
encoder.train()
actor_critic.train()
# Update model
rl_losses = rl_agent.update(rollouts_policy)
# Refresh rollouts
rollouts_policy.after_update()
encoder.eval()
actor_critic.eval()
# ============ Update the ICM dynamics model using past data ===============
icm_fd.train()
action_onehot = torch.zeros(NPROC, envs.action_space.n).to(
device
) # (N, n_actions)
avg_fd_loss = 0
avg_fd_loss_count = 0
icm_update_count = 0
for t in random_range(0, args.num_steps - 1):
phi_st = all_icm_feats[t] # (N, 512)
phi_st1 = all_icm_feats[t + 1] # (N, 512)
action_onehot.zero_()
at = all_icm_acts[t].long() # (N, 1)
action_onehot.scatter_(1, at, 1)
# Forward pass
phi_st1_hat = icm_fd(phi_st, action_onehot)
fd_loss = F.mse_loss(phi_st1_hat, phi_st1)
# Backward pass
icm_optimizer.zero_grad()
fd_loss.backward()
torch.nn.utils.clip_grad_norm_(icm_fd.parameters(), args.max_grad_norm)
# Update step
icm_optimizer.step()
avg_fd_loss += fd_loss.item()
avg_fd_loss_count += phi_st1_hat.shape[0]
avg_fd_loss /= avg_fd_loss_count
all_losses = {"icm_fd_loss": avg_fd_loss}
icm_fd.eval()
# =================== Save model ====================
if (j + 1) % args.save_interval == 0 and args.save_dir != "":
save_path = os.path.join(args.save_dir, "checkpoints")
try:
os.makedirs(save_path)
except OSError:
pass
encoder_state = encoder.state_dict()
actor_critic_state = actor_critic.state_dict()
icm_fd_state = icm_fd.state_dict()
torch.save(
[encoder_state, actor_critic_state, icm_fd_state, j],
f"{save_path}/ckpt.latest.pth",
)
if args.save_unique:
torch.save(
[encoder_state, actor_critic_state, icm_fd_state, j],
f"{save_path}/ckpt.{(j+1):07d}.pth",
)
# =================== Logging data ====================
total_num_steps = (j + 1 - j_start) * NPROC * args.num_steps
if j % args.log_interval == 0:
end = time.time()
fps = int(total_num_steps / (end - start))
logging.info(f"===> Updates {j}, #steps {total_num_steps}, FPS {fps}")
train_metrics = rl_losses
train_metrics.update(all_losses)
train_metrics["exploration_rewards"] = np.mean(episode_expl_rewards)
train_metrics["area_covered"] = np.mean(per_proc_area)
train_metrics["objects_covered"] = np.mean(objects_tracker)
train_metrics["landmarks_covered"] = np.mean(osr_tracker)
train_metrics["collisions"] = np.mean(episode_collisions)
train_metrics["novelty_rewards"] = np.mean(novelty_tracker)
train_metrics["smooth_coverage_rewards"] = np.mean(smooth_coverage_tracker)
# Update statistics
for k, v in train_metrics.items():
train_metrics_tracker[k].append(v)
for k, v in train_metrics_tracker.items():
logging.info(f"{k}: {np.mean(v).item():.3f}")
tbwriter.add_scalar(f"train_metrics/{k}", np.mean(v).item(), j)
# =================== Evaluate models ====================
if args.eval_interval is not None and (j + 1) % args.eval_interval == 0:
if "habitat" in args.env_name:
devices = [
int(dev) for dev in os.environ["CUDA_VISIBLE_DEVICES"].split(",")
]
# Devices need to be indexed between 0 to N-1
devices = [dev for dev in range(len(devices))]
eval_envs = make_vec_envs_habitat(
args.eval_habitat_config_file, device, devices
)
else:
eval_envs = make_vec_envs_avd(
args.env_name,
args.seed + 12,
12,
eval_log_dir,
device,
True,
split="val",
nRef=args.num_pose_refs,
set_return_topdown_map=True,
)
num_eval_episodes = 16 if "habitat" in args.env_name else 30
eval_config = {}
eval_config["num_steps"] = args.num_steps
eval_config["feat_shape_sim"] = args.feat_shape_sim
eval_config["num_processes"] = 1 if "habitat" in args.env_name else 12
eval_config["num_pose_refs"] = args.num_pose_refs
eval_config["num_eval_episodes"] = num_eval_episodes
eval_config["env_name"] = args.env_name
eval_config["actor_type"] = "learned_policy"
eval_config["encoder_type"] = args.encoder_type
eval_config["use_action_embedding"] = args.use_action_embedding
eval_config["use_collision_embedding"] = args.use_collision_embedding
eval_config[
"vis_save_dir"
] = f"{args.save_dir}/policy_vis/update_{(j+1):05d}"
models = {}
models["encoder"] = encoder
models["actor_critic"] = actor_critic
val_metrics, _ = evaluate_visitation(
models, eval_envs, eval_config, device, visualize_policy=False
)
for k, v in val_metrics.items():
tbwriter.add_scalar(f"val_metrics/{k}", v, j)
tbwriter.close()