def train(self) -> None:
r"""Main method for training PPO.
Returns:
None
"""
profiling_wrapper.configure(
capture_start_step=self.config.PROFILING.CAPTURE_START_STEP,
num_steps_to_capture=self.config.PROFILING.NUM_STEPS_TO_CAPTURE,
)
self.envs = construct_envs(self.config,
get_env_class(self.config.ENV_NAME))
ppo_cfg = self.config.RL.PPO
self.device = (torch.device("cuda", self.config.TORCH_GPU_ID)
if torch.cuda.is_available() else torch.device("cpu"))
if not os.path.isdir(self.config.CHECKPOINT_FOLDER):
os.makedirs(self.config.CHECKPOINT_FOLDER)
self._setup_actor_critic_agent(ppo_cfg)
logger.info("agent number of parameters: {}".format(
sum(param.numel() for param in self.agent.parameters())))
rollouts = RolloutStorage(
ppo_cfg.num_steps,
self.envs.num_envs,
self.obs_space,
self.envs.action_spaces[0],
ppo_cfg.hidden_size,
)
rollouts.to(self.device)
observations = self.envs.reset()
batch = batch_obs(observations, device=self.device)
batch = apply_obs_transforms_batch(batch, self.obs_transforms)
for sensor in rollouts.observations:
rollouts.observations[sensor][0].copy_(batch[sensor])
# batch and observations may contain shared PyTorch CUDA
# tensors. We must explicitly clear them here otherwise
# they will be kept in memory for the entire duration of training!
batch = None
observations = None
current_episode_reward = torch.zeros(self.envs.num_envs, 1)
running_episode_stats = dict(
count=torch.zeros(self.envs.num_envs, 1),
reward=torch.zeros(self.envs.num_envs, 1),
)
window_episode_stats: DefaultDict[str, deque] = defaultdict(
lambda: deque(maxlen=ppo_cfg.reward_window_size))
t_start = time.time()
env_time = 0
pth_time = 0
count_steps = 0
count_checkpoints = 0
lr_scheduler = LambdaLR(
optimizer=self.agent.optimizer,
lr_lambda=lambda x: linear_decay(x, self.config.NUM_UPDATES
), # type: ignore
)
with TensorboardWriter(self.config.TENSORBOARD_DIR,
flush_secs=self.flush_secs) as writer:
for update in range(self.config.NUM_UPDATES):
profiling_wrapper.on_start_step()
profiling_wrapper.range_push("train update")
if ppo_cfg.use_linear_lr_decay:
lr_scheduler.step() # type: ignore
if ppo_cfg.use_linear_clip_decay:
self.agent.clip_param = ppo_cfg.clip_param * linear_decay(
update, self.config.NUM_UPDATES)
profiling_wrapper.range_push("rollouts loop")
for _step in range(ppo_cfg.num_steps):
(
delta_pth_time,
delta_env_time,
delta_steps,
) = self._collect_rollout_step(rollouts,
current_episode_reward,
running_episode_stats)
pth_time += delta_pth_time
env_time += delta_env_time
count_steps += delta_steps
profiling_wrapper.range_pop() # rollouts loop
(
delta_pth_time,
value_loss,
action_loss,
dist_entropy,
) = self._update_agent(ppo_cfg, rollouts)
pth_time += delta_pth_time
for k, v in running_episode_stats.items():
window_episode_stats[k].append(v.clone())
deltas = {
k:
((v[-1] -
v[0]).sum().item() if len(v) > 1 else v[0].sum().item())
for k, v in window_episode_stats.items()
}
deltas["count"] = max(deltas["count"], 1.0)
writer.add_scalar("reward", deltas["reward"] / deltas["count"],
count_steps)
# Check to see if there are any metrics
# that haven't been logged yet
metrics = {
k: v / deltas["count"]
for k, v in deltas.items() if k not in {"reward", "count"}
}
if len(metrics) > 0:
writer.add_scalars("metrics", metrics, count_steps)
losses = [value_loss, action_loss]
writer.add_scalars(
"losses",
{k: l
for l, k in zip(losses, ["value", "policy"])},
count_steps,
)
# log stats
if update > 0 and update % self.config.LOG_INTERVAL == 0:
logger.info("update: {}\tfps: {:.3f}\t".format(
update, count_steps / (time.time() - t_start)))
logger.info(
"update: {}\tenv-time: {:.3f}s\tpth-time: {:.3f}s\t"
"frames: {}".format(update, env_time, pth_time,
count_steps))
logger.info("Average window size: {} {}".format(
len(window_episode_stats["count"]),
" ".join("{}: {:.3f}".format(k, v / deltas["count"])
for k, v in deltas.items() if k != "count"),
))
# checkpoint model
if update % self.config.CHECKPOINT_INTERVAL == 0:
self.save_checkpoint(f"ckpt.{count_checkpoints}.pth",
dict(step=count_steps))
count_checkpoints += 1
profiling_wrapper.range_pop() # train update
self.envs.close()
def _init_train(self):
if self.config.RL.DDPPO.force_distributed:
self._is_distributed = True
if is_slurm_batch_job():
add_signal_handlers()
if self._is_distributed:
local_rank, tcp_store = init_distrib_slurm(
self.config.RL.DDPPO.distrib_backend
)
if rank0_only():
logger.info(
"Initialized DD-PPO with {} workers".format(
torch.distributed.get_world_size()
)
)
self.config.defrost()
self.config.TORCH_GPU_ID = local_rank
self.config.SIMULATOR_GPU_ID = local_rank
# Multiply by the number of simulators to make sure they also get unique seeds
self.config.TASK_CONFIG.SEED += (
torch.distributed.get_rank() * self.config.NUM_ENVIRONMENTS
)
self.config.freeze()
random.seed(self.config.TASK_CONFIG.SEED)
np.random.seed(self.config.TASK_CONFIG.SEED)
torch.manual_seed(self.config.TASK_CONFIG.SEED)
self.num_rollouts_done_store = torch.distributed.PrefixStore(
"rollout_tracker", tcp_store
)
self.num_rollouts_done_store.set("num_done", "0")
if rank0_only() and self.config.VERBOSE:
logger.info(f"config: {self.config}")
profiling_wrapper.configure(
capture_start_step=self.config.PROFILING.CAPTURE_START_STEP,
num_steps_to_capture=self.config.PROFILING.NUM_STEPS_TO_CAPTURE,
)
self._init_envs()
ppo_cfg = self.config.RL.PPO
if torch.cuda.is_available():
self.device = torch.device("cuda", self.config.TORCH_GPU_ID)
torch.cuda.set_device(self.device)
else:
self.device = torch.device("cpu")
if rank0_only() and not os.path.isdir(self.config.CHECKPOINT_FOLDER):
os.makedirs(self.config.CHECKPOINT_FOLDER)
self._setup_actor_critic_agent(ppo_cfg)
if self._is_distributed:
self.agent.init_distributed(find_unused_params=True)
logger.info(
"agent number of parameters: {}".format(
sum(param.numel() for param in self.agent.parameters())
)
)
obs_space = self.obs_space
if self._static_encoder:
self._encoder = self.actor_critic.net.visual_encoder
obs_space = spaces.Dict(
{
"visual_features": spaces.Box(
low=np.finfo(np.float32).min,
high=np.finfo(np.float32).max,
shape=self._encoder.output_shape,
dtype=np.float32,
),
**obs_space.spaces,
}
)
self._nbuffers = 2 if ppo_cfg.use_double_buffered_sampler else 1
self.rollouts = RolloutStorage(
ppo_cfg.num_steps,
self.envs.num_envs,
obs_space,
self.envs.action_spaces[0],
ppo_cfg.hidden_size,
num_recurrent_layers=self.actor_critic.net.num_recurrent_layers,
is_double_buffered=ppo_cfg.use_double_buffered_sampler,
)
self.rollouts.to(self.device)
observations = self.envs.reset()
batch = batch_obs(
observations, device=self.device, cache=self._obs_batching_cache
)
batch = apply_obs_transforms_batch(batch, self.obs_transforms)
if self._static_encoder:
with torch.no_grad():
batch["visual_features"] = self._encoder(batch)
self.rollouts.buffers["observations"][0] = batch
self.current_episode_reward = torch.zeros(self.envs.num_envs, 1)
self.running_episode_stats = dict(
count=torch.zeros(self.envs.num_envs, 1),
reward=torch.zeros(self.envs.num_envs, 1),
)
self.window_episode_stats = defaultdict(
lambda: deque(maxlen=ppo_cfg.reward_window_size)
)
self.env_time = 0.0
self.pth_time = 0.0
self.t_start = time.time()
def train(self) -> None:
r"""Main method for DD-PPO.
Returns:
None
"""
self.local_rank, tcp_store = init_distrib_slurm(
self.config.RL.DDPPO.distrib_backend
)
add_signal_handlers()
profiling_wrapper.configure(
capture_start_step=self.config.PROFILING.CAPTURE_START_STEP,
num_steps_to_capture=self.config.PROFILING.NUM_STEPS_TO_CAPTURE,
)
# Stores the number of workers that have finished their rollout
num_rollouts_done_store = distrib.PrefixStore(
"rollout_tracker", tcp_store
)
num_rollouts_done_store.set("num_done", "0")
self.world_rank = distrib.get_rank()
self.world_size = distrib.get_world_size()
self.config.defrost()
self.config.TORCH_GPU_ID = self.local_rank
self.config.SIMULATOR_GPU_ID = self.local_rank
# Multiply by the number of simulators to make sure they also get unique seeds
self.config.TASK_CONFIG.SEED += (
self.world_rank * self.config.NUM_PROCESSES
)
self.config.freeze()
random.seed(self.config.TASK_CONFIG.SEED)
np.random.seed(self.config.TASK_CONFIG.SEED)
torch.manual_seed(self.config.TASK_CONFIG.SEED)
if torch.cuda.is_available():
self.device = torch.device("cuda", self.local_rank)
torch.cuda.set_device(self.device)
else:
self.device = torch.device("cpu")
self.envs = construct_envs(
self.config,
get_env_class(self.config.ENV_NAME),
workers_ignore_signals=True,
)
ppo_cfg = self.config.RL.PPO
if (
not os.path.isdir(self.config.CHECKPOINT_FOLDER)
and self.world_rank == 0
):
os.makedirs(self.config.CHECKPOINT_FOLDER)
self._setup_actor_critic_agent(ppo_cfg)
self.agent.init_distributed(find_unused_params=True)
if self.world_rank == 0:
logger.info(
"agent number of trainable parameters: {}".format(
sum(
param.numel()
for param in self.agent.parameters()
if param.requires_grad
)
)
)
observations = self.envs.reset()
batch = batch_obs(observations, device=self.device)
batch = apply_obs_transforms_batch(batch, self.obs_transforms)
obs_space = self.obs_space
if self._static_encoder:
self._encoder = self.actor_critic.net.visual_encoder
obs_space = spaces.Dict(
{
"visual_features": spaces.Box(
low=np.finfo(np.float32).min,
high=np.finfo(np.float32).max,
shape=self._encoder.output_shape,
dtype=np.float32,
),
**obs_space.spaces,
}
)
with torch.no_grad():
batch["visual_features"] = self._encoder(batch)
rollouts = RolloutStorage(
ppo_cfg.num_steps,
self.envs.num_envs,
obs_space,
self.envs.action_spaces[0],
ppo_cfg.hidden_size,
num_recurrent_layers=self.actor_critic.net.num_recurrent_layers,
)
rollouts.to(self.device)
for sensor in rollouts.observations:
rollouts.observations[sensor][0].copy_(batch[sensor])
# batch and observations may contain shared PyTorch CUDA
# tensors. We must explicitly clear them here otherwise
# they will be kept in memory for the entire duration of training!
batch = None
observations = None
current_episode_reward = torch.zeros(
self.envs.num_envs, 1, device=self.device
)
running_episode_stats = dict(
count=torch.zeros(self.envs.num_envs, 1, device=self.device),
reward=torch.zeros(self.envs.num_envs, 1, device=self.device),
)
window_episode_stats: DefaultDict[str, deque] = defaultdict(
lambda: deque(maxlen=ppo_cfg.reward_window_size)
)
t_start = time.time()
env_time = 0
pth_time = 0
count_steps: float = 0
count_checkpoints = 0
start_update = 0
prev_time = 0
lr_scheduler = LambdaLR(
optimizer=self.agent.optimizer,
lr_lambda=lambda x: linear_decay(x, self.config.NUM_UPDATES), # type: ignore
)
interrupted_state = load_interrupted_state()
if interrupted_state is not None:
self.agent.load_state_dict(interrupted_state["state_dict"])
self.agent.optimizer.load_state_dict(
interrupted_state["optim_state"]
)
lr_scheduler.load_state_dict(interrupted_state["lr_sched_state"])
requeue_stats = interrupted_state["requeue_stats"]
env_time = requeue_stats["env_time"]
pth_time = requeue_stats["pth_time"]
count_steps = requeue_stats["count_steps"]
count_checkpoints = requeue_stats["count_checkpoints"]
start_update = requeue_stats["start_update"]
prev_time = requeue_stats["prev_time"]
with (
TensorboardWriter(
self.config.TENSORBOARD_DIR, flush_secs=self.flush_secs
)
if self.world_rank == 0
else contextlib.suppress()
) as writer:
for update in range(start_update, self.config.NUM_UPDATES):
profiling_wrapper.on_start_step()
profiling_wrapper.range_push("train update")
if ppo_cfg.use_linear_lr_decay:
lr_scheduler.step() # type: ignore
if ppo_cfg.use_linear_clip_decay:
self.agent.clip_param = ppo_cfg.clip_param * linear_decay(
update, self.config.NUM_UPDATES
)
if EXIT.is_set():
profiling_wrapper.range_pop() # train update
self.envs.close()
if REQUEUE.is_set() and self.world_rank == 0:
requeue_stats = dict(
env_time=env_time,
pth_time=pth_time,
count_steps=count_steps,
count_checkpoints=count_checkpoints,
start_update=update,
prev_time=(time.time() - t_start) + prev_time,
)
save_interrupted_state(
dict(
state_dict=self.agent.state_dict(),
optim_state=self.agent.optimizer.state_dict(),
lr_sched_state=lr_scheduler.state_dict(),
config=self.config,
requeue_stats=requeue_stats,
)
)
requeue_job()
return
count_steps_delta = 0
self.agent.eval()
profiling_wrapper.range_push("rollouts loop")
for step in range(ppo_cfg.num_steps):
(
delta_pth_time,
delta_env_time,
delta_steps,
) = self._collect_rollout_step(
rollouts, current_episode_reward, running_episode_stats
)
pth_time += delta_pth_time
env_time += delta_env_time
count_steps_delta += delta_steps
# This is where the preemption of workers happens. If a
# worker detects it will be a straggler, it preempts itself!
if (
step
>= ppo_cfg.num_steps * self.SHORT_ROLLOUT_THRESHOLD
) and int(num_rollouts_done_store.get("num_done")) > (
self.config.RL.DDPPO.sync_frac * self.world_size
):
break
profiling_wrapper.range_pop() # rollouts loop
num_rollouts_done_store.add("num_done", 1)
self.agent.train()
if self._static_encoder:
self._encoder.eval()
(
delta_pth_time,
value_loss,
action_loss,
dist_entropy,
) = self._update_agent(ppo_cfg, rollouts)
pth_time += delta_pth_time
stats_ordering = sorted(running_episode_stats.keys())
stats = torch.stack(
[running_episode_stats[k] for k in stats_ordering], 0
)
distrib.all_reduce(stats)
for i, k in enumerate(stats_ordering):
window_episode_stats[k].append(stats[i].clone())
stats = torch.tensor(
[value_loss, action_loss, count_steps_delta],
device=self.device,
)
distrib.all_reduce(stats)
count_steps += stats[2].item()
if self.world_rank == 0:
num_rollouts_done_store.set("num_done", "0")
losses = [
stats[0].item() / self.world_size,
stats[1].item() / self.world_size,
]
deltas = {
k: (
(v[-1] - v[0]).sum().item()
if len(v) > 1
else v[0].sum().item()
)
for k, v in window_episode_stats.items()
}
deltas["count"] = max(deltas["count"], 1.0)
writer.add_scalar(
"reward",
deltas["reward"] / deltas["count"],
count_steps,
)
# Check to see if there are any metrics
# that haven't been logged yet
metrics = {
k: v / deltas["count"]
for k, v in deltas.items()
if k not in {"reward", "count"}
}
if len(metrics) > 0:
writer.add_scalars("metrics", metrics, count_steps)
writer.add_scalars(
"losses",
{k: l for l, k in zip(losses, ["value", "policy"])},
count_steps,
)
# log stats
if update > 0 and update % self.config.LOG_INTERVAL == 0:
logger.info(
"update: {}\tfps: {:.3f}\t".format(
update,
count_steps
/ ((time.time() - t_start) + prev_time),
)
)
logger.info(
"update: {}\tenv-time: {:.3f}s\tpth-time: {:.3f}s\t"
"frames: {}".format(
update, env_time, pth_time, count_steps
)
)
logger.info(
"Average window size: {} {}".format(
len(window_episode_stats["count"]),
" ".join(
"{}: {:.3f}".format(k, v / deltas["count"])
for k, v in deltas.items()
if k != "count"
),
)
)
# checkpoint model
if update % self.config.CHECKPOINT_INTERVAL == 0:
self.save_checkpoint(
f"ckpt.{count_checkpoints}.pth",
dict(step=count_steps),
)
count_checkpoints += 1
profiling_wrapper.range_pop() # train update
self.envs.close()