[Feature] TD3 compatibility with compile

ghstack-source-id: 6b2f00c347c34bf85b36bab05e66b8e9f4bf7d62
Pull Request resolved: #2656

sota-implementations/td3/config.yaml

@@ -21,7 +21,7 @@ collector:
 replay_buffer:
   prb: 0 # use prioritized experience replay
   size: 1000000
-  scratch_dir: null
+  scratch_dir:
 
 # optim
 optim:
@@ -52,3 +52,8 @@ logger:
   mode: online
   eval_iter: 25000
   video: False
+
+compile:
+  compile: False
+  compile_mode:
+  cudagraphs: False

sota-implementations/td3/td3.py

@@ -12,14 +12,16 @@
 """
 from __future__ import annotations
 
-import time
+import warnings
 
 import hydra
 import numpy as np
 import torch
 import torch.cuda
 import tqdm
-from torchrl._utils import logger as torchrl_logger
+from tensordict.nn import CudaGraphModule
+
+from torchrl._utils import compile_with_warmup, timeit
 
 from torchrl.envs.utils import ExplorationType, set_exploration_type
 
@@ -36,6 +38,9 @@
 )
 
 
+torch.set_float32_matmul_precision("high")
+
+
 @hydra.main(version_base="1.1", config_path="", config_name="config")
 def main(cfg: "DictConfig"):  # noqa: F821
     device = cfg.network.device
@@ -75,23 +80,67 @@ def main(cfg: "DictConfig"):  # noqa: F821
     # Create TD3 loss
     loss_module, target_net_updater = make_loss_module(cfg, model)
 
+    compile_mode = None
+    if cfg.compile.compile:
+        compile_mode = cfg.compile.compile_mode
+        if compile_mode in ("", None):
+            if cfg.compile.cudagraphs:
+                compile_mode = "default"
+            else:
+                compile_mode = "reduce-overhead"
+
     # Create off-policy collector
-    collector = make_collector(cfg, train_env, exploration_policy)
+    collector = make_collector(
+        cfg, train_env, exploration_policy, compile_mode=compile_mode
+    )
 
     # Create replay buffer
     replay_buffer = make_replay_buffer(
         batch_size=cfg.optim.batch_size,
         prb=cfg.replay_buffer.prb,
         buffer_size=cfg.replay_buffer.size,
         scratch_dir=cfg.replay_buffer.scratch_dir,
-        device="cpu",
+        device=device,
     )
 
     # Create optimizers
     optimizer_actor, optimizer_critic = make_optimizer(cfg, loss_module)
 
+    def update(sampled_tensordict, update_actor):
+        # Compute loss
+        q_loss, *_ = loss_module.value_loss(sampled_tensordict)
+
+        # Update critic
+        q_loss.backward()
+        optimizer_critic.step()
+        optimizer_critic.zero_grad(set_to_none=True)
+
+        # Update actor
+        if update_actor:
+            actor_loss, *_ = loss_module.actor_loss(sampled_tensordict)
+
+            actor_loss.backward()
+            optimizer_actor.step()
+            optimizer_actor.zero_grad(set_to_none=True)
+
+            # Update target params
+            target_net_updater.step()
+        else:
+            actor_loss = q_loss.new_zeros(())
+
+        return q_loss.detach(), actor_loss.detach()
+
+    if cfg.compile.compile:
+        update = compile_with_warmup(update, mode=compile_mode, warmup=1)
+
+    if cfg.compile.cudagraphs:
+        warnings.warn(
+            "CudaGraphModule is experimental and may lead to silently wrong results. Use with caution.",
+            category=UserWarning,
+        )
+        update = CudaGraphModule(update, in_keys=[], out_keys=[], warmup=5)
+
     # Main loop
-    start_time = time.time()
     collected_frames = 0
     pbar = tqdm.tqdm(total=cfg.collector.total_frames)
 
@@ -108,70 +157,56 @@ def main(cfg: "DictConfig"):  # noqa: F821
     frames_per_batch = cfg.collector.frames_per_batch
     update_counter = 0
 
-    sampling_start = time.time()
-    for tensordict in collector:
-        sampling_time = time.time() - sampling_start
-        exploration_policy[1].step(tensordict.numel())
+    collector_iter = iter(collector)
+    total_iter = len(collector)
+
+    for _ in range(total_iter):
+        timeit.printevery(num_prints=1000, total_count=total_iter, erase=True)
+
+        with timeit("collect"):
+            tensordict = next(collector_iter)
 
         # Update weights of the inference policy
         collector.update_policy_weights_()
 
-        pbar.update(tensordict.numel())
-
-        tensordict = tensordict.reshape(-1)
         current_frames = tensordict.numel()
-        # Add to replay buffer
-        replay_buffer.extend(tensordict.cpu())
+        pbar.update(current_frames)
+
+        with timeit("rb - extend"):
+            # Add to replay buffer
+            tensordict = tensordict.reshape(-1)
+            replay_buffer.extend(tensordict)
+
         collected_frames += current_frames
 
-        # Optimization steps
-        training_start = time.time()
-        if collected_frames >= init_random_frames:
-            (
-                actor_losses,
-                q_losses,
-            ) = ([], [])
-            for _ in range(num_updates):
-
-                # Update actor every delayed_updates
-                update_counter += 1
-                update_actor = update_counter % delayed_updates == 0
-
-                # Sample from replay buffer
-                sampled_tensordict = replay_buffer.sample()
-                if sampled_tensordict.device != device:
-                    sampled_tensordict = sampled_tensordict.to(
-                        device, non_blocking=True
-                    )
-                else:
-                    sampled_tensordict = sampled_tensordict.clone()
-
-                # Compute loss
-                q_loss, *_ = loss_module.value_loss(sampled_tensordict)
-
-                # Update critic
-                optimizer_critic.zero_grad()
-                q_loss.backward()
-                optimizer_critic.step()
-                q_losses.append(q_loss.item())
-
-                # Update actor
-                if update_actor:
-                    actor_loss, *_ = loss_module.actor_loss(sampled_tensordict)
-                    optimizer_actor.zero_grad()
-                    actor_loss.backward()
-                    optimizer_actor.step()
-
-                    actor_losses.append(actor_loss.item())
-
-                    # Update target params
-                    target_net_updater.step()
-
-                # Update priority
-                if prb:
-                    replay_buffer.update_priority(sampled_tensordict)
-
-        training_time = time.time() - training_start
+        with timeit("train"):
+            # Optimization steps
+            if collected_frames >= init_random_frames:
+                (
+                    actor_losses,
+                    q_losses,
+                ) = ([], [])
+                for _ in range(num_updates):
+                    # Update actor every delayed_updates
+                    update_counter += 1
+                    update_actor = update_counter % delayed_updates == 0
+
+                    with timeit("rb - sample"):
+                        # Sample from replay buffer
+                        sampled_tensordict = replay_buffer.sample()
+
+                    with timeit("update"):
+                        torch.compiler.cudagraph_mark_step_begin()
+                        q_loss, actor_loss = update(sampled_tensordict, update_actor)
+
+                    q_losses.append(q_loss.clone())
+                    if update_actor:
+                        actor_losses.append(actor_loss.clone())
+
+                    # Update priority
+                    if prb:
+                        replay_buffer.update_priority(sampled_tensordict)
+
         episode_end = (
             tensordict["next", "done"]
             if tensordict["next", "done"].any()
@@ -183,45 +218,39 @@ def main(cfg: "DictConfig"):  # noqa: F821
         metrics_to_log = {}
         if len(episode_rewards) > 0:
             episode_length = tensordict["next", "step_count"][episode_end]
-            metrics_to_log["train/reward"] = episode_rewards.mean().item()
-            metrics_to_log["train/episode_length"] = episode_length.sum().item() / len(
+            metrics_to_log["train/reward"] = episode_rewards.mean()
+            metrics_to_log["train/episode_length"] = episode_length.sum() / len(
                 episode_length
             )
 
         if collected_frames >= init_random_frames:
-            metrics_to_log["train/q_loss"] = np.mean(q_losses)
+            metrics_to_log["train/q_loss"] = torch.stack(q_losses).mean()
             if update_actor:
-                metrics_to_log["train/a_loss"] = np.mean(actor_losses)
-            metrics_to_log["train/sampling_time"] = sampling_time
-            metrics_to_log["train/training_time"] = training_time
+                metrics_to_log["train/a_loss"] = torch.stack(actor_losses).mean()
 
         # Evaluation
         if abs(collected_frames % eval_iter) < frames_per_batch:
-            with set_exploration_type(ExplorationType.DETERMINISTIC), torch.no_grad():
-                eval_start = time.time()
+            with set_exploration_type(
+                ExplorationType.DETERMINISTIC
+            ), torch.no_grad(), timeit("eval"):
                 eval_rollout = eval_env.rollout(
                     eval_rollout_steps,
                     exploration_policy,
                     auto_cast_to_device=True,
                     break_when_any_done=True,
                 )
                 eval_env.apply(dump_video)
-                eval_time = time.time() - eval_start
                 eval_reward = eval_rollout["next", "reward"].sum(-2).mean().item()
                 metrics_to_log["eval/reward"] = eval_reward
-                metrics_to_log["eval/time"] = eval_time
         if logger is not None:
+            metrics_to_log.update(timeit.todict(prefix="time"))
             log_metrics(logger, metrics_to_log, collected_frames)
-        sampling_start = time.time()
 
     collector.shutdown()
     if not eval_env.is_closed:
         eval_env.close()
     if not train_env.is_closed:
         train_env.close()
-    end_time = time.time()
-    execution_time = end_time - start_time
-    torchrl_logger.info(f"Training took {execution_time:.2f} seconds to finish")
 
 
 if __name__ == "__main__":