[Feature] timeit.printevery

ghstack-source-id: 19165bb
Pull Request resolved: #2653

Author: Vincent Moens

sota-implementations/a2c/a2c_atari.py

@@ -182,7 +182,10 @@ def update(batch, max_grad_norm=cfg.optim.max_grad_norm):
     lr = cfg.optim.lr
 
     c_iter = iter(collector)
-    for i in range(len(collector)):
+    total_iter = len(collector)
+    for i in range(total_iter):
+        timeit.printevery(1000, total_iter, erase=True)
+
         with timeit("collecting"):
             data = next(c_iter)
 
@@ -261,10 +264,7 @@ def update(batch, max_grad_norm=cfg.optim.max_grad_norm):
                         "test/reward": test_rewards.mean(),
                     }
                 )
-        if i % 200 == 0:
-            log_info.update(timeit.todict(prefix="time"))
-            timeit.print()
-            timeit.erase()
+        log_info.update(timeit.todict(prefix="time"))
 
         if logger:
             for key, value in log_info.items():

sota-implementations/a2c/a2c_mujoco.py

@@ -179,7 +179,10 @@ def update(batch):
     pbar = tqdm.tqdm(total=cfg.collector.total_frames)
 
     c_iter = iter(collector)
-    for i in range(len(collector)):
+    total_iter = len(collector)
+    for i in range(total_iter):
+        timeit.printevery(1000, total_iter, erase=True)
+
         with timeit("collecting"):
             data = next(c_iter)
 
@@ -257,10 +260,7 @@ def update(batch):
                 )
                 actor.train()
 
-        if i % 200 == 0:
-            log_info.update(timeit.todict(prefix="time"))
-            timeit.print()
-            timeit.erase()
+        log_info.update(timeit.todict(prefix="time"))
 
         if logger:
             for key, value in log_info.items():

sota-implementations/cql/cql_offline.py

@@ -11,7 +11,6 @@
 """
 from __future__ import annotations
 
-import time
 import warnings
 
 import hydra
@@ -21,7 +20,7 @@
 import tqdm
 from tensordict.nn import CudaGraphModule
 
-from torchrl._utils import logger as torchrl_logger, timeit
+from torchrl._utils import timeit
 from torchrl.envs.utils import ExplorationType, set_exploration_type
 from torchrl.objectives import group_optimizers
 from torchrl.record.loggers import generate_exp_name, get_logger
@@ -156,9 +155,9 @@ def update(data, policy_eval_start, iteration):
     eval_steps = cfg.logger.eval_steps
 
     # Training loop
-    start_time = time.time()
     policy_eval_start = torch.tensor(policy_eval_start, device=device)
     for i in range(gradient_steps):
+        timeit.printevery(1000, gradient_steps, erase=True)
         pbar.update(1)
         # sample data
         with timeit("sample"):
@@ -192,15 +191,10 @@ def update(data, policy_eval_start, iteration):
                 to_log["evaluation_reward"] = eval_reward
 
         with timeit("log"):
-            if i % 200 == 0:
-                to_log.update(timeit.todict(prefix="time"))
+            to_log.update(timeit.todict(prefix="time"))
             log_metrics(logger, to_log, i)
-        if i % 200 == 0:
-            timeit.print()
-            timeit.erase()
 
     pbar.close()
-    torchrl_logger.info(f"Training time: {time.time() - start_time}")
     if not eval_env.is_closed:
         eval_env.close()
 

sota-implementations/cql/cql_online.py

@@ -170,7 +170,9 @@ def update(sampled_tensordict):
     eval_rollout_steps = cfg.logger.eval_steps
 
     c_iter = iter(collector)
-    for i in range(len(collector)):
+    total_iter = len(collector)
+    for i in range(total_iter):
+        timeit.printevery(1000, total_iter, erase=True)
         with timeit("collecting"):
             tensordict = next(c_iter)
         pbar.update(tensordict.numel())
@@ -222,8 +224,7 @@ def update(sampled_tensordict):
                 "loss_alpha_prime"
             ).mean()
             metrics_to_log["train/entropy"] = log_loss_td.get("entropy").mean()
-            if i % 10 == 0:
-                metrics_to_log.update(timeit.todict(prefix="time"))
+            metrics_to_log.update(timeit.todict(prefix="time"))
 
         # Evaluation
         with timeit("eval"):
@@ -245,9 +246,6 @@ def update(sampled_tensordict):
                     metrics_to_log["eval/reward"] = eval_reward
 
         log_metrics(logger, metrics_to_log, collected_frames)
-        if i % 10 == 0:
-            timeit.print()
-            timeit.erase()
 
     collector.shutdown()
     if not eval_env.is_closed:

sota-implementations/cql/discrete_cql_online.py

@@ -151,7 +151,9 @@ def update(sampled_tensordict):
     frames_per_batch = cfg.collector.frames_per_batch
 
     c_iter = iter(collector)
-    for i in range(len(collector)):
+    total_iter = len(collector)
+    for _ in range(total_iter):
+        timeit.printevery(1000, total_iter, erase=True)
         with timeit("collecting"):
             torch.compiler.cudagraph_mark_step_begin()
             tensordict = next(c_iter)
@@ -224,12 +226,7 @@ def update(sampled_tensordict):
             tds = torch.stack(tds, dim=0).mean()
             metrics_to_log["train/q_loss"] = tds["loss_qvalue"]
             metrics_to_log["train/cql_loss"] = tds["loss_cql"]
-            if i % 100 == 0:
-                metrics_to_log.update(timeit.todict(prefix="time"))
-
-        if i % 100 == 0:
-            timeit.print()
-            timeit.erase()
+            metrics_to_log.update(timeit.todict(prefix="time"))
 
         if logger is not None:
             log_metrics(logger, metrics_to_log, collected_frames)

sota-implementations/crossq/crossq.py

@@ -192,7 +192,9 @@ def update(sampled_tensordict: TensorDict, update_actor: bool):
     update_counter = 0
     delayed_updates = cfg.optim.policy_update_delay
     c_iter = iter(collector)
-    for i in range(len(collector)):
+    total_iter = len(collector)
+    for _ in range(total_iter):
+        timeit.printevery(1000, total_iter, erase=True)
         with timeit("collecting"):
             torch.compiler.cudagraph_mark_step_begin()
             tensordict = next(c_iter)
@@ -258,18 +260,14 @@ def update(sampled_tensordict: TensorDict, update_actor: bool):
             metrics_to_log["train/episode_length"] = episode_length.sum().item() / len(
                 episode_length
             )
-            if i % 20 == 0:
-                metrics_to_log.update(timeit.todict(prefix="time"))
+            metrics_to_log.update(timeit.todict(prefix="time"))
         if collected_frames >= init_random_frames:
             metrics_to_log["train/q_loss"] = tds["loss_qvalue"]
             metrics_to_log["train/actor_loss"] = tds["loss_actor"]
             metrics_to_log["train/alpha_loss"] = tds["loss_alpha"]
 
         if logger is not None:
             log_metrics(logger, metrics_to_log, collected_frames)
-        if i % 20 == 0:
-            timeit.print()
-            timeit.erase()
 
     collector.shutdown()
     if not eval_env.is_closed:

sota-implementations/ddpg/ddpg.py

@@ -156,7 +156,9 @@ def update(sampled_tensordict):
     eval_rollout_steps = cfg.env.max_episode_steps
 
     c_iter = iter(collector)
-    for i in range(len(collector)):
+    total_iter = len(collector)
+    for _ in range(total_iter):
+        timeit.printevery(1000, total_iter, erase=True)
         with timeit("collecting"):
             tensordict = next(c_iter)
         # Update exploration policy
@@ -226,10 +228,7 @@ def update(sampled_tensordict):
                 eval_env.apply(dump_video)
                 eval_reward = eval_rollout["next", "reward"].sum(-2).mean().item()
                 metrics_to_log["eval/reward"] = eval_reward
-        if i % 20 == 0:
-            metrics_to_log.update(timeit.todict(prefix="time"))
-            timeit.print()
-            timeit.erase()
+        metrics_to_log.update(timeit.todict(prefix="time"))
 
         if logger is not None:
             log_metrics(logger, metrics_to_log, collected_frames)

sota-implementations/decision_transformer/dt.py

@@ -128,6 +128,7 @@ def update(data: TensorDict) -> TensorDict:
     # Pretraining
     pbar = tqdm.tqdm(range(pretrain_gradient_steps))
     for i in pbar:
+        timeit.printevery(1000, pretrain_gradient_steps, erase=True)
         # Sample data
         with timeit("rb - sample"):
             data = offline_buffer.sample().to(model_device)
@@ -151,10 +152,7 @@ def update(data: TensorDict) -> TensorDict:
             to_log["eval/reward"] = (
                 eval_td["next", "reward"].sum(1).mean().item() / reward_scaling
             )
-        if i % 200 == 0:
-            to_log.update(timeit.todict(prefix="time"))
-            timeit.print()
-            timeit.erase()
+        to_log.update(timeit.todict(prefix="time"))
 
         if logger is not None:
             log_metrics(logger, to_log, i)

sota-implementations/decision_transformer/online_dt.py

@@ -8,7 +8,6 @@
 """
 from __future__ import annotations
 
-import time
 import warnings
 
 import hydra
@@ -130,8 +129,8 @@ def update(data):
 
     torchrl_logger.info(" ***Pretraining*** ")
     # Pretraining
-    start_time = time.time()
     for i in range(pretrain_gradient_steps):
+        timeit.printevery(1000, pretrain_gradient_steps, erase=True)
         pbar.update(1)
         with timeit("sample"):
             # Sample data
@@ -170,18 +169,14 @@ def update(data):
                 eval_td["next", "reward"].sum(1).mean().item() / reward_scaling
             )
 
-        if i % 200 == 0:
-            to_log.update(timeit.todict(prefix="time"))
-            timeit.print()
-            timeit.erase()
+        to_log.update(timeit.todict(prefix="time"))
 
         if logger is not None:
             log_metrics(logger, to_log, i)
 
     pbar.close()
     if not test_env.is_closed:
         test_env.close()
-    torchrl_logger.info(f"Training time: {time.time() - start_time}")
 
 
 if __name__ == "__main__":

sota-implementations/discrete_sac/discrete_sac.py

@@ -155,7 +155,9 @@ def update(sampled_tensordict):
     frames_per_batch = cfg.collector.frames_per_batch
 
     c_iter = iter(collector)
-    for i in range(len(collector)):
+    total_iter = len(collector)
+    for i in range(total_iter):
+        timeit.printevery(1000, total_iter, erase=True)
         with timeit("collecting"):
             collected_data = next(c_iter)
 
@@ -229,10 +231,7 @@ def update(sampled_tensordict):
                 eval_env.apply(dump_video)
                 eval_reward = eval_rollout["next", "reward"].sum(-2).mean().item()
                 metrics_to_log["eval/reward"] = eval_reward
-        if i % 50 == 0:
-            metrics_to_log.update(timeit.todict(prefix="time"))
-            timeit.print()
-            timeit.erase()
+        metrics_to_log.update(timeit.todict(prefix="time"))
         if logger is not None:
             log_metrics(logger, metrics_to_log, collected_frames)