merge update functions

lerobot/scripts/train.py

@@ -9,7 +9,6 @@ import torch
 from datasets import concatenate_datasets
 from datasets.utils import disable_progress_bars, enable_progress_bars
 from diffusers.optimization import get_scheduler
-from torch import Tensor
 
 from lerobot.common.datasets.factory import make_dataset
 from lerobot.common.datasets.utils import cycle
@@ -25,67 +24,47 @@ from lerobot.common.utils.utils import (
 from lerobot.scripts.eval import eval_policy
 
 
-def update_diffusion(policy, batch: dict[str, Tensor], optimizer, lr_scheduler) -> dict:
+def update_policy(policy, batch, optimizer, lr_scheduler=None):
-    """Run the model in train mode, compute the loss, and do an optimization step."""
     start_time = time.time()
+
+    # Diffusion
     policy.diffusion.train()
+    # Act
+    policy.train()
+
     batch = policy.normalize_inputs(batch)
-    loss = policy.forward(batch)["loss"]
+
+    output_dict = policy.forward(batch)
+    # TODO(rcadene): policy.unnormalize_outputs(out_dict)
+    loss = output_dict["loss"]
     loss.backward()
 
-    # TODO(rcadene): self.unnormalize_outputs(out_dict)
+    # Diffusion
-
+    model = policy.diffusion if hasattr(policy, "diffusion") else policy # TODO: hacky, remove this line
     grad_norm = torch.nn.utils.clip_grad_norm_(
-        policy.diffusion.parameters(),
+        model.parameters(),
         policy.cfg.grad_clip_norm,
         error_if_nonfinite=False,
     )
 
     optimizer.step()
     optimizer.zero_grad()
+    if lr_scheduler is not None:
         lr_scheduler.step()
 
-    if policy.ema is not None:
+    if hasattr(policy, "ema") and policy.ema is not None:
-        policy.ema.step(policy.diffusion)
+        policy.ema.step(model)
     
     info = {
         "loss": loss.item(),
         "grad_norm": float(grad_norm),
-        "lr": lr_scheduler.get_last_lr()[0],
+        "lr": policy.cfg.lr if lr_scheduler is None else lr_scheduler.get_last_lr()[0],
         "update_s": time.time() - start_time,
     }
 
     return info
 
 
-
-def update_act(policy, batch: dict[str, Tensor], optimizer) -> dict:
-    start_time = time.time()
-    policy.train()
-    batch = policy.normalize_inputs(batch)
-    loss_dict = policy.forward(batch)
-    # TODO(rcadene): policy.unnormalize_outputs(out_dict)
-    loss = loss_dict["loss"]
-    loss.backward()
-
-    grad_norm = torch.nn.utils.clip_grad_norm_(
-        policy.parameters(), policy.cfg.grad_clip_norm, error_if_nonfinite=False
-    )
-
-    optimizer.step()
-    optimizer.zero_grad()
-
-    train_info = {
-        "loss": loss.item(),
-        "grad_norm": float(grad_norm),
-        "lr": policy.cfg.lr,
-        "update_s": time.time() - start_time,
-    }
-
-    return train_info
-
-
-
 @hydra.main(version_base=None, config_name="default", config_path="../configs")
 def train_cli(cfg: dict):
     train(
@@ -316,6 +295,7 @@ def train(cfg: dict, out_dir=None, job_name=None):
         optimizer = torch.optim.AdamW(
             optimizer_params_dicts, lr=policy.cfg.lr, weight_decay=policy.cfg.weight_decay
         )
+        lr_scheduler = None
     elif isinstance(policy, DiffusionPolicy):
         optimizer = torch.optim.Adam(
             policy.diffusion.parameters(), cfg.lr, cfg.adam_betas, cfg.adam_eps, cfg.adam_weight_decay
@@ -394,11 +374,7 @@ def train(cfg: dict, out_dir=None, job_name=None):
         for key in batch:
             batch[key] = batch[key].to(cfg.device, non_blocking=True)
 
-        # Temporary hack to move update outside of policy
+        train_info = update_policy(policy, batch, optimizer, lr_scheduler)
-        if isinstance(policy, ActPolicy):
-            train_info = update_act(policy, batch, optimizer)
-        elif isinstance(policy, DiffusionPolicy):
-            train_info = update_diffusion(policy, batch, optimizer, lr_scheduler)
 
         # TODO(rcadene): is it ok if step_t=0 = 0 and not 1 as previously done?
         if step % cfg.log_freq == 0: