move optimizer and scheduler outside policies

lerobot/common/policies/act/modeling_act.py

@@ -127,25 +127,6 @@ class ActionChunkingTransformerPolicy(nn.Module):
         self.action_head = nn.Linear(cfg.d_model, cfg.action_dim)
 
         self._reset_parameters()
-        self._create_optimizer()
-
-    def _create_optimizer(self):
-        optimizer_params_dicts = [
-            {
-                "params": [
-                    p for n, p in self.named_parameters() if not n.startswith("backbone") and p.requires_grad
-                ]
-            },
-            {
-                "params": [
-                    p for n, p in self.named_parameters() if n.startswith("backbone") and p.requires_grad
-                ],
-                "lr": self.cfg.lr_backbone,
-            },
-        ]
-        self.optimizer = torch.optim.AdamW(
-            optimizer_params_dicts, lr=self.cfg.lr, weight_decay=self.cfg.weight_decay
-        )
 
     def _reset_parameters(self):
         """Xavier-uniform initialization of the transformer parameters as in the original code."""

lerobot/common/policies/diffusion/modeling_diffusion.py

@@ -18,7 +18,6 @@ import einops
 import torch
 import torch.nn.functional as F  # noqa: N812
 import torchvision
-from diffusers.optimization import get_scheduler
 from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
 from robomimic.models.base_nets import SpatialSoftmax
 from torch import Tensor, nn
@@ -70,25 +69,7 @@ class DiffusionPolicy(nn.Module):
             self.ema_diffusion = copy.deepcopy(self.diffusion)
             self.ema = _EMA(cfg, model=self.ema_diffusion)
 
-        # TODO(alexander-soare): Move optimizer out of policy.
-        self.optimizer = torch.optim.Adam(
-            self.diffusion.parameters(), cfg.lr, cfg.adam_betas, cfg.adam_eps, cfg.adam_weight_decay
-        )
 
-        # TODO(alexander-soare): Move LR scheduler out of policy.
-        # TODO(rcadene): modify lr scheduler so that it doesn't depend on epochs but steps
-        self.global_step = 0
-
-        # configure lr scheduler
-        self.lr_scheduler = get_scheduler(
-            cfg.lr_scheduler,
-            optimizer=self.optimizer,
-            num_warmup_steps=cfg.lr_warmup_steps,
-            num_training_steps=lr_scheduler_num_training_steps,
-            # pytorch assumes stepping LRScheduler every epoch
-            # however huggingface diffusers steps it every batch
-            last_epoch=self.global_step - 1,
-        )
 
     def reset(self):
         """

lerobot/scripts/train.py

@@ -8,6 +8,7 @@ import hydra
 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
@@ -24,7 +25,7 @@ from lerobot.common.utils.utils import (
 from lerobot.scripts.eval import eval_policy
 
 
-def update_diffusion(self, policy, batch: dict[str, Tensor], **_) -> dict:
+def update_diffusion(policy, batch: dict[str, Tensor], optimizer, lr_scheduler) -> dict:
     """Run the model in train mode, compute the loss, and do an optimization step."""
     start_time = time.time()
     policy.diffusion.train()
@@ -40,9 +41,9 @@ def update_diffusion(self, policy, batch: dict[str, Tensor], **_) -> dict:
         error_if_nonfinite=False,
     )
 
-        policy.optimizer.step()
+    optimizer.step()
-        policy.optimizer.zero_grad()
+    optimizer.zero_grad()
-        policy.lr_scheduler.step()
+    lr_scheduler.step()
 
     if policy.ema is not None:
         policy.ema.step(policy.diffusion)
@@ -50,7 +51,7 @@ def update_diffusion(self, policy, batch: dict[str, Tensor], **_) -> dict:
     info = {
         "loss": loss.item(),
         "grad_norm": float(grad_norm),
-            "lr": policy.lr_scheduler.get_last_lr()[0],
+        "lr": lr_scheduler.get_last_lr()[0],
         "update_s": time.time() - start_time,
     }
 
@@ -58,7 +59,7 @@ def update_diffusion(self, policy, batch: dict[str, Tensor], **_) -> dict:
 
 
 
-def update_act(self, policy, batch: dict[str, Tensor], **_) -> dict:
+def update_act(policy, batch: dict[str, Tensor], optimizer) -> dict:
     start_time = time.time()
     policy.train()
     batch = policy.normalize_inputs(batch)
@@ -71,8 +72,8 @@ def update_act(self, policy, batch: dict[str, Tensor], **_) -> dict:
         policy.parameters(), policy.cfg.grad_clip_norm, error_if_nonfinite=False
     )
 
-    policy.optimizer.step()
+    optimizer.step()
-    policy.optimizer.zero_grad()
+    optimizer.zero_grad()
 
     train_info = {
         "loss": loss.item(),
@@ -83,6 +84,8 @@ def update_act(self, policy, batch: dict[str, Tensor], **_) -> dict:
 
     return train_info
 
+
+
 @hydra.main(version_base=None, config_name="default", config_path="../configs")
 def train_cli(cfg: dict):
     train(
@@ -295,6 +298,43 @@ def train(cfg: dict, out_dir=None, job_name=None):
     logging.info("make_policy")
     policy = make_policy(cfg, dataset_stats=offline_dataset.stats)
 
+    # Temporary hack to move optimizer out of policy
+    if isinstance(policy, ActPolicy):
+        optimizer_params_dicts = [
+            {
+                "params": [
+                    p for n, p in policy.named_parameters() if not n.startswith("backbone") and p.requires_grad
+                ]
+            },
+            {
+                "params": [
+                    p for n, p in policy.named_parameters() if n.startswith("backbone") and p.requires_grad
+                ],
+                "lr": policy.cfg.lr_backbone,
+            },
+        ]
+        optimizer = torch.optim.AdamW(
+            optimizer_params_dicts, lr=policy.cfg.lr, weight_decay=policy.cfg.weight_decay
+        )
+    elif isinstance(policy, DiffusionPolicy):
+        optimizer = torch.optim.Adam(
+            policy.diffusion.parameters(), cfg.lr, cfg.adam_betas, cfg.adam_eps, cfg.adam_weight_decay
+        )
+        # TODO(rcadene): modify lr scheduler so that it doesn't depend on epochs but steps
+        global_step = 0
+        # configure lr scheduler
+        lr_scheduler = get_scheduler(
+            cfg.lr_scheduler,
+            optimizer=optimizer,
+            num_warmup_steps=cfg.lr_warmup_steps,
+            num_training_steps=cfg.offline_steps,
+            # pytorch assumes stepping LRScheduler every epoch
+            # however huggingface diffusers steps it every batch
+            last_epoch=global_step - 1,
+        )
+
+
+
     num_learnable_params = sum(p.numel() for p in policy.parameters() if p.requires_grad)
     num_total_params = sum(p.numel() for p in policy.parameters())
 
@@ -355,11 +395,10 @@ def train(cfg: dict, out_dir=None, job_name=None):
             batch[key] = batch[key].to(cfg.device, non_blocking=True)
 
         # Temporary hack to move update outside of policy
-        if isinstance(policy, DiffusionPolicy):
+        if isinstance(policy, ActPolicy):
-            train_info = update_diffusion(policy, batch)
+            train_info = update_act(policy, batch, optimizer)
-        elif isinstance(policy, ActPolicy):
+        elif isinstance(policy, DiffusionPolicy):
-            train_info = update_act(policy, batch)
+            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: