ready for review

examples/2_evaluate_pretrained_policy.py

@@ -11,6 +11,7 @@ from lerobot.common.utils import init_hydra_config
 from lerobot.scripts.eval import eval
 
 # Get a pretrained policy from the hub.
+# TODO(alexander-soare): This no longer works until we upload a new model that uses the current configs.
 hub_id = "lerobot/diffusion_policy_pusht_image"
 folder = Path(snapshot_download(hub_id))
 # OR uncomment the following to evaluate a policy from the local outputs/train folder.

examples/3_train_policy.py

@@ -11,7 +11,6 @@ import torch
 from omegaconf import OmegaConf
 
 from lerobot.common.datasets.factory import make_dataset
-from lerobot.common.datasets.utils import cycle
 from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
 from lerobot.common.policies.diffusion.modeling_diffusion import DiffusionPolicy
 from lerobot.common.utils import init_hydra_config
@@ -26,8 +25,8 @@ device = torch.device("cuda")
 log_freq = 250
 
 # Set up the dataset.
-cfg = init_hydra_config("lerobot/configs/default.yaml", overrides=["env=pusht"])
-dataset = make_dataset(cfg)
+hydra_cfg = init_hydra_config("lerobot/configs/default.yaml", overrides=["env=pusht"])
+dataset = make_dataset(hydra_cfg)
 
 # Set up the the policy.
 # Policies are initialized with a configuration class, in this case `DiffusionConfig`.
@@ -50,17 +49,25 @@ dataloader = torch.utils.data.DataLoader(
 )
 
 # Run training loop.
-dataloader = cycle(dataloader)
-for step in range(training_steps):
-    batch = {k: v.to(device, non_blocking=True) for k, v in next(dataloader).items()}
-    info = policy(batch)
-    if step % log_freq == 0:
-        num_samples = (step + 1) * cfg.batch_size
-        loss = info["loss"]
-        update_s = info["update_s"]
-        print(f"step: {step} samples: {num_samples} loss: {loss:.3f} update_time: {update_s:.3f} (seconds)")
+step = 0
+done = False
+while not done:
+    for batch in dataloader:
+        batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()}
+        info = policy(batch)
+        if step % log_freq == 0:
+            num_samples = (step + 1) * cfg.batch_size
+            loss = info["loss"]
+            update_s = info["update_s"]
+            print(
+                f"step: {step} samples: {num_samples} loss: {loss:.3f} update_time: {update_s:.3f} (seconds)"
+            )
+        step += 1
+        if step >= training_steps:
+            done = True
+            break
 
 # Save the policy, configuration, and normalization stats for later use.
 policy.save(output_directory / "model.pt")
-OmegaConf.save(cfg, output_directory / "config.yaml")
+OmegaConf.save(hydra_cfg, output_directory / "config.yaml")
 torch.save(dataset.transform.transforms[-1].stats, output_directory / "stats.pth")

lerobot/common/policies/diffusion/modeling_diffusion.py

@@ -1,4 +1,5 @@
-"""
+"""Diffusion Policy as per "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion"
+
 TODO(alexander-soare):
   - Remove reliance on Robomimic for SpatialSoftmax.
   - Remove reliance on diffusers for DDPMScheduler and LR scheduler.

tests/test_examples.py

@@ -29,11 +29,12 @@ def test_examples_3_and_2():
     with open(path, "r") as file:
         file_contents = file.read()
 
-    # Do less steps and use CPU.
+    # Do less steps, use CPU, and don't complicate things with dataloader workers.
     file_contents = _find_and_replace(
         file_contents,
         [
-            ("offline_steps = 5000", "offline_steps = 1"),
+            ("training_steps = 5000", "training_steps = 1"),
+            ("num_workers=4", "num_workers=0"),
             ('device = torch.device("cuda")', 'device = torch.device("cpu")'),
         ],
     )