Merge remote-tracking branch 'origin/main' into user/aliberts/2025_02_10_dataset_v2.1

lerobot/common/datasets/factory.py

@@ -104,7 +104,7 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
         )
         logging.info(
             "Multiple datasets were provided. Applied the following index mapping to the provided datasets: "
-            f"{pformat(dataset.repo_id_to_index , indent=2)}"
+            f"{pformat(dataset.repo_id_to_index, indent=2)}"
         )
 
     if cfg.dataset.use_imagenet_stats:

lerobot/common/datasets/push_dataset_to_hub/dora_parquet_format.py

@@ -72,7 +72,7 @@ def load_from_raw(raw_dir: Path, videos_dir: Path, fps: int, video: bool, episod
             # However, note that "nearest" might synchronize the reference camera with other cameras on slightly future timestamps.
             # are too far appart.
             direction="nearest",
-            tolerance=pd.Timedelta(f"{1/fps} seconds"),
+            tolerance=pd.Timedelta(f"{1 / fps} seconds"),
         )
     # Remove rows with episode_index -1 which indicates data that correspond to in-between episodes
     df = df[df["episode_index"] != -1]

lerobot/common/policies/act/modeling_act.py

@@ -409,9 +409,9 @@ class ACT(nn.Module):
             latent dimension.
         """
         if self.config.use_vae and self.training:
-            assert (
-                "action" in batch
-            ), "actions must be provided when using the variational objective in training mode."
+            assert "action" in batch, (
+                "actions must be provided when using the variational objective in training mode."
+            )
 
         batch_size = (
             batch["observation.images"]

lerobot/common/policies/diffusion/configuration_diffusion.py

@@ -221,7 +221,7 @@ class DiffusionConfig(PreTrainedConfig):
         for key, image_ft in self.image_features.items():
             if image_ft.shape != first_image_ft.shape:
                 raise ValueError(
-                    f"`{key}` does not match `{first_image_key}`, but we " "expect all image shapes to match."
+                    f"`{key}` does not match `{first_image_key}`, but we expect all image shapes to match."
                 )
 
     @property

lerobot/common/policies/pi0/modeling_pi0.py

@@ -300,7 +300,7 @@ class PI0Policy(PreTrainedPolicy):
             self._action_queue.extend(actions.transpose(0, 1))
         return self._action_queue.popleft()
 
-    def forward(self, batch: dict[str, Tensor], noise=None, time=None) -> dict[str, Tensor]:
+    def forward(self, batch: dict[str, Tensor], noise=None, time=None) -> tuple[Tensor, dict[str, Tensor]]:
         """Do a full training forward pass to compute the loss"""
         if self.config.adapt_to_pi_aloha:
             batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
@@ -328,12 +328,12 @@ class PI0Policy(PreTrainedPolicy):
         losses = losses[:, :, : self.config.max_action_dim]
         loss_dict["losses_after_rm_padding"] = losses.clone()
 
-        loss = losses.mean()
         # For backward pass
-        loss_dict["loss"] = loss
+        loss = losses.mean()
         # For logging
         loss_dict["l2_loss"] = loss.item()
-        return loss_dict
+
+        return loss, loss_dict
 
     def prepare_images(self, batch):
         """Apply Pi0 preprocessing to the images, like resizing to 224x224 and padding to keep aspect ratio, and

lerobot/common/policies/tdmpc/modeling_tdmpc.py

@@ -594,9 +594,9 @@ class TDMPCTOLD(nn.Module):
 
         self.apply(_apply_fn)
         for m in [self._reward, *self._Qs]:
-            assert isinstance(
-                m[-1], nn.Linear
-            ), "Sanity check. The last linear layer needs 0 initialization on weights."
+            assert isinstance(m[-1], nn.Linear), (
+                "Sanity check. The last linear layer needs 0 initialization on weights."
+            )
             nn.init.zeros_(m[-1].weight)
             nn.init.zeros_(m[-1].bias)  # this has already been done, but keep this line here for good measure
 

lerobot/common/policies/vqbet/configuration_vqbet.py

@@ -184,7 +184,7 @@ class VQBeTConfig(PreTrainedConfig):
         for key, image_ft in self.image_features.items():
             if image_ft.shape != first_image_ft.shape:
                 raise ValueError(
-                    f"`{key}` does not match `{first_image_key}`, but we " "expect all image shapes to match."
+                    f"`{key}` does not match `{first_image_key}`, but we expect all image shapes to match."
                 )
 
     @property

lerobot/common/policies/vqbet/vqbet_utils.py

@@ -203,9 +203,9 @@ class GPT(nn.Module):
     def forward(self, input, targets=None):
         device = input.device
         b, t, d = input.size()
-        assert (
-            t <= self.config.gpt_block_size
-        ), f"Cannot forward sequence of length {t}, block size is only {self.config.gpt_block_size}"
+        assert t <= self.config.gpt_block_size, (
+            f"Cannot forward sequence of length {t}, block size is only {self.config.gpt_block_size}"
+        )
 
         # positional encodings that are added to the input embeddings
         pos = torch.arange(0, t, dtype=torch.long, device=device).unsqueeze(0)  # shape (1, t)
@@ -273,10 +273,10 @@ class GPT(nn.Module):
         assert len(inter_params) == 0, "parameters {} made it into both decay/no_decay sets!".format(
             str(inter_params)
         )
-        assert (
-            len(param_dict.keys() - union_params) == 0
-        ), "parameters {} were not separated into either decay/no_decay set!".format(
-            str(param_dict.keys() - union_params),
+        assert len(param_dict.keys() - union_params) == 0, (
+            "parameters {} were not separated into either decay/no_decay set!".format(
+                str(param_dict.keys() - union_params),
+            )
         )
 
         decay = [param_dict[pn] for pn in sorted(decay)]
@@ -419,9 +419,9 @@ class ResidualVQ(nn.Module):
         # and the network should be able to reconstruct
 
         if quantize_dim < self.num_quantizers:
-            assert (
-                self.quantize_dropout > 0.0
-            ), "quantize dropout must be greater than 0 if you wish to reconstruct from a signal with less fine quantizations"
+            assert self.quantize_dropout > 0.0, (
+                "quantize dropout must be greater than 0 if you wish to reconstruct from a signal with less fine quantizations"
+            )
             indices = F.pad(indices, (0, self.num_quantizers - quantize_dim), value=-1)
 
         # get ready for gathering
@@ -472,9 +472,9 @@ class ResidualVQ(nn.Module):
         all_indices = []
 
         if return_loss:
-            assert not torch.any(
-                indices == -1
-            ), "some of the residual vq indices were dropped out. please use indices derived when the module is in eval mode to derive cross entropy loss"
+            assert not torch.any(indices == -1), (
+                "some of the residual vq indices were dropped out. please use indices derived when the module is in eval mode to derive cross entropy loss"
+            )
             ce_losses = []
 
         should_quantize_dropout = self.training and self.quantize_dropout and not return_loss
@@ -887,9 +887,9 @@ class VectorQuantize(nn.Module):
                 # only calculate orthogonal loss for the activated codes for this batch
 
                 if self.orthogonal_reg_active_codes_only:
-                    assert not (
-                        is_multiheaded and self.separate_codebook_per_head
-                    ), "orthogonal regularization for only active codes not compatible with multi-headed with separate codebooks yet"
+                    assert not (is_multiheaded and self.separate_codebook_per_head), (
+                        "orthogonal regularization for only active codes not compatible with multi-headed with separate codebooks yet"
+                    )
                     unique_code_ids = torch.unique(embed_ind)
                     codebook = codebook[:, unique_code_ids]
 
@@ -999,9 +999,9 @@ def gumbel_sample(
     ind = sampling_logits.argmax(dim=dim)
     one_hot = F.one_hot(ind, size).type(dtype)
 
-    assert not (
-        reinmax and not straight_through
-    ), "reinmax can only be turned on if using straight through gumbel softmax"
+    assert not (reinmax and not straight_through), (
+        "reinmax can only be turned on if using straight through gumbel softmax"
+    )
 
     if not straight_through or temperature <= 0.0 or not training:
         return ind, one_hot
@@ -1209,9 +1209,9 @@ class EuclideanCodebook(nn.Module):
         self.gumbel_sample = gumbel_sample
         self.sample_codebook_temp = sample_codebook_temp
 
-        assert not (
-            use_ddp and num_codebooks > 1 and kmeans_init
-        ), "kmeans init is not compatible with multiple codebooks in distributed environment for now"
+        assert not (use_ddp and num_codebooks > 1 and kmeans_init), (
+            "kmeans init is not compatible with multiple codebooks in distributed environment for now"
+        )
 
         self.sample_fn = sample_vectors_distributed if use_ddp and sync_kmeans else batched_sample_vectors
         self.kmeans_all_reduce_fn = distributed.all_reduce if use_ddp and sync_kmeans else noop

lerobot/common/robot_devices/control_utils.py

@@ -33,7 +33,7 @@ def log_control_info(robot: Robot, dt_s, episode_index=None, frame_index=None, f
 
     def log_dt(shortname, dt_val_s):
         nonlocal log_items, fps
-        info_str = f"{shortname}:{dt_val_s * 1000:5.2f} ({1/ dt_val_s:3.1f}hz)"
+        info_str = f"{shortname}:{dt_val_s * 1000:5.2f} ({1 / dt_val_s:3.1f}hz)"
         if fps is not None:
             actual_fps = 1 / dt_val_s
             if actual_fps < fps - 1:

lerobot/common/utils/io_utils.py

@@ -58,7 +58,7 @@ def deserialize_json_into_object(fpath: Path, obj: T) -> T:
             # Check that they have exactly the same set of keys.
             if target.keys() != source.keys():
                 raise ValueError(
-                    f"Dictionary keys do not match.\n" f"Expected: {target.keys()}, got: {source.keys()}"
+                    f"Dictionary keys do not match.\nExpected: {target.keys()}, got: {source.keys()}"
                 )
 
             # Recursively update each key.

lerobot/common/utils/wandb_utils.py

@@ -102,7 +102,7 @@ class WandBLogger:
         self._wandb.log_artifact(artifact)
 
     def log_dict(self, d: dict, step: int, mode: str = "train"):
-        if mode in {"train", "eval"}:
+        if mode not in {"train", "eval"}:
             raise ValueError(mode)
 
         for k, v in d.items():
@@ -114,7 +114,7 @@ class WandBLogger:
             self._wandb.log({f"{mode}/{k}": v}, step=step)
 
     def log_video(self, video_path: str, step: int, mode: str = "train"):
-        if mode in {"train", "eval"}:
+        if mode not in {"train", "eval"}:
             raise ValueError(mode)
 
         wandb_video = self._wandb.Video(video_path, fps=self.env_fps, format="mp4")

lerobot/configs/train.py

@@ -85,6 +85,11 @@ class TrainPipelineConfig(HubMixin):
             config_path = parser.parse_arg("config_path")
             if not config_path:
                 raise ValueError("A config_path is expected when resuming a run.")
+            if not Path(config_path).resolve().exists():
+                raise NotADirectoryError(
+                    f"{config_path=} is expected to be a local path. "
+                    "Resuming from the hub is not supported for now."
+                )
             policy_path = Path(config_path).parent
             self.policy.pretrained_path = policy_path
             self.checkpoint_path = policy_path.parent

lerobot/scripts/eval.py

@@ -151,7 +151,9 @@ def rollout(
         if return_observations:
             all_observations.append(deepcopy(observation))
 
-        observation = {key: observation[key].to(device, non_blocking=True) for key in observation}
+        observation = {
+            key: observation[key].to(device, non_blocking=device.type == "cuda") for key in observation
+        }
 
         with torch.inference_mode():
             action = policy.select_action(observation)

lerobot/scripts/train.py

@@ -232,8 +232,10 @@ def train(cfg: TrainPipelineConfig):
         if is_log_step:
             logging.info(train_tracker)
             if wandb_logger:
-                wandb_log_dict = {**train_tracker.to_dict(), **output_dict}
-                wandb_logger.log_dict(wandb_log_dict)
+                wandb_log_dict = train_tracker.to_dict()
+                if output_dict:
+                    wandb_log_dict.update(output_dict)
+                wandb_logger.log_dict(wandb_log_dict, step)
             train_tracker.reset_averages()
 
         if cfg.save_checkpoint and is_saving_step:
@@ -271,6 +273,7 @@ def train(cfg: TrainPipelineConfig):
             logging.info(eval_tracker)
             if wandb_logger:
                 wandb_log_dict = {**eval_tracker.to_dict(), **eval_info}
+                wandb_logger.log_dict(wandb_log_dict, step, mode="eval")
                 wandb_logger.log_video(eval_info["video_paths"][0], step, mode="eval")
 
     if eval_env:

lerobot/scripts/visualize_dataset.py

@@ -111,9 +111,9 @@ def visualize_dataset(
     output_dir: Path | None = None,
 ) -> Path | None:
     if save:
-        assert (
-            output_dir is not None
-        ), "Set an output directory where to write .rrd files with `--output-dir path/to/directory`."
+        assert output_dir is not None, (
+            "Set an output directory where to write .rrd files with `--output-dir path/to/directory`."
+        )
 
     repo_id = dataset.repo_id