optimzize data loading

src/lerobot/policies/rlearn/processor_rlearn.py

@@ -53,11 +53,7 @@ def make_rlearn_processor(
     input_steps = [
         # No renaming by default, but keep for future extensibility
         RenameProcessor(rename_map={}),
-        NormalizerProcessor(
+        # Move heavy normalization to GPU after transfer for better parallelism
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
         ToBatchProcessor(),
         RLearnLanguageFromTaskProcessor(),
         # Use SigLIP2 for tokenizer to keep vocab aligned with text tower
@@ -69,6 +65,12 @@ def make_rlearn_processor(
             padding_side="right",
         ),
         DeviceProcessor(device=config.device),
+        # Move normalization after GPU transfer to use GPU acceleration
+        NormalizerProcessor(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+        ),
     ]
 
     output_steps = [

src/lerobot/policies/rlearn/rlearn_plan.md

@@ -79,9 +79,10 @@ _ Open X-Embodiment (OXE)
   - Exactly similar to: https://github.com/lucidrains/rewind-reward-pytorch/blob/main/rewind_reward_pytorch/rewind_reward.py#L11 [x]
   - Try DINO v2 as encoder Base 86 M: with https://huggingface.co/sentence-transformers/all-MiniLM-L12-v2 [x]
   - Test rewind (evaluate) [x]
+- benchmark siglip 2 vs this implementation forward pass, debug speed [x]
+- use siglip 2 [x]
 - Overfit on one episode []
 - Cleanup code? []
-- benchmark siglip 2 vs this implementation forward pass, debug speed []
 - Convert python -m lerobot.datasets.v21.convert_dataset_v20_to_v21 --repo-id=IPEC-COMMUNITY/bc_z_lerobot and train on 1 percent
 - Then on 10 percent
 - Ablation dino v2 vs dino v3 base 86 M

src/lerobot/scripts/train.py

@@ -14,12 +14,16 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import logging
+import os
 import time
 from contextlib import nullcontext
 from pprint import pformat
 from typing import Any
 
 import torch
+
+# Fix tokenizer parallelism conflicts with multiprocessing
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
 from termcolor import colored
 from torch.amp import GradScaler
 from torch.optim import Optimizer
@@ -240,6 +244,8 @@ def train(cfg: TrainPipelineConfig):
         sampler=sampler,
         pin_memory=device.type == "cuda",
         drop_last=False,
+        persistent_workers=cfg.num_workers > 0,  # Keep workers alive
+        prefetch_factor=2 if cfg.num_workers > 0 else None,  # Prefetch batches
     )
     dl_iter = cycle(dataloader)