less video prefetch

src/lerobot/policies/rlearn/configuration_rlearn.py

@@ -61,7 +61,7 @@ class RLearNConfig(PreTrainedConfig):
     use_tanh_head: bool = False  # when True, bound outputs in [-1, 1]
 
     # Training
-    learning_rate: float = 3e-5
+    learning_rate: float = 1e-4
     weight_decay: float = 0.01
     
     # Performance optimizations

src/lerobot/policies/rlearn/rlearn_plan.md

@@ -82,15 +82,17 @@ _ Open X-Embodiment (OXE)
 - benchmark siglip 2 vs this implementation forward pass, debug speed [x]
 - use siglip 2 [x]
 - Overfit on one episode []
-- Cleanup code? []
+- Cleanup code? [] + enable language loss
 - 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
+- Then on 10 percent []
+- Ablation 16 sucessive frame vs 16 frame samples with stride 2 or 4 []
 - Add more artificial text to dataset generated by vlm (google gemini) []
   - See google gemini vlm caption [] https://gemini.google.com/app/7e332ffaf32580f2
   - Multiple captions per video, creat method to generate as much data as possible etc [] https://arxiv.org/abs/2508.13446, https://arxiv.org/pdf/2412.04453
-- How can we improve spatial aware learning? solve issue of Contrastive learning and position
-- Extend evaluation []
 - Add other datasets from OXE metioned in rewind []
+- Extend evaluation []
 - Ablation for size vision encoder, language encoder, temporal head []
 - Add other datasets metnioned here []
+- How can we improve spatial aware learning? solve issue of Contrastive learning and position []
+
+

src/lerobot/scripts/train.py

@@ -258,7 +258,7 @@ def train(cfg: TrainPipelineConfig):
         pin_memory=device.type == "cuda",
         drop_last=False,
         persistent_workers=cfg.num_workers > 0,  # Keep workers alive between epochs
-        prefetch_factor=12,  # Maximum prefetch for video pipeline
+        prefetch_factor=6,  # Maximum prefetch for video pipeline
         timeout=30,  # Prevent hanging on video decode errors
     )
     dl_iter = cycle(dataloader)