Merge branch 'cache_dataset' into train_tdmpc

lerobot/common/datasets/factory.py

@@ -28,6 +28,7 @@ def make_dataset(
         cfg.dataset_repo_id,
         split=split,
         delta_timestamps=delta_timestamps,
+        use_cache=cfg.training.dataset_use_cache,
     )
 
     if cfg.get("override_dataset_stats"):

lerobot/common/datasets/lerobot_dataset.py

@@ -27,7 +27,14 @@ class LeRobotDataset(torch.utils.data.Dataset):
         split: str = "train",
         transform: callable = None,
         delta_timestamps: dict[list[float]] | None = None,
+        use_cache: bool = False,
     ):
+        """
+        Args:
+            use_cache: Enable this to cache all items as tensors for faster data loading after the first
+                epoch. Useful if you have a small enough dataset to fit into memory. You may set multiple
+                workers for the PyTorch Dataloader but remember to set persistent_workers=True.
+        """
         super().__init__()
         self.repo_id = repo_id
         self.version = version
@@ -44,6 +51,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         self.info = load_info(repo_id, version, root)
         if self.video:
             self.videos_dir = load_videos(repo_id, version, root)
+        self.cache = {} if use_cache else None
 
     @property
     def fps(self) -> int:
@@ -104,9 +112,12 @@ class LeRobotDataset(torch.utils.data.Dataset):
         return 1 / self.fps - 1e-4
 
     def __len__(self):
-        return self.num_samples
+        return self.num_samples // 8
 
     def __getitem__(self, idx):
+        if self.cache is not None and idx in self.cache:
+            item = self.cache[idx]
+        else:
             item = self.hf_dataset[idx]
 
             if self.delta_timestamps is not None:

lerobot/configs/default.yaml

@@ -26,6 +26,12 @@ training:
   save_freq: ???
   log_freq: 250
   save_model: true
+  # `dataset_use_cache` indicates whether to cache all dataset items as Tensors in RAM. Potentially useful for
+  # faster data loading with datasets small enough to fit in memory. If you wish to use dataloader workers,
+  # remember to set `dataloader_persistent_workers to True.
+  dataset_use_cache: false
+  dataloader_num_workers: 4
+  dataloader_persistent_workers: false
 
 eval:
   n_episodes: 1

lerobot/configs/policy/tdmpc.yaml

@@ -10,6 +10,9 @@ training:
   online_steps_between_rollouts: 1
   online_sampling_ratio: 0.5
   online_env_seed: 10000
+  dataset_use_cache: true
+  dataloader_num_workers: 4
+  dataloader_persistent_workers: true
 
   batch_size: 256
   grad_clip_norm: 10.0

lerobot/scripts/train.py

@@ -368,7 +368,8 @@ def train(cfg: dict, out_dir=None, job_name=None):
     # create dataloader for offline training
     dataloader = torch.utils.data.DataLoader(
         offline_dataset,
-        num_workers=4,
+        num_workers=cfg.training.dataloader_num_workers,
+        persistent_workers=cfg.training.dataloader_persistent_workers,
         batch_size=cfg.training.batch_size,
         shuffle=True,
         pin_memory=cfg.device != "cpu",
@@ -415,7 +416,8 @@ def train(cfg: dict, out_dir=None, job_name=None):
     )
     dataloader = torch.utils.data.DataLoader(
         concat_dataset,
-        num_workers=4,
+        num_workers=cfg.training.dataloader_num_workers,
+        persistent_workers=cfg.training.dataloader_persistent_workers,
         batch_size=cfg.training.batch_size,
         sampler=sampler,
         pin_memory=cfg.device != "cpu",