Make using sampler in train.py more explicit

lerobot/scripts/train.py

@@ -348,15 +348,22 @@ def train(cfg: DictConfig, out_dir: str | None = None, job_name: str | None = No
             logging.info("Resume training")
 
     # create dataloader for offline training
+    if cfg.get("drop_n_last_frames"):
+        shuffle = False
+        sampler = EpisodeAwareSampler(
+            offline_dataset.episode_data_index,
+            drop_n_last_frames=cfg.drop_n_last_frames,
+            shuffle=True,
+        )
+    else:
+        shuffle = True
+        sampler = None
     dataloader = torch.utils.data.DataLoader(
         offline_dataset,
         num_workers=4,
         batch_size=cfg.training.batch_size,
-        sampler=EpisodeAwareSampler(
-            offline_dataset.episode_data_index,
-            drop_n_last_frames=cfg.drop_n_last_frames if hasattr(cfg, "drop_n_last_frames") else 0,
-            shuffle=True,
-        ),
+        shuffle=shuffle,
+        sampler=sampler,
         pin_memory=device.type != "cpu",
         drop_last=False,
     )