[Feature] Add training code for the baseline of VLLN Bench (#198)

* add VL-LN Bench training code

* add VL-LN Bench training code

* "Remove VLLN trainer; unify training for VLN and IION datasets."

* solve the issue from kellyiss and kew6688

* solve the issue from Tai-Wang

* (1) Remove `dataset_utils.py`. (2) Add standard docstrings to the main class and key functions.

* solve the issue from Tai-Wang

* solve the issue from Tai-Wang

* refine the docstring

Author: DuangZhu

internnav/dataset/internvla_n1_lerobot_dataset.py

@@ -17,6 +17,7 @@
 from torchcodec.decoders import VideoDecoder
 from transformers.image_utils import to_numpy_array
 
+from .vlln_lerobot_dataset import VLLNDataset
 from .rope2d import get_rope_index_2, get_rope_index_25
 
 # Define placeholders for dataset paths
@@ -150,6 +151,11 @@ def parse_sampling_rate(dataset_name):
     return 1.0
 
 
+def read_jsonl(path):
+    with open(path, "r") as f:
+        return [json.loads(line) for line in f]
+
+
 def data_list(dataset_names):
     config_list = []
     for dataset_name in dataset_names:
@@ -180,11 +186,6 @@ def rank0_print(*args):
         print(*args)
 
 
-def read_jsonl(path):
-    with open(path, "r") as f:
-        return [json.loads(line) for line in f]
-
-
 def preprocess_qwen_2_visual(
     sources,
     tokenizer: transformers.PreTrainedTokenizer,
@@ -1329,11 +1330,50 @@ def __call__(self, instances: Sequence[Dict]) -> Dict[str, torch.Tensor]:
 
         return batch
 
+class CombinedDataset(Dataset):
+    """
+    Combine multiple datasets into a single dataset interface.
+
+    This class is used to merge different datasets for joint training.
+    It concatenates samples from all provided datasets and optionally shuffles
+    the global index mapping (without changing the underlying datasets).
+    """
+    def __init__(self, datasets, shuffle=False):
+        super(CombinedDataset, self).__init__()
+        self.datasets = datasets
+        self.lengths = [len(dataset) for dataset in datasets]
+        self.cum_lengths = np.cumsum(self.lengths)
+        self.total_length = self.cum_lengths[-1]
+        self.shuffle_enabled = shuffle
+        self.indices = np.arange(self.total_length)
+        if self.shuffle_enabled:
+            self.shuffle()
+
+    def shuffle(self):
+        np.random.shuffle(self.indices)
+
+    def _map_index(self, idx):
+        return self.indices[idx]
+
+    def __len__(self):
+        return self.cum_lengths[-1]
+
+    def __getitem__(self, i):
+        real_idx = self._map_index(i)
+        for idx, cum_len in enumerate(self.cum_lengths):
+            if real_idx < cum_len:
+                return self.datasets[idx][real_idx - cum_len + self.lengths[idx]]
+        raise ValueError(f"Index {real_idx} out of bound")
+
 
 def make_supervised_data_module(tokenizer: transformers.PreTrainedTokenizer, data_args) -> Dict:
     """Make dataset and collator for supervised fine-tuning."""
-    train_dataset = NavPixelGoalDataset(tokenizer=tokenizer, data_args=data_args)
-    # train_dataset = LazySupervisedDataset(tokenizer=tokenizer, data_args=data_args)
+    train_datasets = []
+    if data_args.iion_dataset_use:
+        train_datasets.append(VLLNDataset(tokenizer=tokenizer, data_args=data_args))
+    if data_args.vln_dataset_use:
+        train_datasets.append(NavPixelGoalDataset(tokenizer=tokenizer, data_args=data_args))
+    train_dataset = CombinedDataset(train_datasets, shuffle=False)
     if data_args.data_flatten:
         data_collator = FlattenedDataCollatorForSupervisedDataset(tokenizer=tokenizer)
         return dict(train_dataset=train_dataset, eval_dataset=None, data_collator=data_collator)