change all 'iion' to 'iign'

Author: DuangZhu

internnav/dataset/internvla_n1_lerobot_dataset.py

@@ -1371,7 +1371,7 @@ def __getitem__(self, i):
 def make_supervised_data_module(tokenizer: transformers.PreTrainedTokenizer, data_args) -> Dict:
     """Make dataset and collator for supervised fine-tuning."""
     train_datasets = []
-    if data_args.iion_dataset_use:
+    if data_args.iign_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))

internnav/dataset/vlln_lerobot_dataset.py

@@ -15,31 +15,31 @@
 from .rope2d import get_rope_index_2, get_rope_index_25
 
 # Define placeholders for dataset paths
-IION_split1 = {
+IIGN_split1 = {
     "data_path": "projects/VL-LN-Bench/traj_data/mp3d_split1",
     "height": 125,
     "pitch_1": 0,
     "pitch_2": 30,
 }
 
-IION_split2 = {
+IIGN_split2 = {
     "data_path": "projects/VL-LN-Bench/traj_data/mp3d_split2",
     "height": 125,
     "pitch_1": 0,
     "pitch_2": 30,
 }
 
-IION_split3 = {
+IIGN_split3 = {
     "data_path": "projects/VL-LN-Bench/traj_data/mp3d_split3",
     "height": 125,
     "pitch_1": 0,
     "pitch_2": 30,
 }
 
 data_dict = {
-    "iion_split1": IION_split1,
-    "iion_split2": IION_split2,
-    "iion_split3": IION_split3,
+    "iign_split1": IIGN_split1,
+    "iign_split2": IIGN_split2,
+    "iign_split3": IIGN_split3,
 }
 
 IGNORE_INDEX = -100
@@ -55,14 +55,14 @@
 
 class VLLNDataset(Dataset):
     """
-    Dataset for 'Vision-Language'-'Language-Navigation' (VL-LN) / IION-style training.
+    Dataset for 'Vision-Language'-'Language-Navigation' (VL-LN) / IIGN-style training.
 
     Args:
         tokenizer (transformers.PreTrainedTokenizer): Tokenizer used to encode
             the chat template and produce `input_ids` / `labels`.
         data_args: A config-like object that must provide at least:
-            - iion_dataset_use (str): comma-separated dataset names, optionally
-              with sampling rate suffix like `iion_split1%50`.
+            - iign_dataset_use (str): comma-separated dataset names, optionally
+              with sampling rate suffix like `iign_split1%50`.
             - model_type (str): decides which rope-index function to use.
             - sample_step (int): stride for sampling start frames.
             - pixel_goal_only (bool): whether to keep only pixel-goal samples.
@@ -74,7 +74,7 @@ class VLLNDataset(Dataset):
 
     def __init__(self, tokenizer: transformers.PreTrainedTokenizer, data_args):
         super(VLLNDataset, self).__init__()
-        dataset = data_args.iion_dataset_use.split(",")
+        dataset = data_args.iign_dataset_use.split(",")
         dataset_list = data_list(dataset)
         rank0_print(f"Loading datasets: {dataset_list}")
         self.video_max_total_pixels = getattr(data_args, "video_max_total_pixels", 1664 * 28 * 28)

internnav/trainer/internvla_n1_argument.py

@@ -29,7 +29,7 @@ class DataArguments:
     video_min_frame_pixels: int = field(default=4 * 28 * 28)
 
     vln_dataset_use: str = field(default="")
-    iion_dataset_use: str = field(default="")
+    iign_dataset_use: str = field(default="")
     sample_step: int = field(default=4)
     num_history: Optional[int] = field(default=8)
     predict_step_num: Optional[int] = field(default=32)

scripts/train/qwenvl_train/train_system2_vlln.sh

@@ -27,7 +27,7 @@ max_pixels=313600
 min_pixels=3136
 
 # Dataset configuration (replace with public dataset names)
-iion_datasets=iion_split1,iion_split2 #,iion_split3
+iign_datasets=iign_split1,iign_split2 #,iign_split3
 
 # Output configuration
 run_name=InternVLA-N1-vlln
@@ -38,7 +38,7 @@ srun torchrun --nnodes=$SLURM_NNODES --nproc_per_node=8 \
     internnav/trainer/internvla_vlln_trainer.py \
     --deepspeed ${deepspeed} \
     --model_name_or_path "${llm}" \
-    --iion_dataset_use ${iion_datasets} \
+    --iign_dataset_use ${iign_datasets} \
     --data_flatten False \
     --tune_mm_vision True \
     --tune_mm_mlp True \