Fix: Standardize torchtext shared library loading path for multi-platform compatibility

config/detection/ssd_coco/mobilevit.yaml

@@ -7,8 +7,8 @@ common:
   auto_resume: true
   mixed_precision: true
 dataset:
-  root_train: "/mnt/vision_datasets/coco"
-  root_val: "/mnt/vision_datasets/coco"
+  root_train: "/content/data/train2017"
+  root_val: "/content/data/val2017"
   name: "coco_ssd"
   category: "detection"
   train_batch_size0: 32 # effective batch size is 128 (32 * 4 GPUs)

data/datasets/detection/coco_base.py

@@ -52,7 +52,7 @@ def __init__(
         logger.disable_printing()
 
         self.coco = COCO(ann_file)
-        self.img_dir = os.path.join(self.root, "images/{}{}".format(split, year))
+        self.img_dir = self.root
         self.ids = (
             list(self.coco.imgToAnns.keys())
             if self.is_training
@@ -139,9 +139,17 @@ def __getitem__(
 
         transform_fn = self.get_augmentation_transforms(size=(crop_size_h, crop_size_w))
 
-        image_id = self.ids[img_index]
+        num_images = len(self.ids)
+        for _ in range(num_images):
+            image_id = self.ids[img_index]
 
-        image, img_name = self.get_image(image_id=image_id)
+            image, img_name = self.get_image(image_id=image_id)
+            if image is not None:
+                break
+            # corrupted image; try next
+            img_index = (img_index + 1) % num_images
+        else:
+            raise RuntimeError("All images in the dataset are corrupted.")
         im_width, im_height = image.size
 
         boxes, labels, mask = self.get_boxes_and_labels(

data/datasets/segmentation/coco_segmentation.py

@@ -31,7 +31,7 @@ def __init__(self, opts: argparse.Namespace, *args, **kwargs) -> None:
         ann_file = os.path.join(
             self.root, "annotations/instances_{}{}.json".format(split, year)
         )
-        self.img_dir = os.path.join(self.root, "images/{}{}".format(split, year))
+        self.img_dir = self.root
         self.split = split
         self.coco = COCO(ann_file)
         self.coco_mask = mask

data/text_tokenizer/clip_tokenizer.py

@@ -7,16 +7,26 @@
 
 import torch
 from torch import Tensor
-from torchtext.transforms import CLIPTokenizer
 
 from data.text_tokenizer import TOKENIZER_REGISTRY, BaseTokenizer
 from utils import logger
 from utils.download_utils import get_local_path
 
+try:
+    from torchtext.transforms import CLIPTokenizer
+except (ImportError, OSError):
+    CLIPTokenizer = None
+
 
 @TOKENIZER_REGISTRY.register(name="clip")
 class ClipTokenizer(BaseTokenizer):
     def __init__(self, opts, *args, **kwargs):
+        if CLIPTokenizer is None:
+            logger.error(
+                "torchtext is not available. Please install a compatible version "
+                "of torchtext to use the CLIP tokenizer."
+            )
+
         merges_path = getattr(opts, "text_tokenizer.clip.merges_path", None)
         if merges_path is None:
             logger.error(

data/transforms/video.py

@@ -16,9 +16,27 @@
 import torch
 import torchaudio
 from torch.nn import functional as F
-from torchvision.io import write_video
 from torchvision.transforms import functional as FV
 
+try:
+    import cv2
+
+    def write_video(filename: str, video_array: torch.Tensor, fps: float):
+        """Write video using OpenCV (replacement for removed torchvision.io.write_video)."""
+        video_array = video_array.cpu().numpy()  # (N, H, W, C) uint8
+        n_frames = video_array.shape[0]
+        h, w = video_array.shape[1:3]
+        fourcc = cv2.VideoWriter_fourcc(*"mp4v")
+        out = cv2.VideoWriter(str(filename), fourcc, fps, (w, h))
+        for i in range(n_frames):
+            frame = video_array[i]
+            # cv2 expects BGR
+            frame_bgr = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
+            out.write(frame_bgr)
+        out.release()
+except ImportError:
+    from torchvision.io import write_video
+
 from data.transforms import TRANSFORMATIONS_REGISTRY, BaseTransformation
 from data.transforms.utils import *
 from options.parse_args import JsonValidator

main_eval.py

@@ -154,5 +154,5 @@ def main_worker_detection(args: Optional[List[str]] = None, **kwargs):
 
 if __name__ == "__main__":
     # main_worker()
-    main_worker_segmentation()
-    # main_worker_detection()
+    #main_worker_segmentation()
+    main_worker_detection()