Video-LLaVa: handle any number of frames (#31221)

video-llava can handle more frames

src/transformers/models/video_llava/modeling_video_llava.py

@@ -287,7 +287,7 @@ class VideoLlavaForConditionalGeneration(VideoLlavaPreTrainedModel):
         num_images, num_image_patches, embed_dim = visual_features.shape
         batch_size, sequence_length = input_ids.shape
         left_padding = not torch.sum(input_ids[:, -1] == torch.tensor(self.pad_token_id))
-        special_vision_token = self.config.video_token_index if num_frames == 8 else self.config.image_token_index
+        special_vision_token = self.config.video_token_index if num_frames > 1 else self.config.image_token_index
 
         # 1. Create a mask to know where special image tokens are
         special_image_token_mask = input_ids == special_vision_token
@@ -375,14 +375,13 @@ class VideoLlavaForConditionalGeneration(VideoLlavaPreTrainedModel):
         # videos do not need to select features and it's always "full" (as it is done in the orig implementation)
         if pixel_values_videos is not None:
             batch_size_vid, num_frames, channels, height, width = pixel_values_videos.shape
-            if num_frames != 8:
-                raise ValueError(f"Video pixel values should have exactly `8` frames but foung `{num_frames}`")
 
             pixel_values = pixel_values_videos.reshape(batch_size_vid * num_frames, channels, height, width)
             video_outputs = self.video_tower(pixel_values, output_hidden_states=True)
             video_outputs = video_outputs.hidden_states[vision_feature_layer].squeeze(1)
         else:
             video_outputs = None
+            num_frames = 0
 
         if pixel_values_images is not None:
             image_outputs = self.image_tower(pixel_values_images, output_hidden_states=True)
@@ -397,7 +396,7 @@ class VideoLlavaForConditionalGeneration(VideoLlavaPreTrainedModel):
         else:
             image_outputs = None
 
-        return image_outputs, video_outputs
+        return image_outputs, video_outputs, num_frames
 
     @add_start_docstrings_to_model_forward(VIDEO_LLAVA_INPUTS_DOCSTRING)
     @replace_return_docstrings(output_type=VideoLlavaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
@@ -513,7 +512,7 @@ class VideoLlavaForConditionalGeneration(VideoLlavaPreTrainedModel):
 
             # 2. Merge text and images
             if (pixel_values_images is not None or pixel_values_videos is not None) and input_ids.shape[1] != 1:
-                image_outputs, video_outputs = self._get_vision_features(
+                image_outputs, video_outputs, num_frames = self._get_vision_features(
                     pixel_values_images=pixel_values_images,
                     pixel_values_videos=pixel_values_videos,
                     vision_feature_layer=vision_feature_layer,
@@ -546,7 +545,7 @@ class VideoLlavaForConditionalGeneration(VideoLlavaPreTrainedModel):
                         input_ids,
                         attention_mask,
                         labels,
-                        num_frames=8,
+                        num_frames=num_frames,
                     )
             else:
                 # In case input_ids.shape[1] == 1 & past_key_values != None, we are in the case of

tests/models/video_llava/test_modeling_video_llava.py

@@ -487,6 +487,9 @@ class VideoLlavaForConditionalGenerationIntegrationTest(unittest.TestCase):
             repo_id="raushan-testing-hf/videos-test", filename="video_demo.npy", repo_type="dataset"
         )
         video_file = np.load(video_file)
+
+        # let's expand it for 16 frames, to check model can handle any number of frames
+        video_file = video_file.repeat(2, 0)
         inputs = self.processor(prompt, videos=video_file, return_tensors="pt").to(torch_device, torch.float16)
 
         # Make sure that `generate` works