test

app.py

@@ -1,24 +1,52 @@
-# Load model directly
-#from transformers import AutoProcessor, AutoModelForCausalLM
+import av
+import torch
+import numpy as np
+from transformers import LlavaNextVideoForConditionalGeneration, LlavaNextVideoProcessor
 
-#processor = AutoProcessor.from_pretrained("lmms-lab/LLaVA-NeXT-Video-32B-Qwen")
-#model = AutoModelForCausalLM.from_pretrained("lmms-lab/LLaVA-NeXT-Video-32B-Qwen")
+def read_video_pyav(container, indices):
+    '''
+    Decode the video with PyAV decoder.
+    Args:
+        container (`av.container.input.InputContainer`): PyAV container.
+        indices (`List[int]`): List of frame indices to decode.
+    Returns:
+        result (np.ndarray): np array of decoded frames of shape (num_frames, height, width, 3).
+    '''
+    frames = []
+    container.seek(0)
+    start_index = indices[0]
+    end_index = indices[-1]
+    for i, frame in enumerate(container.decode(video=0)):
+        if i > end_index:
+            break
+        if i >= start_index and i in indices:
+            frames.append(frame)
+    return np.stack([x.to_ndarray(format="rgb24") for x in frames])
 
-import gradio as gr
-from transformers import pipeline
+# Load the model in half-precision
+model = LlavaNextVideoForConditionalGeneration.from_pretrained("llava-hf/LLaVA-NeXT-Video-7B-hf", torch_dtype=torch.float16, device_map="auto")
+processor = LlavaNextVideoProcessor.from_pretrained("llava-hf/LLaVA-NeXT-Video-7B-hf")
 
-pipeline = pipeline(task="image-classification", model="llms-lab/LLaVA-NeXT-Video-32BQwen")
+# Load the video as an np.array, sampling uniformly 8 frames (can sample more for longer videos)
+video_path = hf_hub_download(repo_id="raushan-testing-hf/videos-test", filename="sample_demo_1.mp4", repo_type="dataset")
+container = av.open(video_path)
+total_frames = container.streams.video[0].frames
+indices = np.arange(0, total_frames, total_frames / 8).astype(int)
+video = read_video_pyav(container, indices)
 
-def predict(input_img):
-    predictions = pipeline(input_img)
-    return input_img, {p["label"]: p["score"] for p in predictions} 
+conversation = [
+    {
 
-gradio_app = gr.Interface(
-    predict,
-    inputs=gr.Image(label="Select hot dog candidate", sources=['upload', 'webcam'], type="pil"),
-    outputs=[gr.Image(label="Processed Image"), gr.Label(label="Result", num_top_classes=2)],
-    title="Hot Dog? Or Not?",
-)
+        "role": "user",
+        "content": [
+            {"type": "text", "text": "Why is this video funny?"},
+            {"type": "video"},
+            ],
+    },
+]
 
-if __name__ == "__main__":
-    gradio_app.launch()
+prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
+inputs = processor(text=prompt, videos=video, return_tensors="pt")
+
+out = model.generate(**inputs, max_new_tokens=60)
+processor.batch_decode(out, skip_special_tokens=True, clean_up_tokenization_spaces=True)