Update demo_vc3_proactive.py

demo_vc3_proactive.py

@@ -38,9 +38,9 @@ MODEL_PATH = "MCG-NJU/VideoChat3-4B"
 VIDEO_PATH = "example.mp4"
 QUESTION = "Your question"
 # ----------------------------------------------------------
-TARGET_FPS = 4.0
+TARGET_FPS = 2.0
 MAX_PIXELS = 224*224          # pixel budget for normal frames
-MAX_ROUNDS = 32
+MAX_ROUNDS = 16
 MAX_NEW_TOKENS = 128
 # ----------------------------------------------------------
 
@@ -114,7 +114,7 @@ def main():
             with torch.inference_mode():
                 out = model.generate(
                     **inputs, max_new_tokens=max_tokens,
-                    do_sample=True, temperature=0.7
+                    do_sample=False, temperature=1.0, top_p=1.0, top_k=0,
                 )
             trimmed = [o[len(i):] for i, o in zip(inputs.input_ids, out)]
             text = processor.batch_decode(
@@ -129,7 +129,8 @@ def main():
     )
 
     extractor = VideoFrameExtractor(args.video, target_fps=TARGET_FPS)
-    frames_per_round = max(1, round(TARGET_FPS))
+    actual_fps = extractor.actual_fps if extractor.actual_fps > 0 else TARGET_FPS
+    frames_per_round = max(1, round(actual_fps))
     standby_remaining = 0
 
     try:
@@ -150,8 +151,8 @@ def main():
             frames = [_resize_frame(frame, max_px) for frame in raw_frames]
 
             round_idx = frame_start // frames_per_round
-            time_start = frame_start / TARGET_FPS
-            time_end = (frame_start + len(frames)) / TARGET_FPS
+            time_start = frame_start / actual_fps
+            time_end = (frame_start + len(frames)) / actual_fps
             answer = session.step(
                 frames,
                 round_idx=round_idx,