Add `change_video_resolution_and_fps`, and use it to parse the input video (if necessary).

app.py

@@ -1,4 +1,5 @@
 import gradio as gr
+import tempfile
 import torch
 from pytorchvideo.data import make_clip_sampler
 from pytorchvideo.data.clip_sampling import ClipInfoList
@@ -15,6 +16,8 @@ from torchvision.transforms import (
 )
 from transformers import VideoMAEForVideoClassification, VideoMAEFeatureExtractor
 
+from video_utils import change_video_resolution_and_fps
+
 MODEL_CKPT = "omermazig/videomae-finetuned-nba-5-class-4-batch-8000-vid-multiclass"
 DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 CLIPS_FROM_SINGLE_VIDEO = 5
@@ -52,8 +55,13 @@ labels = list(trained_model.config.label2id.keys())
 
 def parse_video_to_clips(video_file):
     """A utility to parse the input videos """
-    video_path_handler = VideoPathHandler()
-    video: EncodedVideoPyAV = video_path_handler.video_from_path(video_file)
+    new_resolution = (320, 256)
+    new_fps = 30
+    with tempfile.NamedTemporaryFile() as new_video:
+        print(new_video.name)
+        change_video_resolution_and_fps(video_file, new_video.name, new_resolution, new_fps)
+        video_path_handler = VideoPathHandler()
+        video: EncodedVideoPyAV = video_path_handler.video_from_path(video_file)
 
     clip_sampler = make_clip_sampler("random_multi", clip_duration, CLIPS_FROM_SINGLE_VIDEO)
     # noinspection PyTypeChecker

video_utils.py

@@ -0,0 +1,60 @@
+import shutil
+from typing import Optional, Tuple
+
+import cv2
+
+
+def change_video_resolution_and_fps(video_path: str, output_path: str,
+                                    new_resolution: Optional[Tuple[int, int]] = None,
+                                    new_fps: Optional[int] = None) -> bool:
+    cap = cv2.VideoCapture(video_path)
+
+    try:
+        original_width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+        original_height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+        resolution = (original_width, original_height)
+        fps = int(cap.get(cv2.CAP_PROP_FPS))
+        # Specify the new_resolution and new_fps for the cut video
+        new_resolution = new_resolution if new_resolution else resolution
+        new_fps = new_fps if new_fps else fps
+        fps_decrease_factor = fps / new_fps
+        if not fps_decrease_factor.is_integer():
+            raise ValueError(f"New fps ({new_fps}) must be a divisor of the current fps ({fps})")
+
+        if fps == new_fps and (original_width, original_height) == new_resolution:
+            shutil.copy(video_path, output_path)
+            return True
+
+        current_frame = 0
+
+        # Read first frame
+        ret, frame = cap.read()
+        if not ret:
+            # Video has ended, without us recording anything
+            return False
+
+        # Initialize the video writer to save the cut video
+        fourcc = cv2.VideoWriter_fourcc(*'XVID')
+        out = cv2.VideoWriter(output_path, fourcc, new_fps, new_resolution, isColor=True)
+        new_video_current_frame = 0
+
+        while ret:
+            if current_frame % fps_decrease_factor == 0:
+                if resolution != new_resolution:
+                    # Resize the frame to the desired new_resolution before writing it
+                    frame = cv2.resize(frame, new_resolution, interpolation=cv2.INTER_AREA)
+                # Write the frame to the cut video
+                out.write(frame)
+                new_video_current_frame += 1
+
+            current_frame += 1
+            # Read next frame
+            ret, frame = cap.read()
+
+        # We're done recording
+        out.release()
+        return new_video_current_frame > 0
+
+    finally:
+        # Release the video capture and close all windows
+        cap.release()