app.py

import os
# Configure Hugging Face caches to use the writable /cache volume in Spaces
os.environ["HF_HOME"] = "/cache"
os.environ["TRANSFORMERS_CACHE"] = "/cache"
os.environ["HF_DATASETS_CACHE"] = "/cache"

from flask import Flask, request, jsonify
import numpy as np
import torch
import av
import cv2
import tempfile
import shutil
import logging
from transformers import VideoMAEForVideoClassification, VideoMAEImageProcessor
from PIL import Image
from torchvision.transforms import Compose, Resize, ToTensor

# Initialize Flask app
app = Flask(__name__)

# Configure logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# Globals for model, processor, and transforms
device = "cuda" if torch.cuda.is_available() else "cpu"
model = None
processor = None
transform = None


def load_model():
    """Load the model and processor into globals"""
    global model, processor, transform
    if model is None:
        model_name = "OPear/videomae-large-finetuned-UCF-Crime"
        logger.info(f"Loading model {model_name} on device {device}")
        # Downloads will go to /cache automatically
        model = VideoMAEForVideoClassification.from_pretrained(model_name).to(device)
        processor = VideoMAEImageProcessor.from_pretrained(model_name)
        transform = Compose([
            Resize((224, 224)),
            ToTensor(),
        ])
        logger.info("Model and processor loaded successfully")
    return model, processor, transform


def sample_frame_indices(clip_len=16, frame_sample_rate=1, seg_len=0):
    """Uniformly sample exactly 16 frame indices from a clip"""
    if seg_len <= clip_len:
        indices = np.linspace(0, seg_len - 1, num=clip_len, dtype=int)
    else:
        end_idx = np.random.randint(clip_len, seg_len)
        start_idx = max(0, end_idx - clip_len)
        indices = np.linspace(start_idx, end_idx - 1, num=clip_len, dtype=int)
    return np.clip(indices, 0, seg_len - 1)


def process_video(video_path):
    """Extract 16 uniformly-sampled frames from the video"""
    try:
        container = av.open(video_path)
        video_stream = container.streams.video[0]
        seg_len = video_stream.frames if video_stream.frames > 0 else int(
            cv2.VideoCapture(video_path).get(cv2.CAP_PROP_FRAME_COUNT)
        )
    except Exception as e:
        logger.error(f"Error opening video: {e}")
        return None, None

    indices = sample_frame_indices(clip_len=16, seg_len=seg_len)
    frames = []

    # Try PyAV decode
    try:
        container.seek(0)
        for i, frame in enumerate(container.decode(video=0)):
            if i > indices[-1]:
                break
            if i in indices:
                frames.append(frame.to_ndarray(format="rgb24"))
    except Exception as e:
        logger.warning(f"PyAV decoding failed, falling back to OpenCV: {e}")

    # Fallback to OpenCV if necessary
    if len(frames) < len(indices):
        cap = cv2.VideoCapture(video_path)
        for i in indices:
            cap.set(cv2.CAP_PROP_POS_FRAMES, i)
            ret, frame = cap.read()
            if ret:
                frames.append(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
        cap.release()

    if len(frames) != 16:
        logger.error(f"Expected 16 frames, got {len(frames)}")
        return None, None

    return np.stack(frames), indices


def predict_video(frames):
    """Run inference on a stack of 16 frames"""
    model, processor, transform = load_model()
    video_tensor = torch.stack([transform(Image.fromarray(f)) for f in frames])
    video_tensor = video_tensor.unsqueeze(0)

    inputs = processor(list(video_tensor[0]), return_tensors="pt", do_rescale=False)
    inputs = {k: v.to(device) for k, v in inputs.items()}

    with torch.no_grad():
        outputs = model(**inputs)
    logits = outputs.logits
    pred_id = logits.argmax(-1).item()
    return model.config.id2label.get(pred_id, "Unknown")


@app.route('/classify-video', methods=['POST'])
def classify_video():
    if 'video' not in request.files:
        return jsonify({'error': 'No video file provided'}), 400

    file = request.files['video']
    if file.filename == '':
        return jsonify({'error': 'Empty filename'}), 400

    temp_dir = tempfile.mkdtemp()
    path = os.path.join(temp_dir, file.filename)
    try:
        file.save(path)
        frames, _ = process_video(path)
        if frames is None:
            return jsonify({'error': 'Failed to extract frames'}), 400
        prediction = predict_video(frames)
        return jsonify({'prediction': prediction})
    except Exception as e:
        logger.exception(f"Error during processing: {e}")
        return jsonify({'error': str(e)}), 500
    finally:
        shutil.rmtree(temp_dir, ignore_errors=True)


@app.route('/health', methods=['GET'])
def health_check():
    return jsonify({'status': 'healthy'}), 200


if __name__ == '__main__':
    # Preload model on startup
    logger.info("Starting application and loading model...")
    load_model()
    port = int(os.environ.get('PORT', 7860))
    app.run(host='0.0.0.0', port=port, debug=False)