handler.py

import torch
from transformers import QwenForVisionLanguage, QwenTokenizer, QwenProcessor
from PIL import Image
import base64
import io
import json
import cv2
import numpy as np

class Qwen2VL7bHandler:
    def __init__(self):
        # Initialize the model and processor
        self.model = None
        self.tokenizer = None
        self.processor = None
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    def initialize(self, ctx):
        # Load the model and processor within the inference environment
        model_dir = ctx.system_properties.get("model_dir")
        self.model = QwenForVisionLanguage.from_pretrained(model_dir)
        self.tokenizer = QwenTokenizer.from_pretrained(model_dir)
        self.processor = QwenProcessor.from_pretrained(model_dir)
        self.model.to(self.device)
        self.model.eval()

    def preprocess(self, data):
        # Process incoming requests and extract video data
        video_data = data.get('video')
        if not video_data:
            raise ValueError("Video data is required")
        
        # Decode the base64 video
        frames = self.extract_frames_from_video(video_data)
        inputs = self.processor(images=frames, return_tensors="pt").to(self.device)
        return inputs

    def extract_frames_from_video(self, video_data):
        # Decode the base64 video data
        video_bytes = base64.b64decode(video_data)
        video_array = np.frombuffer(video_bytes, np.uint8)
        video = cv2.imdecode(video_array, cv2.IMREAD_COLOR)
        
        # Capture frames from the video
        vidcap = cv2.VideoCapture(io.BytesIO(video_bytes))
        frames = []
        success, frame = vidcap.read()
        while success:
            # Convert the frame from BGR to RGB format
            frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
            pil_image = Image.fromarray(frame_rgb)
            frames.append(pil_image)
            success, frame = vidcap.read()

        return frames

    def inference(self, inputs):
        # Perform inference on the preprocessed data
        with torch.no_grad():
            outputs = self.model(**inputs)
        return outputs

    def postprocess(self, inference_output):
        # Convert the model outputs into a format suitable for the response
        predicted_text = self.tokenizer.decode(inference_output.logits.argmax(-1))
        return {"result": predicted_text}

    def handle(self, data, context):
        try:
            # Deserialize the request data
            request_data = json.loads(data[0].get("body"))
            # Preprocess the input data
            inputs = self.preprocess(request_data)
            # Perform inference
            outputs = self.inference(inputs)
            # Postprocess the output
            result = self.postprocess(outputs)
            return [json.dumps(result)]
        except Exception as e:
            return [json.dumps({"error": str(e)})]

# Instantiate the handler for use in deployment
_service = Qwen2VL7bHandler()

def handle(data, context):
    if not _service.model:
        _service.initialize(context)
    return _service.handle(data, context)