handler.py

import torch
from PIL import Image
from io import BytesIO
from realesrgan import RealESRGANer
from typing import Dict, List, Any
import os
from pathlib import Path
from basicsr.archs.rrdbnet_arch import RRDBNet
import numpy as np
import cv2


import torch
import base64
# torch.cuda.empty_cache()
# torch.cuda.set_per_process_memory_fraction(0.99)
# os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:64,garbage_collection_threshold:0.7"


class EndpointHandler:
    def __init__(self, path=""):

        self.model = RealESRGANer(
            scale=4,
            model_path=f"/repository/weights/Real-ESRGAN-x4plus.pth",
            # dni_weight=dni_weight,
            model= RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4),
            tile=1000,
            tile_pad=10,
            # pre_pad=args.pre_pad,
            half=True,
            # gpu_id=args.gpu_id
        )

    def __call__(self, data: Any) -> Dict[str, List[float]]:
        inputs = data.pop("inputs", data)
        outscale = 3
        
        # decode base64 image to PIL
        image = Image.open(BytesIO(base64.b64decode(inputs['image'])))
        # Convert PIL image to NumPy array
        opencv_image = np.array(image)
        # Convert RGB to BGR (PIL uses RGB, OpenCV expects BGR)
        opencv_image = cv2.cvtColor(opencv_image, cv2.COLOR_RGB2BGR)
        output, _ = self.model.enhance(opencv_image, outscale=outscale)
        # BGR to RGB
        output = cv2.cvtColor(output, cv2.COLOR_BGR2RGB)
        
        img_byte_arr = BytesIO()
        output = Image.fromarray(output)
        output.save(img_byte_arr, format='PNG')
        img_str = base64.b64encode(img_byte_arr.getvalue())
        return {"out_image": img_str.decode()}