inpaint.py

import os
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from types import SimpleNamespace
from deepfillv2 import test_dataset, utils
from config import *

class InpaintingTester:
    def __init__(self, save_path, resize_to=None):
        if resize_to is None:
            resize_to = RESIZE_TO
        self.save_path = save_path
        self.setsize = resize_to

        # Build the generator network
        opt = SimpleNamespace(
            pad_type=PAD_TYPE,
            in_channels=IN_CHANNELS,
            out_channels=OUT_CHANNELS,
            latent_channels=LATENT_CHANNELS,
            activation=ACTIVATION,
            norm=NORM,
            init_type=INIT_TYPE,
            init_gain=INIT_GAIN,
            use_cuda=CUDA,
            gpu_device=GPU_DEVICE,
        )

        # Initialize generator (only once)
        self.generator = utils.create_generator(opt).eval()

        # Load pretrained model weights
        # print("-- INPAINT: Loading Pretrained Model --")
        self.load_model_generator(self.generator)

        # Move the generator to GPU
        self.generator = self.generator.to(GPU_DEVICE)

    def load_model_generator(self, generator):
        pretrained_dict = torch.load(
            DEEPFILL_MODEL_PATH, map_location=torch.device(GPU_DEVICE), weights_only=True
        )
        generator.load_state_dict(pretrained_dict)

    def process_image(self, in_image, mask_image, save_image_path):
        # Initialize dataset and dataloader
        trainset = test_dataset.InpaintDataset(in_image, mask_image, self.setsize)
        dataloader = DataLoader(
            trainset,
            batch_size=1,
            shuffle=False,
            num_workers=8,
            pin_memory=True,
        )

        # Testing loop for a single image
        for batch_idx, (img, mask) in enumerate(dataloader):
            img = img.to(GPU_DEVICE)
            mask = mask.to(GPU_DEVICE)

            # Generator output
            with torch.no_grad():
                first_out, second_out = self.generator(img, mask)

            # Combine outputs with input
            first_out_wholeimg = img * (1 - mask) + first_out * mask
            second_out_wholeimg = img * (1 - mask) + second_out * mask

            masked_img = img * (1 - mask) + mask
            mask = torch.cat((mask, mask, mask), 1)
            img_list = [second_out_wholeimg]
            name_list = ["second_out"]

            # Save the sample image
            results_path = os.path.dirname(save_image_path)
            if not os.path.exists(results_path):
                os.makedirs(results_path)

            utils.save_sample_png(
                sample_folder=results_path,
                sample_name=os.path.basename(save_image_path),
                img_list=img_list,
                name_list=name_list,
                pixel_max_cnt=255,
            )

    def process_multiple_images(self, image_mask_pairs):
        # Iterate through a list of image/mask pairs and save results
        png_images=[]
        for img_path, mask_path in image_mask_pairs:
            try:
                save_image_path = os.path.join(self.save_path, os.path.basename(img_path))
                print(f"Processing: {img_path} and {mask_path}")
                self.process_image(img_path, mask_path, save_image_path)
                extention = os.path.splitext(save_image_path)[1]
                save_at=save_image_path.replace(extention, ".png")
                png_images.append(save_at)
            except Exception as e:
                if self.save_path in png_images:
                    png_images.pop()
                    png_images.append(None)
                print(f"Error: {e}")
        # print("-- All Inpainting is finished --")
        return png_images

# Main execution
# if __name__ == "__main__":
#     save_path = "./output"
#     resize_to = None  # Default size from config

#     # List of image and mask pairs
#     image_mask_pairs = [
#         ( "./input/image.jpg", "./input/mask.jpg"),
#     ]

#     tester = InpaintingTester(save_path, resize_to)
    
#     # Process multiple images using a loop
#     results=tester.process_multiple_images(image_mask_pairs)
#     print(results)