app.py

import gradio as gr
import torch
import torchvision.transforms as transforms
import os
from PIL import Image

import mynet

transform = transform = transforms.Compose([
  transforms.ToTensor()
])

models = {
    "kodak_gold_200.pth", "fuji_color_200.pth",
    "kodak_E100.pth"
}

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

MAX_WIDTH = 2457
MAX_HEIGHT = 1843

def resize_to_nearest_multiple_of_32(image):
    width, height = image.size
    
    if width > MAX_WIDTH and height > MAX_HEIGHT:
        gr.Error("Image too large!")
        return image
    
    # 计算最接近的可以被32整除的宽度和高度
    new_width = (width + 31) // 32 * 32  # 向上取整至最接近32的倍数
    new_height = (height + 31) // 32 * 32  # 向上取整至最接近32的倍数
    
    # 调整图像大小
    resized_image = image.resize((new_width, new_height), Image.BILINEAR)
    
    return resized_image

def load_model(model_path):
    if model_path[-4:] == ".pth":
        model = mynet.ResNetGenerator(3,3)
        model.load_state_dict(torch.load(model_path, map_location=device))
    else:
        model = mynet.FilmStyleTransfer()
        checkpoint = torch.load(model_path, map_location=device)
        model.load_state_dict(checkpoint['model_state_dict'])
    model = model.to(device)
    print("[INFO] Open model successfully!")
    return model

def process_images(image, model_choice):
    width, height = image.size

    if width > MAX_WIDTH and height > MAX_HEIGHT:
        raise gr.Error("Image too large!")

    image = resize_to_nearest_multiple_of_32(image)
    image = transform(image)
    print(os.path.join("models", model_choice))
    model = load_model(os.path.join("models", model_choice))
    model.eval()
    with torch.no_grad():
        image = image.unsqueeze(0)
        image = image.to(device)
        output, _ = model(image)
        output = output.squeeze().cpu().clamp(0, 1)
        output = transforms.ToPILImage()(output)
        return output
    
def main():
    with gr.Blocks(title="DigitalFilm App") as demo:
        image_input = gr.Image(type="pil", label="Upload Image(w <= 2457 & h <= 1843)")
        model_choice = gr.Dropdown(models, label="Select Model", allow_custom_value=False)
        image_output = gr.Image(type="pil", label="Generated Image")
        run_button = gr.Button("Run Model")
        run_button.click(process_images, inputs=[image_input, model_choice], outputs=image_output)

    demo.launch()

if __name__ == "__main__":
    main()