app.py

from options.test_options import TestOptions
from models import create_model
import torch
import numpy as np
import gradio as gr
from einops import rearrange
import torchvision
import torchvision.transforms as transforms

def tensor2im(input_image, imtype=np.uint8):
    if not isinstance(input_image, np.ndarray):
        if isinstance(input_image, torch.Tensor):  # get the data from a variable
            image_tensor = input_image.data
        else:
            return input_image
        image_numpy = image_tensor[0].cpu().float().numpy()  # convert it into a numpy array
        if image_numpy.shape[0] == 1:  # grayscale to RGB
            image_numpy = np.tile(image_numpy, (3, 1, 1))
        image_numpy = (np.transpose(image_numpy, (1, 2, 0)) + 1) / 2.0 * 255.0  # post-processing: tranpose and scaling
    else:  # if it is a numpy array, do nothing
        image_numpy = input_image
    return image_numpy.astype(imtype)

def get_model(translation):
    if translation == 'Map to Satellite':
        return 'map2sat'
    elif translation == 'Image to Van Gogh':
        return 'style_vangogh'
    elif translation == 'Image to Monet':
        return 'style_monet'

def unpaired_img2img(translation, image):
    opt = TestOptions().parse()
    m_name = get_model(translation)
    opt.name = m_name + '_pretrained'
    opt.model = 'test'
    opt.no_dropout = True
    opt.num_threads = 0
    opt.batch_size = 1
    opt.no_flip = True
    model = create_model(opt)
    model.setup(opt)
    model.eval()

    normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
    image = torch.from_numpy(image) # Convert image from numpy to PyTorch tensor
    image = rearrange(image, "h w c -> c h w") # Since PyTorch is channel first
    
    # Perform necessary image transforms
    image = transforms.Resize(256)(image)
    image = transforms.CenterCrop(256)(image).float()/255.
    image = normalize(image)
    
    image = rearrange(image, "c h w -> 1 c h w") # Insert batch size of 1 (as required by our model)

    model.set_input(image)
    model.test()
    visuals = model.get_current_visuals()  # get image results
    for i in visuals.values():
        im_data = i
    im = tensor2im(im_data)
    return im

gr.Interface(fn=unpaired_img2img,
            inputs=[gr.inputs.Dropdown(['Map to Satellite', 'Image to Van Gogh', 'Image to Monet']), 
                    gr.inputs.Image(shape=(256,256))],
            outputs=gr.outputs.Image(type="numpy"),
            title="Unpaired Image to Image Translation",
            examples=[['Map to Satellite',"examples/map2.jfif"],
            ['Image to Van Gogh', "examples/img2.jpg"],
            ['Image to Monet', "examples/img1.jpg"]],
            description="<p align='justify'>This is an implementation of the unpaired image to image translation using a pretrained CycleGAN model. To use the app, kindly select first the type of translation you wish to perform among the choices in the dropdown menu. Then, upload the image you wish to translate and click on the 'Submit' button.</p>",
            article="<p align='justify'>The model architecture used in this space is the Cycle-Consistent Adversarial Network, commonly referred to as CycleGAN. CycleGAN aims to perform translation of images between two domains without the need for expensive and difficult-to-acquire paired training data. The architecture consists of two generators, one generates an image from domain X to domain Y while the other generates an image from domain Y back to domain X. These two generators are also paired with a discriminator each that aims to discriminate generated images from real images, thus improving model performance. All credits go to Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros from the Berkeley AI Research (BAIR) laboratory at UC Berkeley for the creation of CycleGAN. To know more about Unpaired Image to Image Translation and CycleGAN, you may access their <a href = https://paperswithcode.com/paper/unpaired-image-to-image-translation-using>Papers with Code</a> page and their <a href = https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix>GitHub</a> repository.</p>",
            allow_flagging="never").launch(inbrowser=True)