app.py

import gradio as gr
import torch
import pickle
import time
from options.test_options import TestOptions
from data.data_loader_test import CreateDataLoader
from models.networks import ResUnetGenerator, load_checkpoint
from models.afwm import AFWM
import torch.nn as nn
import os
import numpy as np
import torch
import cv2
import torch.nn.functional as F
from torchvision import utils
from util import flow_util

def de_offset(s_grid):
    [b,_,h,w] = s_grid.size()


    x = torch.arange(w).view(1, -1).expand(h, -1).float()
    y = torch.arange(h).view(-1, 1).expand(-1, w).float()
    x = 2*x/(w-1)-1
    y = 2*y/(h-1)-1
    grid = torch.stack([x,y], dim=0).float().cuda()
    grid = grid.unsqueeze(0).expand(b, -1, -1, -1)

    offset = grid - s_grid

    offset_x = offset[:,0,:,:] * (w-1) / 2
    offset_y = offset[:,1,:,:] * (h-1) / 2

    offset = torch.cat((offset_y,offset_x),0)
    
    return  offset

def tryon(person,cloth,edge):

    #save images in folders
    cv2.imwrite('./data/test_ma_img/000001_0.jpg', person)
    cv2.imwrite('./data/test_edge/000001_1.jpg', edge)
    cv2.imwrite('./data/test_clothes/000001_1.jpg', cloth)

    with open('opt.pkl', 'rb') as handle:
        opt = pickle.load(handle)  
    
    f2c = flow_util.flow2color()
    start_epoch, epoch_iter = 1, 0

    data_loader = CreateDataLoader(opt)
    dataset = data_loader.load_data()
    dataset_size = len(data_loader)     #must be 1
    print(dataset_size)

    warp_model = AFWM(opt, 3)
    print(warp_model)
    warp_model.eval()
    warp_model.cuda()
    load_checkpoint(warp_model, opt.warp_checkpoint)

    gen_model = ResUnetGenerator(7, 4, 5, ngf=64, norm_layer=nn.BatchNorm2d)
    gen_model.eval()
    gen_model.cuda()
    load_checkpoint(gen_model, opt.gen_checkpoint)

    total_steps = (start_epoch-1) * dataset_size + epoch_iter
    step = 0
    step_per_batch = dataset_size / opt.batchSize
    
    if not os.path.exists('our_t_results'):
        os.mkdir('our_t_results')

    for epoch in range(1,2):

        for i, data in enumerate(dataset, start=epoch_iter):
            iter_start_time = time.time()
            total_steps += opt.batchSize
            epoch_iter += opt.batchSize

            real_image = data['image']
            clothes = data['clothes']
            ##edge is extracted from the clothes image with the built-in function in python
            edge = data['edge']
            edge = torch.FloatTensor((edge.detach().numpy() > 0.5).astype(np.int64))
            clothes = clothes * edge

            flow_out = warp_model(real_image.cuda(), clothes.cuda())
            warped_cloth, last_flow, = flow_out
            warped_edge = F.grid_sample(edge.cuda(), last_flow.permute(0, 2, 3, 1),
                            mode='bilinear', padding_mode='zeros')

            gen_inputs = torch.cat([real_image.cuda(), warped_cloth, warped_edge], 1)
            gen_outputs = gen_model(gen_inputs)
            p_rendered, m_composite = torch.split(gen_outputs, [3, 1], 1)
            p_rendered = torch.tanh(p_rendered)
            m_composite = torch.sigmoid(m_composite)
            m_composite = m_composite * warped_edge
            p_tryon = warped_cloth * m_composite + p_rendered * (1 - m_composite)

            path = 'results/' + opt.name
            os.makedirs(path, exist_ok=True)
            #sub_path = path + '/PFAFN'
            #os.makedirs(sub_path,exist_ok=True)
            print(data['p_name'])

            if step % 1 == 0:
                
                ## save try-on image only

                utils.save_image(
                    p_tryon,
                    os.path.join('./our_t_results', data['p_name'][0]),
                    nrow=int(1),
                    normalize=True,
                    value_range=(-1,1),
                )
                
                ## save person image, garment, flow, warped garment, and try-on image
                
                #a = real_image.float().cuda()
                #b = clothes.cuda()
                #flow_offset = de_offset(last_flow)
                #flow_color = f2c(flow_offset).cuda()
                #c= warped_cloth.cuda()
                #d = p_tryon
                #combine = torch.cat([a[0],b[0], flow_color, c[0], d[0]], 2).squeeze()
                #utils.save_image(
                #    combine,
                #    os.path.join('./im_gar_flow_wg', data['p_name'][0]),
                #    nrow=int(1),
                #    normalize=True,
                #    range=(-1,1),
                #)
                

            step += 1
            if epoch_iter >= dataset_size:
                break
    
    result_img = cv2.imread('./our_t_results/000001_0.jpg')
    return result_img

demo = gr.Interface(fn=tryon, 
             inputs=[gr.inputs.Image(label="Person"),gr.inputs.Image(label="Cloth"),gr.inputs.Image(label="Edge")],
             outputs="image"
             )

# def pp(inp1,inp2):
#     return inp1+" hello "+inp2

# demo2 = gr.Interface(fn=pp, 
#              inputs=[gr.inputs.Textbox(lines=5, label="Input Text"),gr.inputs.Textbox(lines=5, label="Input Text2")],
#              outputs=gr.outputs.Textbox(label="Generated Text"),
            #  )
             
demo.launch()