train.py

import os
import numpy as np
import cv2
import random
import torch
import torch.nn as nn
import time

import sys
sys.path.append("..")
sys.path.append("../..")
from util import mosaic,util,ffmpeg,filt,data
from util import image_processing as impro
from cores import Options
from models import pix2pix_model,video_model,unet_model,loadmodel
from matplotlib import pyplot as plt
import torch.backends.cudnn as cudnn

N = 25
ITER = 10000000
LR = 0.0002
beta1 = 0.5
use_gpu = True
use_gan = True
use_L2 = False
CONTINUE =  True
lambda_L1 = 100.0
lambda_gan = 1

SAVE_FRE = 10000
start_iter = 0
finesize = 256
loadsize = int(finesize*1.2)
batchsize = 1
perload_num = 16
savename = 'MosaicNet_instance_gan_256_D5'
dir_checkpoint = 'checkpoints/'+savename
util.makedirs(dir_checkpoint)

loss_sum = [0.,0.,0.,0.]
loss_plot = [[],[]]
item_plot = []

opt = Options().getparse()
videos = os.listdir('./dataset')
videos.sort()
lengths = []
print('check dataset...')
for video in videos:
    video_images = os.listdir('./dataset/'+video+'/ori')
    lengths.append(len(video_images))
#unet_128
#resnet_9blocks
#netG = pix2pix_model.define_G(3*N+1, 3, 128, 'resnet_6blocks', norm='instance',use_dropout=True, init_type='normal', gpu_ids=[])
netG = video_model.MosaicNet(3*N+1, 3)
loadmodel.show_paramsnumber(netG,'netG')
# netG = unet_model.UNet(3*N+1, 3)
if use_gan:
    #netD = pix2pix_model.define_D(3*2+1, 64, 'pixel', norm='instance')
    #netD = pix2pix_model.define_D(3*2+1, 64, 'basic', norm='instance')
    netD = pix2pix_model.define_D(3*2+1, 64, 'n_layers', n_layers_D=5, norm='instance')

if CONTINUE:
    if not os.path.isfile(os.path.join(dir_checkpoint,'last_G.pth')):
        CONTINUE = False
        print('can not load last_G, training on init weight.')
if CONTINUE:     
    netG.load_state_dict(torch.load(os.path.join(dir_checkpoint,'last_G.pth')))
    if use_gan:
        netD.load_state_dict(torch.load(os.path.join(dir_checkpoint,'last_D.pth')))
    f = open(os.path.join(dir_checkpoint,'iter'),'r')
    start_iter = int(f.read())
    f.close()
if use_gpu:
    netG.cuda()
    if use_gan:
        netD.cuda()
    cudnn.benchmark = True

optimizer_G = torch.optim.Adam(netG.parameters(), lr=LR,betas=(beta1, 0.999))
criterion_L1 = nn.L1Loss()
criterion_L2 = nn.MSELoss()
if use_gan:
    optimizer_D = torch.optim.Adam(netG.parameters(), lr=LR,betas=(beta1, 0.999))
    criterionGAN = pix2pix_model.GANLoss(gan_mode='lsgan').cuda()


def loaddata():
    video_index = random.randint(0,len(videos)-1)
    video = videos[video_index]
    img_index = random.randint(N,lengths[video_index]- N)
    input_img = np.zeros((loadsize,loadsize,3*N+1), dtype='uint8')
    for i in range(0,N):
    
        img = cv2.imread('./dataset/'+video+'/mosaic/output_'+'%05d'%(img_index+i-int(N/2))+'.png')
        img = impro.resize(img,loadsize)
        input_img[:,:,i*3:(i+1)*3] = img
    mask = cv2.imread('./dataset/'+video+'/mask/output_'+'%05d'%(img_index)+'.png',0)
    mask = impro.resize(mask,loadsize)
    mask = impro.mask_threshold(mask,15,128)
    input_img[:,:,-1] = mask

    ground_true = cv2.imread('./dataset/'+video+'/ori/output_'+'%05d'%(img_index)+'.png')
    ground_true = impro.resize(ground_true,loadsize)

    input_img,ground_true = data.random_transform_video(input_img,ground_true,finesize,N)
    input_img = data.im2tensor(input_img,bgr2rgb=False,use_gpu=opt.use_gpu,use_transform = False,is0_1=False)
    ground_true = data.im2tensor(ground_true,bgr2rgb=False,use_gpu=opt.use_gpu,use_transform = False,is0_1=False)
    
    return input_img,ground_true

print('preloading data, please wait 5s...')

if perload_num <= batchsize:
    perload_num = batchsize*2
input_imgs = torch.rand(perload_num,N*3+1,finesize,finesize).cuda()
ground_trues = torch.rand(perload_num,3,finesize,finesize).cuda()
load_cnt = 0

def preload():
    global load_cnt   
    while 1:
        try:
            ran = random.randint(0, perload_num-1)
            input_imgs[ran],ground_trues[ran] = loaddata()
            load_cnt += 1
            # time.sleep(0.1)
        except Exception as e:
            print("error:",e)

import threading
t = threading.Thread(target=preload,args=())  #t为新创建的线程
t.daemon = True
t.start()

time_start=time.time()
while load_cnt < perload_num:
    time.sleep(0.1)
time_end=time.time()
print('load speed:',round((time_end-time_start)/perload_num,3),'s/it')


util.copyfile('./train.py', os.path.join(dir_checkpoint,'train.py'))
util.copyfile('../../models/video_model.py', os.path.join(dir_checkpoint,'model.py'))
netG.train()
time_start=time.time()
print("Begin training...")
for iter in range(start_iter+1,ITER):

    ran = random.randint(0, perload_num-batchsize-1)
    inputdata = input_imgs[ran:ran+batchsize].clone()
    target = ground_trues[ran:ran+batchsize].clone()

    pred = netG(inputdata)

    if use_gan:
        netD.train()
        # print(inputdata[0,3*N,:,:].size())
        # print((inputdata[:,int((N-1)/2)*3:(int((N-1)/2)+1)*3,:,:]).size())
        real_A = torch.cat((inputdata[:,int((N-1)/2)*3:(int((N-1)/2)+1)*3,:,:], inputdata[:,-1,:,:].reshape(-1,1,finesize,finesize)), 1)
        fake_AB = torch.cat((real_A, pred), 1)
        pred_fake = netD(fake_AB.detach())
        loss_D_fake = criterionGAN(pred_fake, False)

        real_AB = torch.cat((real_A, target), 1)
        pred_real = netD(real_AB)
        loss_D_real = criterionGAN(pred_real, True)
        loss_D = (loss_D_fake + loss_D_real) * 0.5
        loss_sum[2] += loss_D_fake.item()
        loss_sum[3] += loss_D_real.item()

        optimizer_D.zero_grad()
        loss_D.backward()
        optimizer_D.step()
        netD.eval()

        # fake_AB = torch.cat((inputdata[:,int((N-1)/2)*3:(int((N-1)/2)+1)*3,:,:], pred), 1)
        real_A = torch.cat((inputdata[:,int((N-1)/2)*3:(int((N-1)/2)+1)*3,:,:], inputdata[:,-1,:,:].reshape(-1,1,finesize,finesize)), 1)
        fake_AB = torch.cat((real_A, pred), 1)
        pred_fake = netD(fake_AB)
        loss_G_GAN = criterionGAN(pred_fake, True)*lambda_gan
        # Second, G(A) = B
        if use_L2:
            loss_G_L1 = (criterion_L1(pred, target)+criterion_L2(pred, target)) * lambda_L1
        else:
            loss_G_L1 = criterion_L1(pred, target) * lambda_L1
        # combine loss and calculate gradients
        loss_G = loss_G_GAN + loss_G_L1
        loss_sum[0] += loss_G_L1.item()
        loss_sum[1] += loss_G_GAN.item()

        optimizer_G.zero_grad()
        loss_G.backward()
        optimizer_G.step()

    else:
        if use_L2:
            loss_G_L1 = (criterion_L1(pred, target)+criterion_L2(pred, target)) * lambda_L1
        else:
            loss_G_L1 = criterion_L1(pred, target) * lambda_L1
        loss_sum[0] += loss_G_L1.item()

        optimizer_G.zero_grad()
        loss_G_L1.backward()
        optimizer_G.step()


    if (iter+1)%100 == 0:
        try:
            data.showresult(inputdata[:,int((N-1)/2)*3:(int((N-1)/2)+1)*3,:,:],
             target, pred,os.path.join(dir_checkpoint,'result_train.png'))
        except Exception as e:
            print(e)
     
    if (iter+1)%1000 == 0:
        time_end = time.time()
        if use_gan:
            print('iter:',iter+1,' L1_loss:', round(loss_sum[0]/1000,4),' G_loss:', round(loss_sum[1]/1000,4),
                ' D_f:',round(loss_sum[2]/1000,4),' D_r:',round(loss_sum[3]/1000,4),' time:',round((time_end-time_start)/1000,2))
            if (iter+1)/1000 >= 10:
                loss_plot[0].append(loss_sum[0]/1000)
                loss_plot[1].append(loss_sum[1]/1000)
                item_plot.append(iter+1)
                try:
                    plt.plot(item_plot,loss_plot[0])
                    plt.plot(item_plot,loss_plot[1])
                    plt.savefig(os.path.join(dir_checkpoint,'loss.png'))
                    plt.close()
                except Exception as e:
                    print("error:",e)
        else:
            print('iter:',iter+1,' L1_loss:',round(loss_sum[0]/1000,4),' time:',round((time_end-time_start)/1000,2))
            if (iter+1)/1000 >= 10:
                loss_plot[0].append(loss_sum[0]/1000)
                item_plot.append(iter+1)
                try:
                    plt.plot(item_plot,loss_plot[0])
                    plt.savefig(os.path.join(dir_checkpoint,'loss.png'))
                    plt.close()
                except Exception as e:
                    print("error:",e)
        loss_sum = [0.,0.,0.,0.]
        time_start=time.time()


    if (iter+1)%SAVE_FRE == 0:
        if iter+1 != SAVE_FRE:
            os.rename(os.path.join(dir_checkpoint,'last_G.pth'),os.path.join(dir_checkpoint,str(iter+1-SAVE_FRE)+'G.pth'))
        torch.save(netG.cpu().state_dict(),os.path.join(dir_checkpoint,'last_G.pth'))
        if use_gan:
            if iter+1 != SAVE_FRE:
                os.rename(os.path.join(dir_checkpoint,'last_D.pth'),os.path.join(dir_checkpoint,str(iter+1-SAVE_FRE)+'D.pth'))
            torch.save(netD.cpu().state_dict(),os.path.join(dir_checkpoint,'last_D.pth'))
        if use_gpu:
            netG.cuda()
            if use_gan:
                netD.cuda()
        f = open(os.path.join(dir_checkpoint,'iter'),'w+')
        f.write(str(iter+1))
        f.close()
        print('network saved.')

        #test
        netG.eval()
        
        test_names = os.listdir('./test')
        test_names.sort()
        result = np.zeros((finesize*2,finesize*len(test_names),3), dtype='uint8')

        for cnt,test_name in enumerate(test_names,0):
            img_names = os.listdir(os.path.join('./test',test_name,'image'))
            img_names.sort()
            inputdata = np.zeros((finesize,finesize,3*N+1), dtype='uint8')
            for i in range(0,N):
                img = impro.imread(os.path.join('./test',test_name,'image',img_names[i]))
                img = impro.resize(img,finesize)
                inputdata[:,:,i*3:(i+1)*3] = img

            mask = impro.imread(os.path.join('./test',test_name,'mask.png'),'gray')
            mask = impro.resize(mask,finesize)
            mask = impro.mask_threshold(mask,15,128)
            inputdata[:,:,-1] = mask
            result[0:finesize,finesize*cnt:finesize*(cnt+1),:] = inputdata[:,:,int((N-1)/2)*3:(int((N-1)/2)+1)*3]
            inputdata = data.im2tensor(inputdata,bgr2rgb=False,use_gpu=opt.use_gpu,use_transform = False,is0_1 = False)
            pred = netG(inputdata)
 
            pred = data.tensor2im(pred,rgb2bgr = False, is0_1 = False)
            result[finesize:finesize*2,finesize*cnt:finesize*(cnt+1),:] = pred

        cv2.imwrite(os.path.join(dir_checkpoint,str(iter+1)+'_test.png'), result)
        netG.train()