train_semi_UAMT.py

from torchvision import transforms, datasets
import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim import lr_scheduler
from torch.autograd import Variable
from torch.utils.data import DataLoader
from models.getnetwork import get_network
import argparse
import time
import os
import numpy as np
from torch.backends import cudnn
import random
from PIL import Image
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel
import sys

from config.dataset_config.dataset_cfg import dataset_cfg
from config.augmentation.online_aug import data_transform_2d, data_normalize_2d
from loss.loss_function import segmentation_loss, softmax_mse_loss
from models.getnetwork import get_network
from dataload.dataset_2d import imagefloder_itn
from config.visdom_config.visual_visdom import visdom_initialization_MT, visualization_MT, visual_image_MT
from config.warmup_config.warmup import GradualWarmupScheduler
from config.train_test_config.train_test_config import print_train_loss_MT, print_val_loss, print_train_eval_sup, print_val_eval, save_val_best_2d, draw_pred_MT, print_best
from warnings import simplefilter
from config.ramps import ramps

simplefilter(action='ignore', category=FutureWarning)


def update_ema_variables(model, ema_model, alpha, global_step):
    # Use the true average until the exponential average is more correct
    alpha = min(1 - 1 / (global_step + 1), alpha)
    for ema_param, param in zip(ema_model.parameters(), model.parameters()):
        ema_param.data.mul_(alpha).add_(param.data, alpha=1 - alpha)


def init_seeds(seed):
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    random.seed(seed)
    np.random.seed(seed)
    os.environ['PYTHONHASHSEED'] = str(0)
    torch.backends.cudnn.benchmark = False
    torch.backends.cudnn.deterministic = True


if __name__ == '__main__':

    parser = argparse.ArgumentParser()
    parser.add_argument('--path_trained_models', default='/mnt/data1/XNet/checkpoints/semi')
    parser.add_argument('--path_seg_results', default='/mnt/data1/XNet/seg_pred/semi')
    parser.add_argument('-pd', '--path_dataset', default='/mnt/data1/XNet/dataset/CREMI')
    parser.add_argument('--dataset_name', default='CREMI', help='CREMI, ISIC-2017, GlaS')
    parser.add_argument('--input1', default='image')
    parser.add_argument('--sup_mark', default='20')
    parser.add_argument('--unsup_mark', default='80')
    parser.add_argument('-b', '--batch_size', default=2, type=int)
    parser.add_argument('-e', '--num_epochs', default=200, type=int)
    parser.add_argument('-s', '--step_size', default=50, type=int)
    parser.add_argument('-l', '--lr', default=0.5, type=float)
    parser.add_argument('-g', '--gamma', default=0.5, type=float)
    parser.add_argument('-u', '--unsup_weight', default=0.05, type=float)
    parser.add_argument('--loss', default='dice')
    parser.add_argument('-w', '--warm_up_duration', default=20)
    parser.add_argument('--momentum', default=0.9, type=float)
    parser.add_argument('--ema_decay', default=0.99, type=float)
    parser.add_argument('--wd', default=-5, type=float, help='weight decay pow')

    parser.add_argument('-i', '--display_iter', default=5, type=int)
    parser.add_argument('-n', '--network', default='unet', type=str)
    parser.add_argument('--local_rank', default=-1, type=int)
    parser.add_argument('--rank_index', default=0, help='0, 1, 2, 3')
    parser.add_argument('-v', '--vis', default=True, help='need visualization or not')
    parser.add_argument('--visdom_port', default=16672)
    args = parser.parse_args()

    torch.cuda.set_device(args.local_rank)
    dist.init_process_group(backend='nccl', init_method='env://')

    rank = torch.distributed.get_rank()
    ngpus_per_node = torch.cuda.device_count()
    init_seeds(rank + 1)

    dataset_name = args.dataset_name
    cfg = dataset_cfg(dataset_name)

    print_num = 77 + (cfg['NUM_CLASSES'] - 3) * 14
    print_num_minus = print_num - 2
    print_num_half = int(print_num / 2 - 1)

    # trained model save
    path_trained_models = args.path_trained_models + '/' + str(os.path.split(args.path_dataset)[1])
    if not os.path.exists(path_trained_models) and rank == args.rank_index:
        os.mkdir(path_trained_models)
    path_trained_models = path_trained_models + '/' + 'UAMT' + '-l=' + str(args.lr) + '-e=' + str(args.num_epochs) + '-s=' + str(args.step_size) + '-g=' + str(args.gamma) + '-b=' + str(args.batch_size) + '-cw' + str(args.unsup_weight) + '-w=' + str(args.warm_up_duration) + '-' + str(args.sup_mark) + '-' + str(args.unsup_mark) + '-' + str(args.input1)
    if not os.path.exists(path_trained_models) and rank == args.rank_index:
        os.mkdir(path_trained_models)

    # seg results save
    path_seg_results = args.path_seg_results + '/' + str(os.path.split(args.path_dataset)[1])
    if not os.path.exists(path_seg_results) and rank == args.rank_index:
        os.mkdir(path_seg_results)
    path_seg_results = path_seg_results + '/' + 'UAMT' + '-l=' + str(args.lr) + '-e=' + str(args.num_epochs) + '-s=' + str(args.step_size) + '-g=' + str(args.gamma) + '-b=' + str(args.batch_size) + '-cw=' + str(args.unsup_weight) + '-w=' + str(args.warm_up_duration) + '-' + str(args.sup_mark) + '-' + str(args.unsup_mark) + '-' + str(args.input1)
    if not os.path.exists(path_seg_results) and rank == args.rank_index:
        os.mkdir(path_seg_results)

    # vis
    if args.vis and rank == args.rank_index:
        visdom_env = str('Semi-UAMT-' + str(os.path.split(args.path_dataset)[1]) + '-' + args.network + '-l=' + str(args.lr) + '-e=' + str(args.num_epochs) + '-s=' + str(args.step_size) + '-g=' + str(args.gamma) + '-b=' + str(args.batch_size) + '-cw=' + str(args.unsup_weight) + '-w=' + str(args.warm_up_duration) + '-' + str(args.sup_mark) + '-' + str(args.unsup_mark) + '-' + str(args.input1))
        visdom = visdom_initialization_MT(env=visdom_env, port=args.visdom_port)

    if args.input1 == 'image':
        input1_mean = 'MEAN'
        input1_std = 'STD'
    else:
        input1_mean = 'MEAN_' + args.input1
        input1_std = 'STD_' + args.input1

    data_transforms = data_transform_2d()
    data_normalize = data_normalize_2d(cfg[input1_mean], cfg[input1_std])

    dataset_train_unsup = imagefloder_itn(
        data_dir=args.path_dataset + '/train_unsup_' + args.unsup_mark,
        input1=args.input1,
        data_transform_1=data_transforms['train'],
        data_normalize_1=data_normalize,
        sup=False,
        num_images=None,
    )
    num_images_unsup = len(dataset_train_unsup)

    dataset_train_sup = imagefloder_itn(
        data_dir=args.path_dataset + '/train_sup_' + args.sup_mark,
        input1=args.input1,
        data_transform_1=data_transforms['train'],
        data_normalize_1=data_normalize,
        sup=True,
        num_images=num_images_unsup,
    )
    dataset_val = imagefloder_itn(
        data_dir=args.path_dataset + '/val',
        input1=args.input1,
        data_transform_1=data_transforms['val'],
        data_normalize_1=data_normalize,
        sup=True,
        num_images=None,
    )

    train_sampler_sup = torch.utils.data.distributed.DistributedSampler(dataset_train_sup, shuffle=True)
    train_sampler_unsup = torch.utils.data.distributed.DistributedSampler(dataset_train_unsup, shuffle=True)
    val_sampler = torch.utils.data.distributed.DistributedSampler(dataset_val, shuffle=False)

    dataloaders = dict()
    dataloaders['train_sup'] = DataLoader(dataset_train_sup, batch_size=args.batch_size, shuffle=False, pin_memory=True, num_workers=8, sampler=train_sampler_sup)
    dataloaders['train_unsup'] = DataLoader(dataset_train_unsup, batch_size=args.batch_size, shuffle=False, pin_memory=True, num_workers=8, sampler=train_sampler_unsup)
    dataloaders['val'] = DataLoader(dataset_val, batch_size=args.batch_size, shuffle=False, pin_memory=True, num_workers=8, sampler=val_sampler)

    num_batches = {'train_sup': len(dataloaders['train_sup']), 'train_unsup': len(dataloaders['train_unsup']), 'val': len(dataloaders['val'])}

    model1 = get_network(args.network, cfg['IN_CHANNELS'], cfg['NUM_CLASSES'])
    model2 = get_network(args.network, cfg['IN_CHANNELS'], cfg['NUM_CLASSES'])

    model1 = model1.cuda()
    model2 = model2.cuda()
    # for param in model2.parameters():
    #     param.detach_()
    model1 = DistributedDataParallel(model1, device_ids=[args.local_rank])
    model2 = DistributedDataParallel(model2, device_ids=[args.local_rank])
    dist.barrier()

    criterion = segmentation_loss(args.loss, False).cuda()

    optimizer1 = optim.SGD(model1.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=5 * 10 ** args.wd)
    exp_lr_scheduler1 = lr_scheduler.StepLR(optimizer1, step_size=args.step_size, gamma=args.gamma)
    scheduler_warmup1 = GradualWarmupScheduler(optimizer1, multiplier=1.0, total_epoch=args.warm_up_duration, after_scheduler=exp_lr_scheduler1)

    since = time.time()
    count_iter = 0

    best_model = model1
    best_result = 'Result1'
    best_val_eval_list = [0 for i in range(4)]

    for epoch in range(args.num_epochs):

        count_iter += 1
        if (count_iter - 1) % args.display_iter == 0:
            begin_time = time.time()

        dataloaders['train_sup'].sampler.set_epoch(epoch)
        dataloaders['train_unsup'].sampler.set_epoch(epoch)
        model1.train()
        model2.train()

        train_loss_sup_1 = 0.0
        train_loss_unsup = 0.0
        train_loss = 0.0

        val_loss_sup_1 = 0.0
        val_loss_sup_2 = 0.0

        unsup_weight = args.unsup_weight * (epoch + 1) / args.num_epochs
        dist.barrier()

        dataset_train_sup = iter(dataloaders['train_sup'])
        dataset_train_unsup = iter(dataloaders['train_unsup'])

        for i in range(num_batches['train_sup']):

            unsup_index = next(dataset_train_unsup)
            img_train_unsup_1 = unsup_index['image']
            img_train_unsup_1 = Variable(img_train_unsup_1.cuda(non_blocking=True))

            noise = torch.clamp(torch.randn_like(img_train_unsup_1) * 0.1, -0.2, 0.2)
            img_train_unsup_2 = img_train_unsup_1 + noise

            optimizer1.zero_grad()

            pred_train_unsup1 = model1(img_train_unsup_1)
            with torch.no_grad():
                pred_train_unsup2 = model2(img_train_unsup_2)

            T = 8

            _, _, w, h = img_train_unsup_1.shape
            volume_batch_r = img_train_unsup_1.repeat(2, 1, 1, 1)
            stride = volume_batch_r.shape[0] // 2
            preds = torch.zeros([stride * T, cfg['NUM_CLASSES'], w, h]).cuda()
            for i_ in range(T // 2):
                ema_inputs = volume_batch_r + torch.clamp(torch.randn_like(volume_batch_r) * 0.1, -0.2, 0.2)
                with torch.no_grad():
                    preds[2 * stride * i_:2 * stride * (i_ + 1)] = model2(ema_inputs)
            preds = torch.softmax(preds, dim=1)
            preds = preds.reshape(T, stride, cfg['NUM_CLASSES'], w, h)
            preds = torch.mean(preds, dim=0)
            uncertainty = -1.0 * torch.sum(preds * torch.log(preds + 1e-6), dim=1, keepdim=True)

            consistency_dist = softmax_mse_loss(pred_train_unsup1, pred_train_unsup2)  # (batch, 2, 112,112,80)
            threshold = (0.75 + 0.25 * ramps.sigmoid_rampup(epoch, args.num_epochs)) * np.log(2)
            mask = (uncertainty < threshold).float()
            loss_train_unsup = torch.sum(mask * consistency_dist) / (2 * torch.sum(mask) + 1e-16)

            loss_train_unsup = loss_train_unsup * unsup_weight
            loss_train_unsup.backward(retain_graph=True)
            torch.cuda.empty_cache()

            sup_index = next(dataset_train_sup)
            img_train_sup = sup_index['image']
            img_train_sup = Variable(img_train_sup.cuda(non_blocking=True))
            mask_train_sup = sup_index['mask']
            mask_train_sup = Variable(mask_train_sup.cuda(non_blocking=True))

            pred_train_sup1 = model1(img_train_sup)

            if count_iter % args.display_iter == 0:
                if i == 0:
                    score_list_train1 = pred_train_sup1
                    mask_list_train = mask_train_sup
                # else:
                elif 0 < i <= num_batches['train_sup'] / 32:
                    score_list_train1 = torch.cat((score_list_train1, pred_train_sup1), dim=0)
                    mask_list_train = torch.cat((mask_list_train, mask_train_sup), dim=0)

            loss_train_sup1 = criterion(pred_train_sup1, mask_train_sup)

            loss_train_sup = loss_train_sup1
            loss_train_sup.backward()

            optimizer1.step()
            update_ema_variables(model1, model2, args.ema_decay, epoch)
            torch.cuda.empty_cache()

            loss_train = loss_train_unsup + loss_train_sup
            train_loss_unsup += loss_train_unsup.item()
            train_loss_sup_1 += loss_train_sup1.item()
            train_loss += loss_train.item()

        scheduler_warmup1.step()
        torch.cuda.empty_cache()

        if count_iter % args.display_iter == 0:

            score_gather_list_train1 = [torch.zeros_like(score_list_train1) for _ in range(ngpus_per_node)]
            torch.distributed.all_gather(score_gather_list_train1, score_list_train1)
            score_list_train1 = torch.cat(score_gather_list_train1, dim=0)

            mask_gather_list_train = [torch.zeros_like(mask_list_train) for _ in range(ngpus_per_node)]
            torch.distributed.all_gather(mask_gather_list_train, mask_list_train)
            mask_list_train = torch.cat(mask_gather_list_train, dim=0)

            if rank == args.rank_index:
                torch.cuda.empty_cache()
                print('=' * print_num)
                print('| Epoch {}/{}'.format(epoch + 1, args.num_epochs).ljust(print_num_minus, ' '), '|')
                train_epoch_loss_sup1, train_epoch_loss_cps, train_epoch_loss = print_train_loss_MT(train_loss_sup_1, train_loss_unsup, train_loss, num_batches, print_num, print_num_half, print_num_minus)
                train_eval_list1, train_m_jc1 = print_train_eval_sup(cfg['NUM_CLASSES'], score_list_train1, mask_list_train, print_num_minus)
                torch.cuda.empty_cache()

            with torch.no_grad():
                model1.eval()
                model2.eval()

                for i, data in enumerate(dataloaders['val']):

                    # if 0 <= i <= num_batches['val'] / 16:

                    inputs_val = Variable(data['image'].cuda(non_blocking=True))
                    mask_val = Variable(data['mask'].cuda(non_blocking=True))
                    name_val = data['ID']

                    optimizer1.zero_grad()

                    outputs_val1 = model1(inputs_val)
                    outputs_val2 = model2(inputs_val)
                    torch.cuda.empty_cache()

                    if i == 0:
                        score_list_val1 = outputs_val1
                        score_list_val2 = outputs_val2
                        mask_list_val = mask_val
                        name_list_val = name_val
                    else:
                        score_list_val1 = torch.cat((score_list_val1, outputs_val1), dim=0)
                        score_list_val2 = torch.cat((score_list_val2, outputs_val2), dim=0)
                        mask_list_val = torch.cat((mask_list_val, mask_val), dim=0)
                        name_list_val = np.append(name_list_val, name_val, axis=0)

                    loss_val_sup1 = criterion(outputs_val1, mask_val)
                    loss_val_sup2 = criterion(outputs_val2, mask_val)

                    val_loss_sup_1 += loss_val_sup1.item()
                    val_loss_sup_2 += loss_val_sup2.item()

                torch.cuda.empty_cache()
                score_gather_list_val1 = [torch.zeros_like(score_list_val1) for _ in range(ngpus_per_node)]
                torch.distributed.all_gather(score_gather_list_val1, score_list_val1)
                score_list_val1 = torch.cat(score_gather_list_val1, dim=0)

                score_gather_list_val2 = [torch.zeros_like(score_list_val2) for _ in range(ngpus_per_node)]
                torch.distributed.all_gather(score_gather_list_val2, score_list_val2)
                score_list_val2 = torch.cat(score_gather_list_val2, dim=0)

                mask_gather_list_val = [torch.zeros_like(mask_list_val) for _ in range(ngpus_per_node)]
                torch.distributed.all_gather(mask_gather_list_val, mask_list_val)
                mask_list_val = torch.cat(mask_gather_list_val, dim=0)

                name_gather_list_val = [None for _ in range(ngpus_per_node)]
                torch.distributed.all_gather_object(name_gather_list_val, name_list_val)
                name_list_val = np.concatenate(name_gather_list_val, axis=0)

                if rank == args.rank_index:
                    val_epoch_loss_sup1, val_epoch_loss_sup2 = print_val_loss(val_loss_sup_1, val_loss_sup_2, num_batches, print_num, print_num_half)
                    val_eval_list1, val_eval_list2, val_m_jc1, val_m_jc2 = print_val_eval(cfg['NUM_CLASSES'], score_list_val1, score_list_val2, mask_list_val, print_num_half)
                    best_val_eval_list, best_model, best_result = save_val_best_2d(cfg['NUM_CLASSES'], best_model, best_val_eval_list, best_result, model1, model2, score_list_val1, score_list_val2, name_list_val, val_eval_list1, val_eval_list2, path_trained_models, path_seg_results, cfg['PALETTE'])
                    torch.cuda.empty_cache()

                    if args.vis:
                        draw_img = draw_pred_MT(cfg['NUM_CLASSES'], mask_train_sup, mask_val, pred_train_sup1, outputs_val1, outputs_val2, train_eval_list1, val_eval_list1, val_eval_list2)
                        visualization_MT(visdom, epoch + 1, train_epoch_loss, train_epoch_loss_sup1, train_epoch_loss_cps, train_m_jc1, val_epoch_loss_sup1, val_epoch_loss_sup2, val_m_jc1, val_m_jc2)
                        visual_image_MT(visdom, draw_img[0], draw_img[1], draw_img[2], draw_img[3], draw_img[4])

                    print('-' * print_num)
                    print('| Epoch Time: {:.4f}s'.format((time.time() - begin_time) / args.display_iter).ljust(print_num_minus, ' '), '|')
            torch.cuda.empty_cache()
        torch.cuda.empty_cache()

    if rank == args.rank_index:
        time_elapsed = time.time() - since
        m, s = divmod(time_elapsed, 60)
        h, m = divmod(m, 60)

        print('=' * print_num)
        print('| Training Completed In {:.0f}h {:.0f}mins {:.0f}s'.format(h, m, s).ljust(print_num_minus, ' '), '|')
        print('-' * print_num)
        print_best(cfg['NUM_CLASSES'], best_val_eval_list, best_model, best_result, path_trained_models, print_num_minus)
        print('=' * print_num)