main_ce_ecg.py

from __future__ import print_function

import os
import sys
import argparse
import time
import math

import tensorboard_logger as tb_logger
from tensorboardX import SummaryWriter
import torch
import torch.backends.cudnn as cudnn
from torchvision import transforms, datasets


from util import AverageMeter
from util import adjust_learning_rate, warmup_learning_rate, accuracy, calculate_auc, calculate_other_metrics
from util import set_optimizer, save_model
from util import EarlyStopping
from networks.resnet_big import SupCEResNet
from networks.CNN import CNN
from networks.CLOCSNET import cnn_network_contrastive, second_cnn_network
from datasets import Chapman

import numpy as np

try:
    import apex
    from apex import amp, optimizers
except ImportError:
    pass



os.environ["CUDA_VISIBLE_DEVICES"] = '2, 3'

def parse_option():
    parser = argparse.ArgumentParser('argument for training')

    parser.add_argument('--print_freq', type=int, default=10,
                        help='print frequency')
    parser.add_argument('--save_freq', type=int, default=50,
                        help='save frequency')
    parser.add_argument('--batch_size', type=int, default=256,
                        help='batch_size')
    parser.add_argument('--num_workers', type=int, default=16,
                        help='num of workers to use')
    parser.add_argument('--epochs', type=int, default=500,
                        help='number of training epochs')

    # optimization
    parser.add_argument('--learning_rate', type=float, default=0.2,
                        help='learning rate')
    parser.add_argument('--lr_decay_epochs', type=str, default='350,400,450',
                        help='where to decay lr, can be a list')
    parser.add_argument('--lr_decay_rate', type=float, default=0.1,
                        help='decay rate for learning rate')
    parser.add_argument('--weight_decay', type=float, default=1e-4,
                        help='weight decay')
    parser.add_argument('--momentum', type=float, default=0.9,
                        help='momentum')

    # model dataset
    parser.add_argument('--model', type=str, default='resnet50')
    parser.add_argument('--dataset', type=str, default='cifar10',
                        choices=['cifar10', 'cifar100', 'chapman'], help='dataset')

    # method
    parser.add_argument('--method', type=str, default='CE',
                        choices=['SupCon', 'SimCLR', 'CMSC', 'CMSC-P', 'CE'], help='choose method')

    # leads of data
    parser.add_argument('--lead', type=int, default=4, help='choose method')

    # other setting
    parser.add_argument('--cosine', action='store_true',
                        help='using cosine annealing')
    parser.add_argument('--syncBN', action='store_true',
                        help='using synchronized batch normalization')
    parser.add_argument('--warm', action='store_true',
                        help='warm-up for large batch training')
    parser.add_argument('--trial', type=str, default='0',
                        help='id for recording multiple runs')

    opt = parser.parse_args()

    # set the path according to the environment
    opt.data_folder = './datasets/'
    opt.model_path = './save/SupCon/{}_models'.format(opt.dataset)
    opt.tb_path = './save/SupCon/{}_tensorboard'.format(opt.dataset)

    iterations = opt.lr_decay_epochs.split(',')
    opt.lr_decay_epochs = list([])
    for it in iterations:
        opt.lr_decay_epochs.append(int(it))

    opt.model_name = 'SupCE_{}_{}_lr_{}_decay_{}_bsz_{}_trial_{}'.\
        format(opt.dataset, opt.model, opt.learning_rate, opt.weight_decay,
               opt.batch_size, opt.trial)

    if opt.cosine:
        opt.model_name = '{}_cosine'.format(opt.model_name)

    # warm-up for large-batch training,
    if opt.batch_size > 256:
        opt.warm = True
    if opt.warm:
        opt.model_name = '{}_warm'.format(opt.model_name)
        opt.warmup_from = 0.01
        opt.warm_epochs = 10
        if opt.cosine:
            eta_min = opt.learning_rate * (opt.lr_decay_rate ** 3)
            opt.warmup_to = eta_min + (opt.learning_rate - eta_min) * (
                    1 + math.cos(math.pi * opt.warm_epochs / opt.epochs)) / 2
        else:
            opt.warmup_to = opt.learning_rate

    opt.tb_folder = os.path.join(opt.tb_path, opt.model_name)
    if not os.path.isdir(opt.tb_folder):
        os.makedirs(opt.tb_folder)

    opt.save_folder = os.path.join(opt.model_path, opt.model_name)
    if not os.path.isdir(opt.save_folder):
        os.makedirs(opt.save_folder)

    if opt.dataset == 'cifar10':
        opt.n_cls = 10
    elif opt.dataset == 'cifar100':
        opt.n_cls = 100
    elif opt.dataset == 'chapman':
        opt.n_cls = 4
    else:
        raise ValueError('dataset not supported: {}'.format(opt.dataset))

    return opt


def set_loader(opt):
    # construct data loader
    if opt.dataset == 'cifar10':
        mean = (0.4914, 0.4822, 0.4465)
        std = (0.2023, 0.1994, 0.2010)
    elif opt.dataset == 'cifar100':
        mean = (0.5071, 0.4867, 0.4408)
        std = (0.2675, 0.2565, 0.2761)
    elif opt.dataset == 'chapman':
        pass
    else:
        raise ValueError('dataset not supported: {}'.format(opt.dataset))
    #normalize = transforms.Normalize(mean=mean, std=std)

    train_transform = transforms.Compose([
        #transforms.RandomResizedCrop(size=32, scale=(0.2, 1.)),
        #transforms.RandomHorizontalFlip(),
        transforms.ToTensor(),
     #   normalize,
    ])

    val_transform = transforms.Compose([
        transforms.ToTensor(),
      #  normalize,
    ])

    if opt.dataset == 'cifar10':
        train_dataset = datasets.CIFAR10(root=opt.data_folder,
                                         transform=train_transform,
                                         download=True)
        val_dataset = datasets.CIFAR10(root=opt.data_folder,
                                       train=False,
                                       transform=val_transform)
    elif opt.dataset == 'cifar100':
        train_dataset = datasets.CIFAR100(root=opt.data_folder,
                                          transform=train_transform,
                                          download=True)
        val_dataset = datasets.CIFAR100(root=opt.data_folder,
                                        train=False,
                                        transform=val_transform)
    elif opt.dataset == 'chapman':
        train_dataset = Chapman(train=True,
                                opt=opt,
                                #transform=train_transform
                                )
        val_dataset = Chapman(train=False,
                              opt=opt,
                              #transform=val_transform
                              )
    else:
        raise ValueError(opt.dataset)

    train_sampler = None
    train_loader = torch.utils.data.DataLoader(
        train_dataset, batch_size=opt.batch_size, shuffle=(train_sampler is None),
        num_workers=opt.num_workers, pin_memory=True, sampler=train_sampler)
    val_loader = torch.utils.data.DataLoader(
        val_dataset, batch_size=256, shuffle=False,
        num_workers=8, pin_memory=True)

    return train_loader, val_loader


def set_model(opt):
    if opt.model == 'resnet50':
        model = SupCEResNet(name='resnet50_ecg', num_classes=opt.n_cls)
    elif opt.model == 'CNN':
        model = CNN(num_classes=opt.n_cls)
    elif opt.model == 'CLOCSNET':
        device = (torch.device('cuda' if torch.cuda.is_available() else 'cpu'))
        model = second_cnn_network(first_model=cnn_network_contrastive(
            dropout_type='drop1d',
            p1=0.1,
            p2=0.1,
            p3=0.1,
            embedding_dim=128,
            device=device), embedding_dim=320, noutputs=opt.n_cls)
    else:
        raise ValueError('model not supported: {}'.format(opt.model))

    criterion = torch.nn.CrossEntropyLoss()

    # enable synchronized Batch Normalization
    if opt.syncBN:
        model = apex.parallel.convert_syncbn_model(model)

    if torch.cuda.is_available():
        if torch.cuda.device_count() > 1:
            model = torch.nn.DataParallel(model)
        model = model.cuda()
        criterion = criterion.cuda()
        cudnn.benchmark = True

    return model, criterion


def train(train_loader, model, criterion, optimizer, epoch, opt):
    """one epoch training"""
    model.train()

    batch_time = AverageMeter()
    data_time = AverageMeter()
    losses = AverageMeter()
    top1 = AverageMeter()

    end = time.time()
    for idx, (images, labels, pids) in enumerate(train_loader):
        data_time.update(time.time() - end)

        #print(images.shape, labels.shape)
        images = images.cuda(non_blocking=True)
        labels = labels.cuda(non_blocking=True)
        bsz = labels.shape[0]

        # warm-up learning rate
        warmup_learning_rate(opt, epoch, idx, len(train_loader), optimizer)

        # compute loss
        output = model(images)

        #a = output.cpu().detach().numpy()
        loss = criterion(output, labels)

        # update metric
        losses.update(loss.item(), bsz)
        #acc1, acc5 = accuracy(output, labels, topk=(1, 5))
        acc1 = accuracy(output, labels)[0]
        top1.update(acc1[0], bsz)
        auc = calculate_auc(n_class=opt.n_cls,
                            outputs_list=output,
                            labels_list=labels)
        precision, recall, f1 = calculate_other_metrics(output, labels)


        # SGD
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        # measure elapsed time
        batch_time.update(time.time() - end)
        end = time.time()


        # print info
        #print(losses.val, losses.avg)
        #print(top1.val, top1.avg)
        if (idx + 1) % opt.print_freq == 0:
            print('Train: [{0}][{1}/{2}]\t'
                  'BT {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
                  'DT {data_time.val:.3f} ({data_time.avg:.3f})\t'
                  'loss {loss.val:.3f} ({loss.avg:.3f})\t'
                  'Acc@1 {top1.val:.3f} ({top1.avg:.3f})\t'
                  'Auc {auc:.3f}\t'
                  'precision {precision:.3f}\t'
                  'recall {recall:.3f}\t'
                  'f1 {f1:.3f}'.format(
                   epoch, idx + 1, len(train_loader), batch_time=batch_time,
                   data_time=data_time, loss=losses, top1=top1, auc=auc,
                    precision=precision, recall=recall, f1=f1
            ))
            sys.stdout.flush()

    return losses.avg, top1.avg, auc, precision, recall, f1


def validate(val_loader, model, criterion, opt):
    """validation"""
    model.eval()

    batch_time = AverageMeter()
    losses = AverageMeter()
    top1 = AverageMeter()

    with torch.no_grad():
        end = time.time()
        for idx, (images, labels, pids) in enumerate(val_loader):
            images = images.float().cuda()
            labels = labels.cuda()
            bsz = labels.shape[0]

            # forward
            output = model(images)
            loss = criterion(output, labels)

            # update metric
            losses.update(loss.item(), bsz)
            #acc1, acc5 = accuracy(output, labels, topk=(1, 5))
            acc1 = accuracy(output, labels)[0]
            top1.update(acc1[0], bsz)

            auc = calculate_auc(n_class=opt.n_cls,
                                outputs_list=output,
                                labels_list=labels)
            precision, recall, f1 = calculate_other_metrics(output, labels)

            # measure elapsed time
            batch_time.update(time.time() - end)
            end = time.time()

            if idx % opt.print_freq == 0:
                print('Test: [{0}/{1}]\t'
                      'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
                      'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
                      'Acc@1 {top1.val:.3f} ({top1.avg:.3f})\t'
                      'Auc {auc:.3f}\t'
                      'precision {precision:.3f}\t'
                      'recall {recall:.3f}\t'
                      'f1 {f1:.3f}'.format(
                       idx, len(val_loader), batch_time=batch_time,
                       loss=losses, top1=top1, auc=auc,precision=precision, recall=recall, f1=f1))

    print(' * Acc@1 {top1.avg:.3f}'.format(top1=top1))
    print(' * Auc {auc:.3f}'.format(auc=auc))
    print(' * precision {precision:.3f}'.format(precision=precision))
    print(' * recall {recall:.3f}'.format(recall=recall))
    print(' * f1 {f1:.3f}'.format(f1=f1))
    return losses.avg, top1.avg, auc, precision, recall, f1


def main():
    best_acc = 0
    best_auc = 0
    best_loss = 1e5
    metrics = dict()
    best_acc_acc = 0
    best_acc_auc = 0
    best_auc_acc = 0
    best_auc_auc = 0
    opt = parse_option()

    # build data loader
    train_loader, val_loader = set_loader(opt)

    # build model and criterion
    model, criterion = set_model(opt)

    # build optimizer
    optimizer = set_optimizer(opt, model)

    # tensorboard
    logger = tb_logger.Logger(logdir=opt.tb_folder, flush_secs=2)
    writer = SummaryWriter(comment='supervised')

    # early_stopping
    patience = 20
    path = os.path.join(
        opt.save_folder, 'earlystop.pth')
    early_stopping = EarlyStopping(patience=patience, verbose=True, path=path)

    # training routine
    for epoch in range(1, opt.epochs + 1):
        adjust_learning_rate(opt, optimizer, epoch)

        # train for one epoch
        time1 = time.time()
        loss, train_acc, auc, precision, recall, f1 = train(train_loader, model, criterion, optimizer, epoch, opt)
        time2 = time.time()
        print('epoch {}, total time {:.2f}'.format(epoch, time2 - time1))

        # tensorboard logger
        logger.log_value('train_loss', loss, epoch)
        logger.log_value('train_acc', train_acc, epoch)
        logger.log_value('train_auc', auc, epoch)
        logger.log_value('train_precision', precision, epoch)
        logger.log_value('train_recall', recall, epoch)
        logger.log_value('train_f1', f1, epoch)
        logger.log_value('learning_rate', optimizer.param_groups[0]['lr'], epoch)
        #writer.add_graph(model, input_to_model=None, verbose=False)
        writer.add_scalar('train_loss', loss, epoch)
        writer.add_scalar('train_acc', train_acc, epoch)
        writer.add_scalar('train_auc', auc, epoch)
        writer.add_scalar('learning_rate', optimizer.param_groups[0]['lr'], epoch)

        # evaluation
        loss, val_acc, val_auc, val_precision, val_recall, val_f1 = validate(val_loader, model, criterion, opt)
        logger.log_value('val_loss', loss, epoch)
        logger.log_value('val_acc', val_acc, epoch)
        logger.log_value('val_auc', val_auc, epoch)
        logger.log_value('val_precision', val_precision, epoch)
        logger.log_value('val_recall', val_recall, epoch)
        logger.log_value('val_f1', val_f1, epoch)
        writer.add_scalar('val_loss', loss, epoch)
        writer.add_scalar('val_acc', val_acc, epoch)
        writer.add_scalar('val_auc', val_auc, epoch)



        if loss < best_loss:
            metrics['acc'] = val_acc
            metrics['auc'] = val_auc
            metrics['precision'] = val_precision
            metrics['recall'] = val_recall
            metrics['f1'] = val_f1

        if val_acc > best_acc:
            best_acc = val_acc
            best_acc_acc = val_acc
            best_acc_auc = val_auc
        if val_auc > best_auc:
            best_auc = val_auc
            best_auc_acc = val_acc
            best_auc_auc = val_auc

        if epoch % opt.save_freq == 0:
            save_file = os.path.join(
                opt.save_folder, 'ckpt_epoch_{epoch}.pth'.format(epoch=epoch))
            save_model(model, optimizer, opt, epoch, save_file)


        early_stopping(loss, model)
        if early_stopping.early_stop:
            print('Early stopping')
            save_file = os.path.join(
                opt.save_folder, 'ckpt_epoch_{epoch}.pth'.format(epoch=epoch))
            save_model(model, optimizer, opt, epoch, save_file)
            break

    # save the last model
    save_file = os.path.join(
        opt.save_folder, 'last-0306-ce.pth')
    save_model(model, optimizer, opt, opt.epochs, save_file)


    #print('best accuracy: {:.2f}'.format(best_acc))
    #print('best auc: {:.2f}'.format(best_auc))
    print('accuracy: {:.4f}'.format(metrics['acc']))
    print('auc: {:.4f}'.format(metrics['auc']))
    print('precision: {:.4f}'.format(metrics['precision']))
    print('recall: {:.4f}'.format(metrics['recall']))
    print('f1: {:.4f}'.format(metrics['f1']))

    print('best_acc: acc {:.4f}, auc {:.4f}'.format(best_acc_acc, best_acc_auc))
    print('best_auc: acc {:.4f}, auc {:.4f}'.format(best_auc_acc, best_auc_auc))

if __name__ == '__main__':
    main()