对比学习extra feature + classifier

main_linear

• Resnet + Classifier + CELoss

def set_model(opt):
    model = SupConResNet(name=opt.model)
    criterion = torch.nn.CrossEntropyLoss()

    classifier = LinearClassifier(name=opt.model, num_classes=opt.n_cls)

    ckpt = torch.load(opt.ckpt, map_location='cpu')
    state_dict = ckpt['model']

    if torch.cuda.is_available():
        if torch.cuda.device_count() > 1:
            model.encoder = torch.nn.DataParallel(model.encoder)
        else:
            new_state_dict = {}
            for k, v in state_dict.items():
                k = k.replace("module.", "")
                new_state_dict[k] = v
            state_dict = new_state_dict
        model = model.cuda()
        classifier = classifier.cuda()
        criterion = criterion.cuda()
        cudnn.benchmark = True

        model.load_state_dict(state_dict)
    else:
        raise NotImplementedError('This code requires GPU')

    return model, classifier, criterion


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

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

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

        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
        with torch.no_grad():
            features = model.encoder(images)
        output = classifier(features.detach())
        loss = criterion(output, labels)

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

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

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

        # print info
        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})'.format(
                   epoch, idx + 1, len(train_loader), batch_time=batch_time,
                   data_time=data_time, loss=losses, top1=top1))
            sys.stdout.flush()

    return losses.avg, top1.avg

main_supcon

def set_model(opt):
    model = SupConResNet(name=opt.model)
    criterion = SupConLoss(temperature=opt.temp)

    # 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.encoder = torch.nn.DataParallel(model.encoder)
        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()

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

        images = torch.cat([images[0], images[1]], dim=0)
        if torch.cuda.is_available():
            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
        features = model(images)
        # 使用torch.cat函数将切分后的两个子特征f1和f2在第一个维度上进行拼接，即将它们作为两个
        # 通道（unsqueeze(1)）拼接在一起，得到最终的特征features
        f1, f2 = torch.split(features, [bsz, bsz], dim=0)
        features = torch.cat([f1.unsqueeze(1), f2.unsqueeze(1)], dim=1)
        
        if opt.method == 'SupCon':
            loss = criterion(features, labels)
        elif opt.method == 'SimCLR':
            loss = criterion(features)
        else:
            raise ValueError('contrastive method not supported: {}'.
                             format(opt.method))

        # update metric
        losses.update(loss.item(), bsz)

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

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

        # print info
        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})'.format(
                   epoch, idx + 1, len(train_loader), batch_time=batch_time,
                   data_time=data_time, loss=losses))
            sys.stdout.flush()

    return losses.avg

• main函数

def main():
    opt = parse_option()

    # build data loader
    train_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)

    # 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(train_loader, model, criterion, optimizer, epoch, opt)
        time2 = time.time()
        print('epoch {}, total time {:.2f}'.format(epoch, time2 - time1))

        # tensorboard logger
        logger.log_value('loss', loss, epoch)
        logger.log_value('learning_rate', optimizer.param_groups[0]['lr'], epoch)

        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)

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

---

写的很好的utils

def warmup_learning_rate(args, epoch, batch_id, total_batches, optimizer):
    if args.warm and epoch <= args.warm_epochs:
        p = (batch_id + (epoch - 1) * total_batches) / \
            (args.warm_epochs * total_batches)
        lr = args.warmup_from + p * (args.warmup_to - args.warmup_from)

        for param_group in optimizer.param_groups:
            param_group['lr'] = lr

def adjust_learning_rate(args, optimizer, epoch):
    lr = args.learning_rate
    if args.cosine:
        eta_min = lr * (args.lr_decay_rate ** 3)
        lr = eta_min + (lr - eta_min) * (
                1 + math.cos(math.pi * epoch / args.epochs)) / 2
    else:
        steps = np.sum(epoch > np.asarray(args.lr_decay_epochs))
        if steps > 0:
            lr = lr * (args.lr_decay_rate ** steps)

    for param_group in optimizer.param_groups:
        param_group['lr'] = lr

def accuracy(output, target, topk=(1,)):
    """Computes the accuracy over the k top predictions for the specified values of k"""
    with torch.no_grad():
        maxk = max(topk)
        batch_size = target.size(0)

        _, pred = output.topk(maxk, 1, True, True)
        pred = pred.t()
        correct = pred.eq(target.view(1, -1).expand_as(pred))

        res = []
        for k in topk:
            correct_k = correct[:k].view(-1).float().sum(0, keepdim=True)
            res.append(correct_k.mul_(100.0 / batch_size))
        return res

class AverageMeter(object):
    """Computes and stores the average and current value"""
    def __init__(self):
        self.reset()

    def reset(self):
        self.val = 0
        self.avg = 0
        self.sum = 0
        self.count = 0

    def update(self, val, n=1):
        self.val = val
        self.sum += val * n
        self.count += n
        self.avg = self.sum / self.count

def set_optimizer(opt, model):
    optimizer = optim.SGD(model.parameters(),
                          lr=opt.learning_rate,
                          momentum=opt.momentum,
                          weight_decay=opt.weight_decay)
    return optimizer


def save_model(model, optimizer, opt, epoch, save_file):
    print('==> Saving...')
    state = {
        'opt': opt,
        'model': model.state_dict(),
        'optimizer': optimizer.state_dict(),
        'epoch': epoch,
    }
    torch.save(state, save_file)
    del state

Loss

对比损失：Supervised Contrastive Loss（监督对比损失）是一种在监督对比学习中使用的损失函数。它旨在学习既具有区分性又具有对同一类别内变化具有不变性的表示。

监督对比学习的目标是最大化正样本对（同一类别的样本）的一致性，并最小化负样本对（不同类别的样本）的一致性。监督对比损失通过鼓励正样本对的表示在嵌入空间中更加接近，同时将负样本对的表示推开来实现这一目标。

• If both labels and mask are None, it degenerates to SimCLR unsupervised loss:

 Args:
     features: hidden vector of shape [bsz, n_views, ...].
     labels: ground truth of shape [bsz].
     mask: contrastive mask of shape [bsz, bsz], mask_{i,j}=1 if sample j
           has the same class as sample i. Can be asymmetric.
Returns:
     A loss scalar

class SupConLoss(nn.Module):
    """Supervised Contrastive Learning: https://arxiv.org/pdf/2004.11362.pdf.
    It also supports the unsupervised contrastive loss in SimCLR"""
    def __init__(self, temperature=0.07, contrast_mode='all',
                 base_temperature=0.07):
        super(SupConLoss, self).__init__()
        self.temperature = temperature
        self.contrast_mode = contrast_mode
        self.base_temperature = base_temperature

    def forward(self, features, labels=None, mask=None):
        """Compute loss for model. If both `labels` and `mask` are None,
        it degenerates to SimCLR unsupervised loss:
        https://arxiv.org/pdf/2002.05709.pdf

        Args:
            features: hidden vector of shape [bsz, n_views, ...].
            labels: ground truth of shape [bsz].
            mask: contrastive mask of shape [bsz, bsz], mask_{i,j}=1 if sample j
                has the same class as sample i. Can be asymmetric.
        Returns:
            A loss scalar.
        """
        device = (torch.device('cuda')
                  if features.is_cuda
                  else torch.device('cpu'))

        if len(features.shape) < 3:
            raise ValueError('`features` needs to be [bsz, n_views, ...],'
                             'at least 3 dimensions are required')
        if len(features.shape) > 3:
            features = features.view(features.shape[0], features.shape[1], -1)

        batch_size = features.shape[0]
        if labels is not None and mask is not None:
            raise ValueError('Cannot define both `labels` and `mask`')
        elif labels is None and mask is None:
            mask = torch.eye(batch_size, dtype=torch.float32).to(device)
        elif labels is not None:
            labels = labels.contiguous().view(-1, 1)
            if labels.shape[0] != batch_size:
                raise ValueError('Num of labels does not match num of features')
            mask = torch.eq(labels, labels.T).float().to(device)
        else:
            mask = mask.float().to(device)

        contrast_count = features.shape[1]
        contrast_feature = torch.cat(torch.unbind(features, dim=1), dim=0)
        if self.contrast_mode == 'one':
            anchor_feature = features[:, 0]
            anchor_count = 1
        elif self.contrast_mode == 'all':
            anchor_feature = contrast_feature
            anchor_count = contrast_count
        else:
            raise ValueError('Unknown mode: {}'.format(self.contrast_mode))

        # compute logits
        anchor_dot_contrast = torch.div(
            torch.matmul(anchor_feature, contrast_feature.T),
            self.temperature)
        # for numerical stability
        logits_max, _ = torch.max(anchor_dot_contrast, dim=1, keepdim=True)
        logits = anchor_dot_contrast - logits_max.detach()

        # tile mask
        mask = mask.repeat(anchor_count, contrast_count)
        # mask-out self-contrast cases
        logits_mask = torch.scatter(
            torch.ones_like(mask),
            1,
            torch.arange(batch_size * anchor_count).view(-1, 1).to(device),
            0
        )
        mask = mask * logits_mask

        # compute log_prob
        exp_logits = torch.exp(logits) * logits_mask
        log_prob = logits - torch.log(exp_logits.sum(1, keepdim=True))

        # compute mean of log-likelihood over positive
        mean_log_prob_pos = (mask * log_prob).sum(1) / mask.sum(1)

        # loss
        loss = - (self.temperature / self.base_temperature) * mean_log_prob_pos
        loss = loss.view(anchor_count, batch_size).mean()

        return loss

Net

model_dict = {
    'resnet18': [resnet18, 512],
    'resnet34': [resnet34, 512],
    'resnet50': [resnet50, 2048],
    'resnet101': [resnet101, 2048],
}

class SupConResNet(nn.Module):
    """backbone + projection head"""
    def __init__(self, name='resnet50', head='mlp', feat_dim=128):
        super(SupConResNet, self).__init__()
        model_fun, dim_in = model_dict[name]
        self.encoder = model_fun()
        if head == 'linear':
            self.head = nn.Linear(dim_in, feat_dim)
        elif head == 'mlp':
            self.head = nn.Sequential(
                nn.Linear(dim_in, dim_in),
                nn.ReLU(inplace=True),
                nn.Linear(dim_in, feat_dim)
            )
        else:
            raise NotImplementedError(
                'head not supported: {}'.format(head))

    def forward(self, x):
        feat = self.encoder(x)
        feat = F.normalize(self.head(feat), dim=1)
        return feat


class SupCEResNet(nn.Module):
    """encoder + classifier"""
    def __init__(self, name='resnet50', num_classes=10):
        super(SupCEResNet, self).__init__()
        model_fun, dim_in = model_dict[name]
        self.encoder = model_fun()
        self.fc = nn.Linear(dim_in, num_classes)

    def forward(self, x):
        return self.fc(self.encoder(x))