hongyi-zhang/mixup

Implementation of the mixup training method

mixup's Issues

How to get results reported in paper for cifar-10?

Thanks for a very interesting paper!! I was trying out your code and was not able to get the results posted in paper but got the results shown on GitHub page.

CIFAR10 PreAct ResNet-18 results of 3.9% are reported in the paper. But the GitHub page gives it 4.24%. Are there any other setting to change like epochs, weight decay etc?

Thanks

Can't reproduce the results with sample run provided for CIFAR-10

I ran the command as provided in readme:

python easy_mixup.py --sess my_session_1 --seed 11111

Since, I am using pytorch >= 1.0
I had to fix the loss calculation:

train_loss += loss.data[0] to train_loss += loss.data.item()
and test_loss += loss.data[0] to test_loss += loss.data.item()

The maximum accuracy I got was 93.90% which translates to an error of 6.1

What am I doing wrong here?

样本数量请教

在每个batch中使用mixup方法，然后返回mixup后的样本，那么请问下这个batch中的样本数量是没有变吗，那整个训练集的样本数量是没有变吗？

Different implement of mixup between paper and repo

There are 2 different implementation of mixup, which version should I use?

From paper

From repo

The learning rate adjust code looks wrong

After 100 epoch lr *= 0.1 on every epoch so quickly it become zero.
After 150 epoch lr *= 0.01 on every epoch.

    if epoch >= 100:
        lr /= 10
    if epoch >= 150:
        lr /= 10

Ubuntu system restarts at loss.backward()

Ubuntu system restarts at loss.backward(). I am using Ubuntu 16.04 and cuda 9.0. Could assistance be provided to resolve this issue? Much appreciated.

TypeError: tight_layout() takes 0 positional arguments but 3 were given

I've tried to run the code inside colab. however, it shows the following Error:

Do I need to have any specific things in my colab? Thank you

0%|          | 10/20001 [00:00<08:23, 39.67it/s]
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
[<ipython-input-2-51480497a079>](https://localhost:8080/#) in <module>
    156                                'samples/example_z=%d_%s_%1.1f_%06d.pt' %
    157                                (n_latent, shape, mixup, iteration))
--> 158                     plot(plot_real, plot_fake, mixup, iteration)

[<ipython-input-2-51480497a079>](https://localhost:8080/#) in plot(x, y, mixup, iteration)
     51                 plt.xlim(*lims)
     52                 plt.ylim(*lims)
---> 53                 plt.tight_layout(0, 0, 0)
     54                 plt.show()
     55                 plt.savefig("images/example_z=%d_%s_%1.1f_%06d.png" %

TypeError: tight_layout() takes 0 positional arguments but 3 were given

Always get the correct 0.

HI I have tried to implement your code in my experiment on UCF101.
but I did't get any improvement and always get the zero correct.
It is a little strange.
my code
``
def train_1epoch(self):
print('==> Epoch:[{0}/{1}][training stage]'.format(self.epoch, self.nb_epochs))
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter() #switch to train mode
self.model.train()
end = time.time()
train_loss = AverageMeter()
total = 0
correct = 0
# mini-batch training
progress = tqdm(self.train_loader)
for i, (data_dict,label) in enumerate(progress):

       # measure data loading time
        data_time.update(time.time() - end)
       # generate mixed inputs, two one-hot label vectors and mixing coefficient 
        # transfer the label into one-hot Encoder

        # label = torch.zeros(label.shape[0], 101).scatter_(1, label.reshape(-1, 1), 1).cuda()
        # print(label.shape[0])
        label = label.cuda()
        # compute output
        output = Variable(torch.zeros(len(data_dict['img1']),101).float()).cuda()
        # print(len(data_dict['img1'])

        for i in range(len(data_dict)):
            key = 'img'+str(i)
            input_var = (data_dict[key]).cuda()
            # generate mixed inputs, two one-hot label vectors and mixing coefficient 
            input_var, label_a, label_b, lam = mixup_data(input_var, label, args.alpha, True)
            input_var, label_a, label_b = Variable(input_var), Variable(label_a), Variable(label_b)
            output += self.model(input_var)

        criterion = self.criterion
        loss = mixup_criterion(criterion, output, label_a, label_b, lam)
        # print(label_a.argmax(dim=1).data,label_b.argmax(dim=1).data)
        # loss = loss_func(criterion, output)
        # print(args.alpha, lam)
        # compute gradient and do SGD step
        self.optimizer.zero_grad()
        loss.backward()
        self.optimizer.step()

        # measure accuracy and record loss
        # train_loss += loss.data[0]
        train_loss.update(loss.data[0], data_dict[key].size(0))
        # print(loss.data[0])
        _, predicted = torch.max(output.data, 1)
        total += label.size(0)
        # print(label.size(0))
        correct += lam * predicted.eq(label_a.data).cpu().sum() + (1 - lam) * predicted.eq(label_b.data).cpu().sum()
        # print(predicted.eq(label_a.argmax(dim=1).data).cpu().sum())
        # print(label_a.data)
        # measure elapsed time
        batch_time.update(time.time() - end)
        end = time.time()
    
    info = {'Epoch':[self.epoch],
            'Batch Time':[round(batch_time.avg,3)],
            'Data Time':[round(data_time.avg,3)],
            'Loss':[round(train_loss.avg,5)],
            'correct':[round(correct,4)],
            'Prec@1':[round(correct/total,4)],
            'Prec@5':[round(correct/total,4)],
            'lr': self.optimizer.param_groups[0]['lr'],
	        'weight-decay': args.decay
            }
    record_info(info, 'record/spatial/rgb_train.csv','train')

``
I did't know where is wrong....