'AssertionError: Batch size should be even when using this' of Mixup funcution

[kuri54]

I am trying to train my own dataset using swin_small_patch4_window7_224.

Data set

Number of classes: 2 classes (imbalanced data with class 0:class 1=1:5)
Train size: 6365
Validation size: 1128

Using timm rand_augment_transform and Mixup

# rand_augment
rand_augment_transform(config_str='rand-m9-mstd0.5', hparams={})

# mixup + cutmix 
mixup_args = {
     'mixup_alpha': 0.3,
     'cutmix_alpha': 0.3,
     'cutmix_minmax': None,
     'prob': 1.0,
     'switch_prob': 0.,
     'mode': 'batch',
     'label_smoothing': 0,
     'num_classes': 2}

mixup_fn = Mixup(**mixup_args)

CosineLRScheduler from timm.scheduler was used for scheduler and SoftTargetCrossEntropy from timm.loss was used for loss function.

What is happening

When I started the training, I got the following error message.

File "/usr/local/lib/python3.7/dist-packages/timm/data/mixup.py", line 210, in __call__
    assert len(x) % 2 == 0, 'Batch size should be even when using this'
AssertionError: Batch size should be even when using this

I get an error in the following places.

for idx, (inputs, labels) in enumerate(train_dataloader):
        inputs = inputs.to(device)
        labels = labels.to(device)

        inputs, labels_mixup = mixup_fn(inputs, labels)  <- 

        optimizer.zero_grad()

        outputs = model(inputs)
        _, preds = torch.max(outputs, 1)
        loss = criterion(outputs, labels_mixup)

        loss.backward()
        optimizer.step()
        scheduler.step(epoch+1)

Batch size should be even when using this. (For example 1, 19, 67, 32, 64...)
Why is this?Is there any way to resolve this?

If my code is wrong, I would appreciate it if you could point it out.

[rwightman]

@kuri54 it's exactly what it says, train with an even batch size

[kuri54]

@rwightman
Thanks for answering!

Yeah, you're right.
I know exactly what you mean, but what are the possible cases?
Is an imbalanced data set included in this case?

I tried many things and solved the problem. I downsampled the dataset to make it a balanced dataset and it worked.

[rwightman]

It shouldn't have anything to do with your dataset size as long as drop _last=True in dataloader, which is common to enable.

[kuri54]

Oh, I see.
I'm finally starting to understand.

I decided to train with a balanced data set this time, and I will try again when I have more time.
If I have any questions at that time, please let me ask again.

Thank you very much.