Hi Harry,

Thanks for your implementation of our work!

Because mixup has a strong regularization effect, the best weight decay parameter for mixup is usually smaller than normal training. For example, you should see mixup accuracy improves if the weight decay is set to 1e-4, which is what we used in the paper.

We didn't search for the best weight decay though, so other choices may lead to further improvement.

Best,

Hello. Thank you for your share.

First of all, My English is not so fluent. Please understand it.

There is a mistake in shuffle_minibatch function in main_cifar10.py, main_cifar100.py

def shuffle_minibatch(inputs, targets, n_class=88, mixup=True):
    """Shuffle a minibatch and do linear interpolation between images and labels.
    Args:
        inputs: a numpy array of images with size batch_size x H x W x 3.
        targets: a numpy array of labels with size batch_size x 1.
        mixup: a boolen as whether to do mixup or not. If mixup is True, we
            sample the weight from beta distribution using parameter alpha=1,
            beta=1. If mixup is False, we set the weight to be 1 and 0
            respectively for the randomly shuffled mini-batches.
    """

According to your comment, inputs parameter shape is batch_size x H x W x 3. but your input shape is batch_sizex 3 x H x W, isn't it?

I think you change comment

Thank you for your work again.

Hello.

Thank you for your nice implementation and share.

There is a typo in Usage section in readme.md file.

#Train and test CIFAR 10 with mixup.
python main_cifar10.py --mixup --exp='cifar10_nomixup'

I think you intended 'cifar10_mixup' for exp name.

Thank you for your work again.