yu4u / mixup-generator Goto Github PK

why the training accuracy goes done when I add mixup?

Attempting to employ multithreading throws the warning/error
"UserWarning: Using a generator with use_multiprocessing=True and multiple workers may duplicate your data. Please consider using thekeras.utils.Sequence class. UserWarning('Using a generator with use_multiprocessing=True`'"

I've attempted to rewrite it a bit to bring it into line with the keras sequence class but I'm still getting that warning and am admittedly new to this generator.
https://keras.io/utils/#sequence

Before going further I wanted to ask the obvious question of whether there's a constraint somewhere that makes it hard to convert to thread safe? If not, I'd like to put it in as an enhancement request.

I ran your code and got some different results，

Without mixup:
Test loss: 0.5639845668315887
Test accuracy: 0.9017000198364258

With mixup alpha = 0.2:
Test loss: 0.5545909198760987
Test accuracy: 0.90420001745224

There doesn't seem to be much difference between the two.Can you analyze the reason?

Is batch_ids[:self.batch_size] and batch_ids[self.batch_size:] trying to mirror each other?
However, batch_ids[self.batch_size:] should return an empty array.
a = [1,2,3,4,5]
a[:5] -> [1, 2, 3, 4, 5]
a[5:] -> []

Hi @yu4u!
Thank you for your work!

After studying the repo, I still have one question about label processing.
In the original implementation , the processing of mixing up for labels happens at the time of loss computing:

def mixup_criterion(criterion, pred, y_a, y_b, lam):
    return lam * criterion(pred, y_a) + (1 - lam) * criterion(pred, y_b)

In your implementation, you're mixing up the labels:

y1 = self.y_train[batch_ids[:self.batch_size]]
y2 = self.y_train[batch_ids[self.batch_size:]]
y = y1 * y_l + y2 * (1 - y_l)

After inserting the resulting labels into the equation even for binary_cross_entropy, the resulting equation isn't the same.
So, the question is, what was the motivation for changing the place for performing the mixup for labels?

hello I wonder if it can be implemented in pytorch ?

how to change the source code to fix detection, about two image size and box label add.
can it support keras Sequence, make the data generator to multi-processes version?

thank you for this!