How to train each batch to have the same class

[94Andrew (로또곧당첨)]

I want to train each batch to have the same class. For example, when batch size is 32, update once with 32 data of class 0, and then learn with 32 data of randomly selected class.
If you learn all data in class 0 and then learn about class 1, it can’t be learning.

[Eta_C]

Load the same number of data per class

I know weighted sampler can solve imbalanced data problem. However, I wonder is there a way to load exactly the same number of data per class ? What I need now is, for example, a batch of 10 samples from class A, 10 from class B, 10 from class C, ETC…( I mean “not probablistically” but deterministically make sure to load 10 sample per class. ) I also want to know how to combine that solution with torch.utils.data.DataLoader

Load the same number of data per class

I had the same problem when I re-implement PFE, which needs 64 classes, which have 4 data in each class, as an input. Here’s my solution (not very good but useful for me). #class Dataset def __init__(self): self.label_to_index = {'A': [1, 3, 4], 'B': [0, 2, 5], 'C': [6, 7, 8]} def __getitem__(self, item): images = np.empty((1, *image_shape)) labels = np.empty((1,)) for i, (label, indices) in enumerate(self.label_to_index.items()): index = np.random.choice(indices, 1, …

[Mika_Spilner (Mika)]

You can make custom Sampler based on torch.utils.data.WeightedRandomSampler.
I think you can set a class probability as 1 and rest of the classes 0, then in next iteration you can choose another class and set its probability 1 and rests 0, and so on.