Question for ClassStratifiedSampler

[min9kwak]

Thank you for the code.

I am trying to reproduce the results and found some strange things in the code. I ran main.py with cifar10_train.yaml. As demonstrated, I thought that ClassStratifiedSamplershould produce [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, ... ], I believe that it makes us use label_matrix for convenience. However, when I checked the actual labels (sdata[-1] on line 280 of paws_train.py), the labels were just sampled in a stratified manner, but were not in the appropriate order. For example, [1, 8, 7, 9, 5, 2, 6, 0, 4, 3] is repeated. sdata[-1] and labels on paws_train.py do not match.... Nevertheless, the model is trained properly. Did I miss something? Thank you in advance.

• I am using pytorch 1.9.0

[MidoAssran]

Hi @min9kwak, thanks for your interest.

Yes your observation is correct, but this is actually the correct behaviour. To clarify, ClassStratifiedSampler should sample torch.cat[[a,b,c,...], [a,b,c,...], ..., [a,b,c,...]], but where [a,b,c,...] are randomly sampled classes in each iteration.

Since the loss is non-parametric, the use of label_matrix is just to identify which samples belong to the same class. For example, if on CIFAR10 you set classes_per_batch: 7 in your config, then in each iteration, ClassStratifiedSampler will only sample 7 random classes in each iteration, and label_matrix will be of shape [num_support_imgs x 7] so in that case you cannot compare it to sdata[-1], which is of shape [num_support_imgs x 10].

[min9kwak]

@MidoAssran Thanks for your quick and clear explanation. As I understood, regardless of whether the order is [a, b, c, a, b, c, ...] or [b, c, a, b, c, a, ...], it is okay to use the label_matrix because it doesn't really have to indicate the certain class information. It's like a cluster assignment for every iteration or batch. It was really helpful. Thank you!!

[MidoAssran]

Yes precisely :)