You were right! by increasing the batch_size now the distribution is closer to 50/50 (sometimes more like 60/50). uff I am glad I asked! thank you!
Hi @ptrblck , Thanks for answering all those questions! I was just wondering if balancing the samples in batches is a good idea. From a brief view, it looks like this would change the negative sample distribution and potentially create a bias, especially for very few shot cases where perhaps the positive proportion is << 1%. Is there some ways to just “stratify” each batch according to labels as we do sklearn.model_selection.train_test_split, but not necessarily changing the weights of samples?
I think you could directly use the scikit-learn stratify split functionality and create the stratified indices for the splits. Once this is done you could use these indices in a custom
BatchSampler and load the entire batch of (stratified) samples in the