maxhalford / pytorch-resample Goto Github PK

Hello,

I have classification dataset of 32k samples, where 89% of my data is class 0 and the other 11% is class 1. I tried to use pytorch-resample to Undersample my 0-class, with a distribution of 0.5/0.5. It did work great, but I noticed that the dataset differs at each epoch, making the training useless. I know there is a random seed initialized the first time we call the Undersample class, but I didn't manage to reset this seed after each epoch since the dataset is part of the DataLoader.

Have you not encountered a similar behavior? Do you have an idea for a fix? Thanks in advance

Nice work! I was about to use this handy tool until I realize my problem was even trickier.

I'm dealing with a very large (TB scale) WebDataset, which also inherits IterableDataset. And the dataset can keep growing. My goal is to do balanced sampling based on some attributes of my samples. In other words, I want to have N classes with each of them having equal weights. It would be straightforward to do with this resample tool if I knew all the possible classes. However, that requires me to iterate through the whole dataset, which can take hours - which would have been fine if I only do it once, but I may have to do it over and over since my dataset will grow in size in the future (and new classes will come). I'm also aware that one workaround is to manage it out of the loop by maintaining an incremental list of classes on disk.

Still, it would be even better if this can be handled during training. I was thinking of building the desired_dist dynamically by initializing it with an empty dict and adding unseen classes to it with equal constant weights on-the-fly. It seems this might work but I've not tested so. Do you think this is something worth having in the repo? And do you see any caveats of doing so? Any suggestions are appreciated.

Hi

Thanks for the wonderful code.
Can you please recheck you HybridSampler implementation?
I could not run the examples and the benchmark code from anywhere especially for the HybridSampler.
The other parts of the code runs absolutely fine.

Thanks
Devraj