Issues with training using GMP about str HOT 3 CLOSED

Hi @chrundle,

Thanks for bringing this to my notice. GMP isn't well tested, and I agree that args.curr_prune_epoch should be replaced by epoch.

However, the second aspect of the question seems surprising as if you set the prune rate to 0 and train it with GMP, it should be similar to Dense training. Can you please let me know if you face the same issues with the Dense training or a pseudo Dense training with GMP + 0% prune rate? This is not the official implementation of GMP, but the prune mechanism for GMP is exactly the same. I am pushing the one with the minor edit you have suggested.

Aditya

from str.

Hi @adityakusupati,

Thanks for the quick response! Following your suggestion, I tried training a Conv model on CIFAR-10 using your code where the conv_type is DenseConv. Initially, I was getting the same issue but I thought to check the hyperparameters in the config file against the baseline configs from the hidden-networks repo. Using their baseline configs resolved the issue with the model not learning. Glad to know it was a simple fix. When I found the first issue I mentioned I wasn't sure if something else more complicated than hyperparameters was going on which is why I reached out first. Thanks again for your timeliness! I'm closing the issue with this comment since you said you will push the fix for the first issue.

from str.

That's great. The hparams in this repo are all for ImageNet. I don't expect them to work on CIFAR-10. Glad to know that you were able to find the right hparams from the hidden-networks repo.

from str.