amlab-amsterdam / l0_regularization Goto Github PK

Would it be possible to get a README file that explicitly describes how to run the experiments from the paper?

Please remove .data[0] and put .item() because of this post and Accumulating losses

It seems that there is a little bug in the code which causes the training to stop after one epoch at line 172. The args,type_net in train_wide_resnet.py is not defined. I think that the args,type_net argument should be added right after defining the argument parser in line 17.

The wide-resnet experiments provided in the repository does not appear to replicate the results in the paper.
Upon running the experiment, neither the expected L0 norm or the expected flops changes at all

This is corroborated by the following excerpt from the paper The State of Sparsity in Deep Neural Networks (https://arxiv.org/pdf/1902.09574.pdf).

Would greatly appreciate any feedback on how to utilize this repository to replicate L0 regularization on Wide Resnets for the CIFAR datasets.

Dear authors,

Please have a look at third term in equation(26) in the paper. I don't think the normalization (Q(1)-Q(0)) should be included. Please check if I were wrong.

Best,
Tianjian