Yang He

Hi @zeliu98 , Thanks for your great work! I met a similar problem with @lcmeng . On four V-100 GPUs, I used the default commander below. It turns out the...

@Andy1621 Hi, may I ask any difference between this repo and [your official repo](https://github.com/Sense-X/UniFormer)? Thanks a lot!

Sorry for the unclear code. This line is not necessary and would not be active. The following pruning operations are conducted on the layers within the range(args.layer_begin, args.layer_end + 1,...

Thank you!

The reason might come from the version of PyTorch. For the new PyTorch version, the batch-norm layer has five state dicts, including 'bn1.weight', 'bn1.bias', 'bn1.running_mean', 'bn1.running_var', 'bn1.num_batches_tracked'. The old version...

Thanks for your interest. The following paragraphs explain why we not directly calculate GM.  We turn to find the filter with the most "GM" property to reduce the computation.

Hi. Let me explain. First, "pruning filters" almost equals to "set them to zero". Will the pruned ones are recovered (become non-zero) during fine-tuning? That depends on the gradients. Gradients...

Thanks for your interest. Several different Downsample layers are provided in your mentioned files. Using [DownsampleC](https://github.com/he-y/filter-pruning-geometric-median/blob/76f0ffefbdba7335251bfbafaadce49a2f08d2b9/models/res_utils.py#L14) could solve the mismatch problem.

Thanks for the reply. Is it possible to release the student model based on the accuracy table? This might be easier for us to use. For example, for every accuracy...

Hi, please see [updates here](https://github.com/he-y/filter-pruning-geometric-median#whats-new) to use FPGM in pytorch and NNI.