kovelja009/mixed_precision_dnn_pytorch

Pytorch implementation of the paper Mixed Precision DNNs: All you need is a good parametrization.

mixed_precision_dnn_pytorch's Issues

About Training Settings

Thanks for your job!
When I train the resnet-18 using ''python training.py PARAMETRIC_FP_WAQ_DELTA_XMAX_RLR --cfg train_resnet_quant_fp.cfg'' that you have provided, the training acc just holds up 10% after the 10 epochs and 8 bit is all it learned.

I wonder why the epoch is set to 10 without learning rate dropping scheduler. By the way, if I want to change cifar-10 to imagenet, could I just replace the dataset settings and add the learning rate dropping scheduler without anything else to change?
Looking forward to your reply.

About Sony Implementation

Excuse me, may I take a few minutes of your time again？
I'd like to ask you if you can run the code of Sony correctly. Because of the old version of nnabla-ext-cuda101 that I couldn't download (.whl and so on) directly, I can't run that implementation correctly. Do you have any ideas?
Looking forward to your reply.