Tweeted By @jeremyphoward

Thanks to @ptrblck_de we have fast fp16 depthwise convs in @PyTorch - and he even created a create benchmark spreadsheet to show where they help.

However, that spreadsheet is a bit tricky to interpret, so I've made a couple of summary tables below.
1/3https://t.co/Uy5SYs1VwD

— Jeremy Howard (@jeremyphoward) October 26, 2019