Argmax performance slower than numpy?

tbenst · June 8, 2018, 2:26am

For some reason, torch.argmax is slower for me than transferring an array to CPU and then calling np.argmax. Any ideas why? Should I file a bug report?

gist.github.com

https://gist.github.com/tbenst/de3f61ac9956778a2ca6d5db94b526ef

argmax_bench.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",

This file has been truncated. show original

richard · June 20, 2018, 4:47pm

That sounds bad, please file an issue.

tbenst · June 23, 2018, 1:21am

Thanks, filed: https://github.com/pytorch/pytorch/issues/8817