Reproducing the results for quantisation

[thrypuro]

Hello, I am a student Research assistant from University of Edinburgh. I been trying to reproduce the results from the slalom paper for the past month. I had a few questions about the codebase.

1. Do you know which exact version (i.e. which year) of imagenet is used for this paper?
2. I have managed to get the codebase to run, I have to run it without the SGX features, which innolved replacing the modulo operation in quant_layers.py line 31 to the pythonic x % p, do you think there might be any issues with the accuracy with such changes?
3. My results seem to be different from the paper, in the sense that without quantisation, the baseline is more accurate than what the paper says, do you know if over the years the model accuracy has changed from when you've done it?

Thank you.