clevercool/ANT-Quantization

code for figure 14

Opened this issue · 0 comments

Hello,
I was wondering if you are able to provide the code used to generate figure 14 in your paper? I was able to mostly replicate the plot for the ResNet-18 weights but some layers are still off (esp. the downsampling layers). I used the default weights (ResNet18_Weights.DEFAULT from torchvision) without any fine-tuning or optimized range clipping. Did you use custom weights or clipping ranges?
Which input image(s) did you use to get the activation plots?

Thanks for any answers in advance.