dichotomies / proxy-nca

PyTorch Implementation of `No Fuss Distance Metric Learning using Proxies`.
MIT License
181 stars 33 forks source link

Where the proxies are changed during training #18

Closed kifjj closed 3 years ago

kifjj commented 3 years ago

First, bug thank you for this. I find it super useful.

Then the my beginner question:

In another thread you wrote: The proxies are randomly initialized and change during training.

but where and how is this happening? Is this line called every time? Because I cannot see any other line updating proxies.

Thank again

kifjj commented 3 years ago

Ha, I found the answer. They are updated by the training loop because they are Parameter and by default requires grad is true