Giters

cfchen-duke / ProtoPNet

This code package implements the prototypical part network (ProtoPNet) from the paper "This Looks Like That: Deep Learning for Interpretable Image Recognition" (to appear at NeurIPS 2019), by Chaofan Chen* (Duke University), Oscar Li* (Duke University), Chaofan Tao (Duke University), Alina Jade Barnett (Duke University), Jonathan Su (MIT Lincoln Laboratory), and Cynthia Rudin (Duke University) (* denotes equal contribution).

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Is it real that 1000 epochs are need

lovekeyczw opened this issue · comments

commented

In my lab, one epoch uses about 800s, 1000 epochs are too large.

commented

In my lab, one epoch uses about 800s, 1000 epochs are too large.

Don't you get a KeyError somewhere after the 10th epoch? Mine takes a long time too but stops due to this error.

commented

I also find this strange, I guess you're supposed to just stop it yourself when it converges? Seems to make reproducibility difficult to navigate maybe.

Also the epochs seem to reset to 0 often so it's hard to know which one you're on.

It's very confusing to know when to stop the training.

commented

I just finished running 1000 epochs on a machine with 4 Titan V cards, which took 15 days. However, it reached maximum test accuracy at epoch 15 (~78.3%). After that it seems to oscillate between 77% and 76% through the end of the training.

commented

Hello, the experiments in the paper were done on 50 epochs or less.

commented

Thanks @alinajadebarnett !

So that means that 30-50ish epochs with the augmented dataset were done (which is x30 times bigger), and then an additional 20 epochs for fine tuning the output linear layer after projection?