Reproduction of evaluation result about rude-carnie HOT 2 CLOSED

It has been a long time since I have run with the same defaults as the original paper, but these numbers you post seem quite low. I typically see quite higher precision @ 1 than in the paper and slightly lower @ 2 (for me, this tradeoff is worth it), but this will depend on the params you pass. The numbers you show here are much lower than I typically see... It didnt help that I made it hard to tweak the hyper parameters without modding the code (see below). I think the performance really depends on the hyper params (batch size LR, LR decay, steps per decay).

To make it a bit easier, I have changed some of the defaults for the hyper params to be closer to original paper, and I made it easier to change the LR decay and steps per decay by offering flags. This should help you match closer to the original paper (and tweak as you see fit). I will probably change the default batch size as well shortly (lowering to 50).

In terms of the implementation, I believe that the L/H model should be identical to the original Caffe code. There is a slight difference in how we handle the mean image -- in TF I use https://www.tensorflow.org/api_docs/python/tf/image/per_image_standardization where original code uses the true mean image I believe.

from rude-carnie.

I have been able now to improve over the paper @ 1 and @ 2 with a higher learning rate, but similar decay. Try eta=0.01 and rest defaulted. You should see better results

from rude-carnie.