Selecting checkpoint for evaluation

feidfoe / learning-not-to-learn

[CVPR2019]Learning Not to Learn : An adversarial method to train deep neural networks with biased data

112 stars 13 forks source link

Selecting checkpoint for evaluation #1

Closed olliethomas closed 5 years ago

olliethomas commented 5 years ago

Hi, thanks for publishing this repository to accompany your CVPR paper. I'm trying to reproduce some of your results as I think it would be an excellent state-of-the-art benchmark for a project I'm working on, but I'm having some trouble training a new model to match the pre-trained checkpoint provided. Were the settings used the same as in the option.py file? Also, I notice that the benchmark checkpoint is from epoch 90, how did you decide to use this checkpoint? Thanks!

feidfoe commented 5 years ago

I'm sorry about my late response! Settings used for pre-training is the same. Since time cost for training is not the matter of interest, I like to train a network over sufficient epoch. The specific number of epoch is not very important, because the network would converge. If your algorithm does not require some early stopping criteria, I recommend you to do the same. Thanks.

olliethomas commented 5 years ago

Thanks!