search

icoz69 / DeepEMD

Code for paper "DeepEMD: Few-Shot Image Classification with Differentiable Earth Mover's Distance and Structured Classifiers", CVPR2020
MIT License
577 stars 82 forks source link

hyperparameters for pre-training on TieredImageNet? #38

Closed zhanyuanyang closed 3 years ago

zhanyuanyang commented 3 years ago

What are the best hyperparameters for pre-training on TieredImageNet?

Should I keep the hyperparameters in the train_pretrain.py unchanged？

icoz69 commented 3 years ago

hi generally the parameters to adjust are lr scheduler, such as decay step and decay strength. I have recently found that using cosine lr scheduler without annealing is good as no hyper-parameters about lr scheduler is needed any more, and you only need to find a good start lr.