search

jiangtaoxie / fast-MPN-COV

@CVPR2018: Efficient unrolling iterative matrix square-root normalized ConvNets, implemented by PyTorch (and code of B-CNN,Compact bilinear pooling etc.) for training from scratch & finetuning.
http://peihuali.org/iSQRT-COV/index.html
MIT License
270 stars 56 forks source link

what are suggested training parameters for CUB_200_2001? #16

Closed wjtan99 closed 5 years ago

wjtan99 commented 5 years ago

I tried lr=0.1,0.01, 0.001, batch=16, after about 10 epochs, it runs into significant over fitting. what are suggested training parameters for CUB_200_2001 to achieve top1 acc = 0.88? Thanks a lot.

wjtan99 commented 5 years ago

BTW, I started new training, not fine tuning from your model.

jiangtaoxie commented 5 years ago

@wjtan99 I have no idea about achieving top1 acc = 0.88 when train a model from scratch on CUB. The #sample of CUB is too small to fit model with billions of parameters.

wjtan99 commented 5 years ago

@jiangtaoxie Hi, I of course still started from Imagenet pretrained model. Then how did you get acc=0.88 on CUB_200_2001? Thanks.

wjtan99 commented 5 years ago

I got it wrong. The script trainingFromScratch is really training from scratch without loading pre-trained model. I switched to use finetune.sh and load VGG16_BN, now the results look much better.
Thanks for your nice work.