search

PingoLH / FCHarDNet

Fully Convolutional HarDNet for Segmentation in Pytorch
MIT License
195 stars 52 forks source link

pretrained model: number of epochs? #2

Closed trougnouf closed 4 years ago

trougnouf commented 4 years ago

What is the number of epochs for your pretrained model? Are the parameters those listed in your config files and default args? Thank you.

PingoLH commented 4 years ago

Hello, thanks for asking. Yes, the hyperparameters for training FCHarDNet-70 were all listed in config/hardnet.yml. You can also add color jittering augmentation to improve test-set accuracy. For the ImageNet pretrained model, it was trained with lr=0.1, batch size=512, weight decay=1e-4. The training time was a little bit tricky: Max epoch=150, but early stopped at epoch 110~120, since I noticed that there is no big difference for FCHarDNet training with a base model that is fully pretrained or not.

trougnouf commented 4 years ago

Thank you @PingoLH !