Closed chihkuanyeh closed 6 years ago
Hi @yankeesrules
During the training weights were always overwritten with the weights from the last iteration (to make it possible to recover the training process from the last saved point) We observe that for big models (k=5) with our regularization after the last learning rate drop model starts to overfit. In the report we put results after 31 training epochs for each setting.
If you are interested, I uploaded weights after 40 epochs of training: https://drive.google.com/drive/folders/1SuTz4MZqEeU2U21pscQWvy45eTjusiWR?usp=sharing
Hope it will help, in case it is not compatible with the current github code write me back
Hi @yankeesrules Does it work?
Will the author kindly release pretrained weights for the Imagenet64?