Course assignment. In this assignment, it tested the how pretraining network helps in general.
Pretrainin network is to do reconstruction given the original images with randomly removing some pixels.
python pretrain_tiny_imagenet.py
it saved the .pth file.
-
train from scratch
train_cifar.py
-
train based on pretrained weight. If we want to use pth/models_570.pth.
train_cifar.py --models_path pth/models_570.pth
-
train based on pretrained weight, but freeze the backbone weight.
train_cifar.py --models_path pth/models_570.pth --freeze_backbone True
-
train on part of full size data, eg 5%.
train_cifar.py --models_path pth/models_570.pth --freeze_backbone True --fraction 5