The train_model.ipynb code in DFTND can not train a good backdoor model on ResNet-18 (CIFAR10)? #6
Description
In the train_model.ipynb code, there is no transforms on CIFAR10 images as in some other works. It seems the step size scheduler is also not used (the learning rate is set as 1e-2). I try to use this code to train a ResNet-18 model on CIFAR10 with square trigger (by changing the json file), but I could not get either a good accuracy on clean images or a good accuracy on poisoned images. I am wondering how to train a ResNet-18 backdoor model on CIFAR10 with the train_model.ipynb code?
Many thanks for your help in advance!