search

OmidPoursaeed / Generative_Adversarial_Perturbations

Generative Adversarial Perturbations (CVPR 2018)
https://arxiv.org/abs/1712.02328
135 stars 22 forks source link

Training time? #5

Closed hcguoO0 closed 4 years ago

hcguoO0 commented 5 years ago

Hi, I try to run your code with a subset of Imagenet dataset about 8000 pictures to generate non-targeted imdep perturbation, but it takes so long time, about 10 hours. I wonder how long time do you spend to train such model ?

OmidPoursaeed commented 5 years ago

Hello, are you setting --MaxIter correctly (to run on a subset of the training set)? How many epochs are you using for training? Also, please make sure you are running it on GPU.

hcguoO0 commented 5 years ago

Thank you for your answer, I have another question, is the weights_init and the torch.log(the loss function) necessary?

OmidPoursaeed commented 5 years ago

Initializing the weights (weights_init) helps with better convergence. We found that using log helps, but it's not necessary.

hcguoO0 commented 4 years ago

Thanks for your response.