Is there a way to download ImageNet weights for WRN-28-2?

[varunnair18]

I see that the repository provides a method to download ImageNet in a 32x32 format, however, would you have access to the weights for the WRN-28-2 model that was trained on ImageNet and used in the transfer learning section of the paper?

I would like to use the ImageNet weights in another application, but if you had a link or another method for me to download them it would save much time and cost in training the network on ImageNet from scratch.

Thanks!

[craffel]

No. Note that since it is Imagenet 32x32, pre-training on imagenet is not really any more expensive than any of the other semi-supervied training runs.

[varunnair18]

@craffel Would you know approximately how long (using a single GPU) I should expect the model to train for on the ImageNet 32x32 to achieve the same model weights that you use for the transfer learning experiments?

[craffel]

@DoctorTeeth might know.

[TheRevanchist]

It took me around half a day to reproduce their ImageNet results (using their code) in a Titan Xp.

[DoctorTeeth]

Oh gosh, it was a long time ago. I want to say that around a day sounds right. I don’t remember having to wait longer than that for the imagenet model to finish training.

On Fri, Oct 18, 2019 at 3:28 PM Ismail Elezi notifications@github.com wrote:

It took me around half a day to reproduce their ImageNet results (using their code) in a Titan Xp.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/brain-research/realistic-ssl-evaluation/issues/28?email_source=notifications&email_token=AABQ6VM5ROXT6NN5VQF3SATQPI2BFA5CNFSM4H5D3L62YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEBWIHMY#issuecomment-543982515, or unsubscribe https://github.com/notifications/unsubscribe-auth/AABQ6VP2FWHGVJ645C3D34LQPI2BFANCNFSM4H5D3L6Q .

[craffel]

Yeah a day-ish sounds right. This would have been on a k80.

[varunnair18]

Thanks for that, also while trying to use the training command in /runs/section-4-3-imagenet-pre-training.yml, I found that I didn't have the zca_mean.npy file necessary to perform the ZCA whitening. Is there a command that I need to run to generate it or somewhere online I can download it?

I was able to successfully follow the instruction for creating tfrecords for ImageNet, but didn't get this part.

[craffel]

You need to generate the CIFAR-10 records to get the ZCA files. They are computed over the CIFAR-10 dataset.