Reproducibility on CIFAR-10 about deepinversion

[slala2121]

I also tried inverting ResNet-34 with accuracy ~92% on CIFAR10, using the same parameters as in the paper. The loss has converged, the teacher network attains 100% accuracy on synthesized images, but the images don't appear to be comparable in quality (below).

Are there other changes that need to be made? Thanks.

Cat:

Dog:

Horse:

Here's a sample loss plots. Seems like there's a trade off between the other losses and l2 as the iterations increase. Is this typical?

history.pdf

[sccbhxc]

I also tried inverting ResNet-34 with accuracy ~92% on CIFAR10, using the same parameters as in the paper. The loss has converged, the teacher network attains 100% accuracy on synthesized images, but the images don't appear to be comparable in quality (below).

Are there other changes that need to be made? Thanks.

Cat:

Dog:

Horse:

Here's a sample loss plots. Seems like there's a trade off between the other losses and l2 as the iterations increase. Is this typical?

history.pdf

Could you share the code with me?

[hongxuyin]

Hi, the hypers in paper for CIFAR-10 are for scheme without multi-resolution, in setting_id 1 or 2. If networks/setting_id are different, a quick search for hyper will work: r_feature, lr, iterations, etc.

[pamolchanov]

We have updated the code with an example of DeepInversion for ResNet34 on CIFAR10. We did not optimize hyper parameters and there is a room to improve the visual quality of images.

[pamolchanov]

I also tried inverting ResNet-34 with accuracy ~92% on CIFAR10, using the same parameters as in the paper. The loss has converged, the teacher network attains 100% accuracy on synthesized images, but the images don't appear to be comparable in quality (below).
Are there other changes that need to be made? Thanks.
Cat:

Dog:

Horse:

Here's a sample loss plots. Seems like there's a trade off between the other losses and l2 as the iterations increase. Is this typical?
history.pdf

Could you share the code with me?

Please see the example in the folder cifar10