vislearn / ib-inn Goto Github PK

Hi - thanks for sharing your code! Quick conceptual question: I believe the architectures in this work do split off their latent codes using intermediate output nodes (unlike what's recommend in the FrEIA repo). I'm wondering why this was not done? Is it not really necessary?

https://github.com/VLL-HD/IB-INN/blob/440b75d5fbafec75842b18c9e7b6f03a8d76d16d/dct_transform.py#L117

It should be

transf = dct_layer([x])

Hi, thanks for a very appreciated code base.

It seems, that some code is missing from the repo.
In particular, the generate argument in main.py doesn't work due to an import of a not existing evaluation.generation function.

Is that, by chance, something that you forgot to commit?

Thank you in advance!

Since FrEIA went through few major updates, this repository do not work with the latest FrEIA.

Are there any quick fix for forward compatibility?

For the time being, I am using the version of FrEIA with framework.py in the root directory (i.e. https://github.com/VLL-HD/FrEIA/tree/550257b10af7d8772b08d4aa9b18772e2c0225f9)

Hi,

There may be a mistake in the reported bits per dimension for your CIFAR10 experiments.
In particular, it seems like they are computed by this code, which either accounts for normalization (dividing by 256) or adding "sigma noise".

Code copy-pasted here:

bpd = output['L_x_val'].item()
if eval(args['data']['dequantize_uniform']):
    bpd += np.log(256)
else:
    bpd += np.log(data.sigma) + 0.5 * np.log(2 * np.pi)
bpd -= 0.5 * np.log(2*np.pi)
bpd /= np.log(2)

However, it seems from the config files that both cfg['data']['dequantize_uniform'] and cfg['data']['sigma_noise'] are turned on at once.

Please correct me if I am wrong.

Once again, thanks for a very well done code base.

Hi,
This paper is a great work. Thanks for the codes.
I would like to enquire about the line 193 in VIB.py.
Why is the variable ‘‘statistics’’ subtract 5 ?
Thanks