truncation trick

[KK666-AI]

Dear author,

I am reading your implementation on latent sampling from sample.py (function: sample_latents). For a gaussian sampling as implemented by latents = torch.randn(batch_size, dim, device=device)/truncated_factor.

I notice that the above implementation is not a standard truncation trick, which is defined by

The Truncation Trick is a latent sampling procedure for generative adversarial networks, where we sample from a truncated normal (where values which fall outside a range are resampled to fall inside that range)

[mingukkang]

Thank you very much.

I will reflect your reporting ASAP.

Best,

Minguk

[KK666-AI]

Let me know if you have fixed it.

[mingukkang]

Hi,

Thank you so much.

I have corrected the wrong implementation of truncation trick.

Please refer to "src/utils/sample.py" for more details.

[KK666-AI]

values = truncnorm.rvs(-threshold, threshold, size=size) is not defined as seen in line ?

[KK666-AI]

also, i think using latents_eps = (1-perturb)*latents + perturb*sample_normal(batch_size, dim, -1.0, device) in line should be more reasonable, because such definition maintains a realistic distribution, whose cumulative distribution function is 1.

What do you think?

[mingukkang]

Hi,

Thank you for your suggestion.

I think that the method you suggested is reasonable.

However, I already conducted BigGAN-Mod/ContraGAN + ICR experiments using the "latents + pertub*sample_normal" way.

We have also decided to keep StudioGAN code in its current state in order to exactly implement the original ICR regularization.

Nice suggestion and thank you.

Best,

Minguk

[KK666-AI]

Yes. I see. The experimental results on large-scale image datasets are very expensive.

[MTandHJ]

Hi,

It seems that the truncated_normal returns a ndarray rather than a tensor.

[KK666-AI]

@MTandHJ I report the similar bug, as seen here

[mingukkang]

Thank you.

I have fixed the problem above.

Best,

Minguk

[KK666-AI]

@mingukkang this function may want to return latents_eps ? see details in this line.