Re: [tensorflow-compression] GDN Negative Parameters

[Johannes Ballé]

Hi Scott,

the parameters should not become negative, at least in the default implementation. We enforce this constraint through reparameterization, so the reparameterization might be negative, while the actual parameter isn't. In other words, if you look inside the tf.Variable that is trained, it might be negative, but if you evaluate `gdn_layer.beta` etc., it should return positive values.

Please let me know if this isn't consistent with what you observe.

Thanks!

Johannes 

On Tue, Jul 26, 2022 at 3:13 AM Scott Ransom <scott.l...@gmail.com> wrote:

Hi,

I noticed that training a model can lead to negative GDN parameters. In particular I've noticed this on the gamma parameter quite extensively, despite the code suggesting that it should be non-negative. Is this intended or a bug?

Best regards,

Scott Ransom

--
You received this message because you are subscribed to the Google Groups "tensorflow-compression" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tensorflow-compre...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/tensorflow-compression/111d6568-dc74-4212-80e0-a6d54a2d086fn%40googlegroups.com.