Question about your PCS 2021 paper "3D SCENE COMPRESSION THROUGH ENTROPY PENALIZED NEURAL REPRESENTATION FUNCTIONS"

[Linfeng Zheng]

Hi,

Thanks for the paper 3D SCENE COMPRESSION THROUGH ENTROPY PENALIZED NEURAL REPRESENTATION FUNCTIONS.

In your paper, you mentioned that "Note that the parameters of this MLP can be discarded after training (once the probability mass functions have been built). "

However, doesn't it need for the entropy decoder to get the output of the pmf estimator MLP to decode the binary codes? Did I understand it wrong?

Thank you!

Best regards,

Linfeng

[Johannes Ballé]

Hi Linfeng,

there's an MLP that is used to implement the probability density model that provides the gradients for optimization (used in this distribution class: https://tensorflow.github.io/compression/docs/api_docs/python/tfc/DeepFactorized.html). Once training is finished, this MLP is used to create the PMF/CDF tables for a range coder (done in this class: https://tensorflow.github.io/compression/docs/api_docs/python/tfc/ContinuousBatchedEntropyModel.html). Once that is done, the MLP parameters can be discarded, because for encoding or decoding, all that's necessary are the tables.

I hope this clarifies it!

Best,

Johannes.

--
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/ad921251-2e3d-450e-b106-75597b320a27n%40googlegroups.com.

[Linfeng Zheng]

Hi Johannes,

Thanks for your clear explanation!

The MLP is used to create the PMF/CDF tables after training. Therefore, only the PMF/CDF tables are needed to save, which have a small size and do not have a big impact on total bitrates. Is that right?

Best,

Linfeng

[Johannes Ballé]

Yes, that is correct.

To view this discussion on the web visit https://groups.google.com/d/msgid/tensorflow-compression/e5e89e93-8a37-4ea1-8262-7e6c58df6e3dn%40googlegroups.com.