About coding with CABAC in compression_Hyperprior

[zhang lu]

Hi  Johannes,

I recently read your ICLR18 paper "VARIATIONAL IMAGE COMPRESSION WITH A SCALE HYPERPRIOR" again, and I have a question about how to encode a symbol with CABAC and entropy model.

I could not find where we use the arithmetic coding and decoding in tensorflow_compression code. Someone tell me you used C languages to .so file and used it,but I did not find .so files.I want to know which are the AD's input and AE's output.

Very Very looking forward your reply!Thanks for your time.

Best,

zhanglu

[Johannes Ballé]

Hi Zhang Lu,

the arithmetic coder we are using is not based on CABAC. It's a standard multi-symbol range coder, and it has an overflow functionality for out-of-range values. Values outside of the range where we have probabilities, are encoded using a code similar to an exponential Golomb code.

The code is here:

https://github.com/tensorflow/compression/blob/master/tensorflow_compression/cc/kernels/range_coder.cc

The documentation for the corresponding TensorFlow op is here:

https://tensorflow.github.io/compression/docs/api_docs/python/tfc/unbounded_index_range_encode.html

https://tensorflow.github.io/compression/docs/api_docs/python/tfc/unbounded_index_range_decode.html

The .so files are not on Github. They show up in the binaries we provide on pypi (https://pypi.org/project/tensorflow-compression/2.2/#files) or if you compile the library yourself.

Hope this helps,

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/34adc07d-4b56-4d81-800b-996f6da66f23n%40googlegroups.com.

[zhang lu]

Hi  Johannes,

I recently read your ICLR18 paper "VARIATIONAL IMAGE COMPRESSION WITH A SCALE HYPERPRIOR" again, and I have a question about how to draw the bpp-PSNR curve and bpp--MS-SSIM.

How can I  change the value of bpp to get the value of PSNR or MS-SSIM of same picture? I would like to know how to get the line chart，the figure 5: Rate–distortion curves aggregated over the Kodak dataset.

Very looking forward your reply!Thanks for your time.

Best wishes,

zhanglu

[Johannes Ballé]

Hi,

the plots were generated by sweeping over the lambda parameter, training one model for each lambda. Then the bpp and the MS-SSIM were averaged over all images encoded with that lambda.

Hope this helps!

Johannes

To view this discussion on the web visit https://groups.google.com/d/msgid/tensorflow-compression/f89eae6c-d10f-48a5-8a23-977c6703ec85n%40googlegroups.com.

[zhang lu]

Hi，

Thanks for your reply.And how can I chose the lambda parameter to get all value of bpp from 0 to 2.0? and how many points I will plot?  

Best wishes，

ZhangLu

[Johannes Ballé]

Hi Zhang Lu,

the lambda values you need you need to determine experimentally. It is hard to know what rate vs. distortion a model will have ahead of time.

Best,

Johannes

To view this discussion on the web visit https://groups.google.com/d/msgid/tensorflow-compression/1bd827a9-47cc-4ab9-8115-ab783e5b70e7n%40googlegroups.com.

[zhang lu]

Hi Johannes,
Thank you very much for your reply！

And I have another problem about the articalVARIATIONAL IMAGE COMPRESSION WITH A SCALE HYPERPRIOR. If I will optimized for MS-SSIM,what the loss function I will use? I know the loss=λ*(1-D)+R,but how could I get the 'D'?

When I used the train_msssim=tf.squeeze(tf.image.ssim_multiscale(x_tilde,x,255)),the program reported an error.

How do i find this parameter？

Looking forward to your reply！

Best wishes,

ZhangLu

[zhang lu]

Hi Johannes,

And I have another problem about the artical VARIATIONAL IMAGE COMPRESSION WITH A SCALE HYPERPRIOR. If I will optimized for MS-SSIM,what the loss function I will use? I know the loss=λ*(1-D)+R,but how could I get the 'D'?

When I used the train_msssim=tf.reduce_mean(tf.image.ssim_multiscale(x_tilde,x,255)) to get the average,the result is not well. Is the loss function wrong or the value of λ is wrong?   Are the values of λ to optimized for MS-SSIM same as optimized for MSE?

Very Very looking forward your reply!Thanks for your time.

Best wishes,

ZhangLu

[Johannes Ballé]

Hi Zhang,

most likely you need to change the value of lambda, since it will be different from the MSE based model.

Johannes

To view this discussion on the web visit https://groups.google.com/d/msgid/tensorflow-compression/2e696b61-bdeb-428d-98ee-681d5eea4c0an%40googlegroups.com.