Measure the size of compressed image

[danishnazir]

Hi All,

I just wanted to ask how can we get exact file size of the compressed image? For example if an image x is passed to a pre-trained model, which produces the latent space at some fixed bpp, then how can we measure the size of the compressed image itself? Do we save the actual bits (in the form of strings) produced by the entropy encoder? Can someone please point me to the right direction?

Regards,

[YodaEmbedding]

Two methods, which we use in the given [Code] links:

• [Code]. During training, we calculate the negative log likelihoods via:

   bpp_loss = sum( -log2(likelihoods) ) / num_pixels
  

  The likelihoods representing $p_{\hat{y}}(\hat{y})$ are determined by feeding a monotonic differentiable function $c : \mathbb{R} \to [0, 1]$ (i.e., the CDF) with a single latent tensor element $\hat{y}_i$. Further details can be found in the Appendix of "Variational image compression with a scale hyperprior".

• [Code]. During inference, when using a real entropy coder,

  bpp = len(bytestring) * 8 / num_pixels
  

If measured on the exact same $\hat{y}$, the aforementioned bpp_loss and bpp should be very, very close to each other, for a well-trained model that generates reasonable distributions from reasonable input images.

Please follow the training tutorial if you want to try this out on your own custom models.

---

Related:

• How to measure compressed video size #191 (comment)
• About Rate Distortion Loss #199 (comment)

[RizerKayseri]

Hello, during the process of using, I encountered something quite confusing. In the case where num_pixels equals 1024, the calculated bpp_loss consistently remains around 0.9. However, for the length of the bitstring (len(bitstring)), its result always approximates to about 120. This means that when I actually compress data, the bpp (bits per pixel) obtained by calculating bpp = len(bitstring) / num_pixels is significantly smaller than the bpp_loss.

[YodaEmbedding]

len_bitstring = 8 * len(bytestring)

I've updated my earlier post for clarity.

[RizerKayseri]

Thank you very much

[RizerKayseri]

I made a foolish mistake.

[danishnazir]

Thank you for the explaination @YodaEmbedding , just wanted to double check my understanding: The size of bpp in Bytes is the actual (during testing ofcourse) filesize on any input x?

[YodaEmbedding]

The file size in bytes of the compressed image is bpp * num_pixels / 8.