Issues with FNet

[gchhablani]

This is a tracking issue for problems/enhancements needed with FNet.

So far, @stefan-it @ontocord have mentioned that:

• --fp16 precision throws an error with the model.

• It is not possible to keep variable sequence length during training because of the way DFT matrices are initialized (via a config variable).

  • For this @ontocord suggested that we keep different buffers ready. A major drawback to this is that we can't possibly keep dft matrices for all sequence lengths. However, the values of the DFT matrices vary according to the total sequence length expected.
  • So, one efficient way of handling this is that we can try to create a generic matrix (not DFT) and then modify it on the fly accordingly based on the sequence length.
    • For example, the DFT matrix is defined as:
      
      • We can create a matrix without the multiplier, and then on the fly take the portion of the matrix that is needed for the batch, and then multiply with the correct multiplier. Wdyt @patrickvonplaten @sgugger?
      • Or multiply the matrix with sqrt(N)/sqrt(seq_length) and take mat[:seq_length, :seq_length] while multiplying.

• Need to verify if pushing to GPU pushes everything to the device (including buffers).

I will be adding more issues/problems here as and when they arise.

[patrickvonplaten]

I think if PyTorch doesn't support --fp16 for the fourier transform, we sadly can't do much here.
Regarding the second point, I think keeping a buffer would indeed be a good solution here!

[huu4ontocord]

I think you can keep N (maybe 5) fixed buffer, like 64, 128, 256, 512, 1024. You would need to pad for those buffers that are too long.

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[gchhablani]

Unstale

[patrickvonplaten]

Should we implement a buffer solution here? @gchhablani - do you think this makes sense? How big would the buffer be in terms of size?

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.