Value error: image size of 32800x65600 pixels is too large.

[LiuJinzhe-Keepgoing]

Very meaningful work, I have a question.
When I want to visualize the attention matrix of a very long prompt, the error message displays:
Value error: image size of 32800x65600 pixels is too large. It must be less than 2 16 in each direction.

It may be because there are too many words in each matrix to fit the pictures. Is there any good solution?

[wln20]

Thanks for your question. Here I plot the attention maps for all layers/heads in a single figure for clarity, and that would make the figure too large when the seqence becomes long. I suppose the problem could be solved by plotting the attention map for each layer/head separately (in multiple figures).

If there's still need for this feature, please let me know and I'll implement it at convenience!