Vision Transformer

I think it would be interesting to visualize how the vision transformer works by splitting an image into a bunch of patches.

from manimml.

Self-Attention

This is the most important component (imo) to visualize property in the Transformer Architecture. I can think of two levels of visualization for this.

In-depth visualization

This visualization will show (1) the Key, Value, and Query feed-forward layers, (2) the matrices returned by these layers that are then multiplied, (3) the softmax operation combining the Key and Query into a score (4) the linear combination of the values into final values.

A high-level conceptual visualization

This is the layer that I think should be a NeuralNetworkLayer.

It should take in either text (broken down into tokens), an image (broken into patches), or vectors (output of a feed forward layer). These should then be passed into a self-attention layer. This layer should put the tokens (whatever type) onto the left and top side of a matrix visualization. The matrix visualization should be a 2D heatmap of the softmaxed (normalized) attention scores. Finally, the scores should be combined with the values to form the output of the self-attention module.

ImageToPatches

I will need to make a layer for splitting up an image into patches. The patches are necessary to represent the image as a sequence.

from manimml.