Purpose and implementation of 're-flatten' Layer in Vision Transformer

In MATLAB 2023b's Computer Vision Toolbox, there is a pre-trained vision Transformer model that includes patch embedding and re-flatten layers. What is the purpose of the re-flatten layer? Additionally, it seems that there is no "re-flatten" function in the Deep Learning Toolbox. Do I need to write this function myself?

[xingxingcui]

I only see patch embedding layer,flatten layer,can't find re-flaten layers in MATLAB 2023b,can you post the re-flaten layer link?

[HK ZHOU]

Thanks for your comment.

After executing the following codes, in the NetworkAnalyzer, the 're-flatten' function is visible, but cannot find the source code.

net = visionTransformer;

analyzeNetwork(net)

Answer 1

[Tijue]

Hello Zhou,

You can open the "visionTransformer" function and find the "re-flatten" layer at line 67.