a question for defining Custom Deep Learning Layer

Hi,

I want to know if we can directly use the fucntion SVD to achieve the singular vectors and Singular values in the defined cumtome deep learning layer.

Looking forward to your reply

Answer 1

[Christine Tobler]

You can use the svd in a custom layer, however, if the SVD is used in the forward method of a custom layer, this will likely require the backward method of that layer to compute what amounts to the derivative of this SVD, which could be quite hard to compute.

[li]

Thanks very much for your answer. I will test it by myself.

[Damien T]

See this reference (Section 3.2) for the derivatives of the SVD:

An extended collection of matrix derivative results for forward and reverse mode algorithmic differentiation (PDF file)