[theano-users] Cost function based on real eigenvalues

[Ming-Yee Tsang]

I am trying to build my model using a cost function based on eigenvalues of
a symmetric (or possibly Hermitian) matrix. I get the eigenvalues using
theano.tensor.nlinalg.eigh (and I tried theano.tensor.slinalg.eigvalsh with
similar results), but when I try to take the gradient of the cost with
respect to two shared variables, I get the error
TypeError: Eigh{UPLO='L'}.grad illegally returned an integer-valued
variable. (Input index 0, dtype complex128)
I verified that cost and the shared variables have type float64, so where
is the complex128 coming from?

[Ming-Yee Tsang]

I am on version 1.0.1.

Post by Ming-Yee Tsang
I am trying to build my model using a cost function based on eigenvalues
of a symmetric (or possibly Hermitian) matrix. I get the eigenvalues using
theano.tensor.nlinalg.eigh (and I tried theano.tensor.slinalg.eigvalsh with
similar results), but when I try to take the gradient of the cost with
respect to two shared variables, I get the error
TypeError: Eigh{UPLO='L'}.grad illegally returned an integer-valued
variable. (Input index 0, dtype complex128)
I verified that cost and the shared variables have type float64, so where
is the complex128 coming from?