How to find norm of a sparse matrix in mxnet ? Please note that is should not be the normal norm operation in ndarray it should be for sparse matrices, is there any support as such in mxnet ?
There is no
norm() operator for sparse NDArray, however you can calculate it yourself using
sqrt(). However be careful with sum because it is only implemented for CSR sparse array (for row-sparse it falls back to NDArray and has to convert the storage to a dense one). Something like the following code should do:
def csr_norm(csr_array): assert isinstance(csr_array, mx.nd.sparse.CSRNDArray), "array is not CSRNDArray" return csr_array.square().sum(0).sum().sqrt()
Actually there’s an implementation sitting in my fork but I don’t have the time to PR it yet.
This implements the norm operator, which can be used to inspect row sparse weight/gradient