MakeSparseTF

Tensorflow function to make a tensor sparse (gradients supported)

The make_sparse op makes the input tensor sparse along its last dimension. make_sparse_grad op is used to compute the gradients for make_sparse.

Why the custom function?

There is no inbuilt support for making a tensor sparse in tensorflow . The top_k op of tensorflow can be used to achieve sparsity but does not support gradient computation.

This project reuses code for top_k to achieved the desired operation._

How to use?

Copy make_sparse.so, make_sparse_grad.so, and make_sparse_op.py to the working directory of your project.

Then in the python code, import the op

	from make_sparse_op import make_sparse

The op works as follows

	x_sparse = make_sparse(x, k)

where,

• x is input tensor along the last dimension of which the sparsity operation is applied.
• k is a scalar tensor which specifies the max number of non-zero values to keep (i.e. the desired sparsity).
• x_sparse is the output which is the sparse version of x.