SGD_op: Learning rate should be tensor of float instead of float

[abhinavarora]

Learning rate should be tensor of float instead of float. This is because a standalone float variable will not be be accessible on the GPU.