How to optimize input variable?


#1

Hi there,

I’m writing a network structure which need an initialized input variable to be optimized.

The example code is below:

scale = mx.symbol.var(shape = (1,1,9), init = mx.initializer.Constant(1))
scale_input = mx.symbol.broadcast_mul(input, scale_exp)
… to some loss function for optimization

But when I load the resulting trained params, the variable scale are all still constant 1. I want the input variable also get the gradient and be optimized except the input training data.

Can someone instruct me how to do this?


#2

I think I solved it.

The variables actually get update, but its zero due to my network initialization to constant 1.

After I initialized it to some random small number, the variables can get update normally.