Created by: Superjomn

When the variable that is both input and output of some operator(such as sigmoid), if it is contained in the fetch_list, the prune API in save_inference_model will fail to include that operator.

A real example:

y = fc(x, act='sigmoid')
# that will add a sigmoid after the fc, and because the sigmoid's input and output share the same variable
# when `prune`d, the framework will drop the sigmoid

Temporary fix

Add a scale op to force include the sigmoid like operator:

y = fc(x, act='sigmoid')
y0 = scale(y, 1)

that works!

Long-term fix

Delete the buggy prune method, save the original model, and in CPP inference, load that model in IR and prune it with an SSA graph.