How to define trainable weights in Flux

I want to define a network mode with W=[w11 0 w13; 0 w22 0; w31 0 w33],
i.e. set some weights to fixed number and they are not trained,

for example

m3 = Chain(Dense(3,3), Dense(3,2))

how to set m3.layer[1].weights, and define which weight is trainable or not?

Thank you very much!