How to manually change pre-trained CNN's (AlexNet) weights?

To whom can help:

I am using MATLAB r2017a. I am trying to change the weights of a pre-trained CNN, like alexnet, using the following codes. Unfortunately, I got the following errors. Are there any suggestions to by-pass this read-only constrain?

Thanks a lot! Jason

" You cannot set the read-only property 'Layers' of SeriesNetwork"

    % code
    >> net=alexnet
    >> net.Layers(2,1).Weights=net.Layers(2,1).Weights*10;

Answer 1

[J]

Bump! I have a hard time believing that one cannot take an existing SeriesNetwork, and either:

1. Modify properties of it, or...

2. ... create a new SeriesNetwork that is a copy of the existing one, but has different values for certain properties

Also, I would image there is a way to take an existing SeriesNetwork, then addLayers/removeLayers/connectLayers and convert back to a SeriesNetwork, rather than being stuck with the DAGNetwork that is generated (if it really is naturally represented as SeriesNetwork)? Or if not, that there is some workaround?

I mean, it makes NO SENSE that if, for example, I have a SeriesNetwork with a dropoutLayer at default 0.5 that I can't just modify that to 0.25 or 0.6 or whatever I want and then re-train without having to convert to a DAGNetwork and then be stuck with a DAGNetwork (certain functions will only operate on SeriesNetwork--take for example, deepDreamImage).

Answer 2

[mizuki]

In R2017a, the only supported way to create a SeriesNetwork is to use trainNetwork, and it is not possible to modify it.

[Tom Hope]

OK - that seems a pretty strict limitation. What about 2017b?

Answer 3

[DL]

You can convert the DAG net to an object and then change that object with is a structure, then load that back. It is a possible way.