How to resolve shape mismatch in fine tuning?

[Ali]

I have a pretrained FCN model in which the shape of the first conv layer is:

(x, 3, x, x),

where these 3 channels are traditional RGB channels.

I want to fine tune this model using my own data set that has 2 extra channels, so my first conv layer has the shape of:

(x, 5, x, x).

I read somewhere that it is possible to load pretrained parameters for the first 3 channels and initialize those two extras, but there was no explanation how to do that.

Any useful suggestion will be appreciated. 

[Ali]

Actually, the net surgery is the answer.

You can find it here. 

Thanks to Shai.