Loading weights - what is stored and fetched ? about nolearn HOT 6 CLOSED

The implementation of NeuralNet.load_weights_from will try to copy weights from layers that match up; they should be in the same position, and have the same number of parameters. That's what the implementation does here: https://github.com/dnouri/nolearn/blob/master/nolearn/lasagne.py#L349

from nolearn.

Hmm - thanks. I had a network with 32,64,128 kernels in 3 convpool layers (followed by 3 dense layers). I loaded it into another network with 64,138,256 kernels (rest remaining the same). So ideally that should have been as good as a fresh start ? or only the dense layers will have a trained start while the rest have a fresh start ?

from nolearn.

Yes, your dense layers will be initialized with those weights. I think @BenjaminBossan has code that adds a little bit of feedback to load_weights_from, so that as a user, you'll know which weights were copied and which weren't.

from nolearn.

Thanks

from nolearn.

I have the code here:
https://github.com/BenjaminBossan/nolearn
Just take that code or copy from the lasagne.py module. There are no tests yet, I hope it works. It did for me.

I would have made a pull request but I don't exactly know how to yet and am short on time. Good luck.

from nolearn.

Thanks Benjamin

from nolearn.