How to export weights in Matcaffe correctly

[Fábio Ferreira]

Hi,

I've been trying to export .caffemodel in Matcaffe by using "net.save(weights)" in function1 and then, initiating it in another function2 using "net=caffe.Net(deployFile, weights, 'test')" as described in the documentation for Matcaffe. However, I recognized that assigning the weights in function1, saving end reloading them in function2 provides a different outcome than simply assigning the same weights in function2 in the forward pass.

Is there a misuse in saving the weights file in my approach?
And is there a way to save a .caffemodel instead of a .mat file (which is produced by net.save())?

Thank you.

Best

[Fábio Ferreira]

Or put in other words: I don't train my net, but create weights manually by initiating them through accessing the layers one by one. After I've done that, how can I export my weights for, let's say, finetuning a net in a next step?

Thank you.