How to run non-Caffe model in Caffe

Question

I want to run inference using some ConvNet on Caffe. The only issue is I already have the weights and bias in raw format and I don't want to re-train it on Caffe.

Now Caffe requires two input file:

• .prototxt (for Network info)
• .caffemodel (for weights and bias)

Here .prototxt file can be generated easily, but is it possible to generate .caffemodel using raw weights and bias?

Answer 1

The process of manually manipulating a Caffe model is called Net Surgery.

Specifically, net.params() allows you to access the net's parameters directly.