Implementing pretrained network weights for classification ONLY

[SRQ]

I have a neural network that I want to implement in Caffe. It has some standard layers that I think will be easily implemented using google protobuf. However, the network is pretrained, i.e. I already have weights and bias values in a .txt file that I want to use to classify images.

Considering I will not have the ".caffemodel" file, how can I use the weights? Am I right to assume that this network can be implemented in google protobuf even though I just want to classify and not do any training. Please make suggestions and correct me where my approach is wrong.

[Jan]

My approach would probably be to define the network, load it in into pycaffe, also load your pretrained weights somehow. Put them into the layers (net.params['layername']), and save the network again in binary. And voila, you have a caffemodel.

Jan