Note: This converter has been adapted from code in Marc Bolaños fork of Caffe. See acks for code provenance.

This is intended to serve as a conversion module for Caffe models to Keras models. It only works with Ye Olde Caffe Classic™ (which isn't really a thing, but which probably should be a thing to prevent confusion with the Caffe 2).

Please be aware that this module is not regularly maintained. Thus, some layers or parameter definitions introduced in newer versions of either Keras or Caffe might not be compatible with the converter. Pull requests welcome!

In order to convert a model you just need the .caffemodel weights and the .prototxt deploy file. You will need to include the input image dimensions as a header to the .prototxt network structure, preferably as an Input layer:

layer {
  name: "image"
  type: "Input"
  top: "image"
  input_param {shape {dim: 1, dim: 3, dim: 128, dim: 128}}
}

Given the differences between Caffe and Keras when applying the max pooling operation, in some occasions the max pooling layers must include a pad: 1 value even if they did not include them in their original .prototxt.

The module caffe2keras can be used as a command line interface for converting any model the following way:

python -m caffe2keras models/train_val_for_keras.prototxt models/bvlc_googlenet.caffemodel keras-output-model.h5

To use the produced model from Keras, simply load the output file (i.e. keras-output-model.h5) using keras.models.load_model.

This code is yet another iteration of a tool which many people have contributed to. Previous authors:

• Marc Bolaños (email, Github)
• Pranav Shyam (Github)
• Antonella Cascitelli (Github)