README.md

dagnn_caffe_deploy

MatConvNet-DAGNN to Caffe converter

Initial EXPERIMENTAL version of a converter from a MatConvNet DagNN model to Caffe.

Only a limited subset of layer types is currently supported: dagnn.Conv, dagnn.ReLU, dagnn.Concat, dagnn.BatchNorm, dagnn.Sum, dagnn.Pooling, dagnn.LRN, dagnn.SoftMax, dagnn.Loss, dagnn.DropOut

Sucessfully tested with:

• imagenet-resnet-50-dag
• imagenet-matconvnet-alex
• imagenet-googlenet-dag

Classification tests with the converted version of imagenet-resnet-50-dag are the same than with the MSRA version of Resnet-50 caffe model.

Using:

• Caffe rc-3
• MatConvNet 1.0-beta23

Installation

Just copy the files in this repository within the MatConvNet folder.

Before using dagnn_caffe_deploy, remember to add MatConvNet's "matlab" folder to the path. MatCaffe must also be available.

Sample Usage

net = load('imagenet-resnet-50-dag.mat');          % Load the MatConvNet model
dagnn_caffe_deploy(net,'imagenet-resnet-50-dag');  % Run the converter speciiying the model
                                                   % and the prefix for the output files

The instructions above will produce the following files:

• imagenet-resnet-50-dag.caffemodel
• imagenet-resnet-50-dag.prototxt
• imagenet-resnet-50-dag_mean_image.binaryproto