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:
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:
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.
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