This tool helps convert MXNet models into Apple CoreML format which can then be run on Apple devices.
In order to use this tool you need to have these:
pip install mxnet-to-coreml
Let‘s say you want to use your MXNet model in an iPhone App. For the purpose of this example, let’s assume it is a squeezenet-v1.1 model.
mxnet_coreml_converter.py --model-prefix='squeezenet_v1.1' --epoch=0 --input-shape='{"data":"3,227,227"}' --mode=classifier --pre-processing-arguments='{"image_input_names":"data"}' --class-labels synset.txt --output-file="squeezenetv11.mlmodel"
The above command will save the converted model in CoreML format to file squeezenet-v11.mlmodel. Internally, the model is first loaded by MXNet recreating the entire symbolic graph in memory. The converter walks through this symbolic graph converting each operator into its CoreML equivalent. Some of the supplied arguments to the converter are used by MXNet to generate the graph while others are used by CoreML either to pre-process the input (before passing it to the neural network) or to process the output of the neural network in a particular way.
In the command above:
You could provide a file containing class labels (as above) so that CoreML will return the category a given image belongs to. The file should have a label per line and labels can have any special characters. The line number of the label in the file should correspond with the index of softmax output. E.g.
mxnet_coreml_converter.py --model-prefix='squeezenet_v1.1' --epoch=0 --input-shape='{"data":"3,227,227"}' --mode=classifier --class-labels synset.txt --output-file="squeezenetv11.mlmodel"
You could ask CoreML to pre-process the images before passing them through the model. The following command provides image re-centering parameters for red, blue and green channel.
mxnet_coreml_converter.py --model-prefix='squeezenet_v1.1' --epoch=0 --input-shape='{"data":"3,224,224"}' --pre-processing-arguments='{"red_bias":127,"blue_bias":117,"green_bias":103}' --output-file="squeezenet_v11.mlmodel"
If you are building an app for a model that takes “Image” as an input, you will have to provide image_input_names as pre-processing arguments. This tells CoreML that a particular input variable is of type Image. E.g.:
mxnet_coreml_converter.py --model-prefix='squeezenet_v1.1' --epoch=0 --input-shape='{"data":"3,224,224"}' --pre-processing-arguments='{"red_bias":127,"blue_bias":117,"green_bias":103,"image_input_names":"data"}' --output-file="squeezenet_v11.mlmodel"
List of MXNet layers that can be converted into their CoreML equivalent:
Any MXNet model that uses the above operators can be converted easily. For instance, the following standard models can be converted:
mxnet_coreml_converter.py --model-prefix='Inception-BN' --epoch=126 --input-shape='{"data":"3,224,224"}' --mode=classifier --pre-processing-arguments='{"image_input_names":"data"}' --class-labels synset.txt --output-file="InceptionBN.mlmodel"
mxnet_coreml_converter.py --model-prefix='nin' --epoch=0 --input-shape='{"data":"3,224,224"}' --mode=classifier --pre-processing-arguments='{"image_input_names":"data"}' --class-labels synset.txt --output-file="nin.mlmodel"
mxnet_coreml_converter.py --model-prefix='resnet-50' --epoch=0 --input-shape='{"data":"3,224,224"}' --mode=classifier --pre-processing-arguments='{"image_input_names":"data"}' --class-labels synset.txt --output-file="resnet50.mlmodel"
mxnet_coreml_converter.py --model-prefix='squeezenet_v1.1' --epoch=0 --input-shape='{"data":"3,227,227"}' --mode=classifier --pre-processing-arguments='{"image_input_names":"data"}' --class-labels synset.txt --output-file="squeezenetv11.mlmodel"
mxnet_coreml_converter.py --model-prefix='vgg16' --epoch=0 --input-shape='{"data":"3,224,224"}' --mode=classifier --pre-processing-arguments='{"image_input_names":"data"}' --class-labels synset.txt --output-file="vgg16.mlmodel"