Converting an MXNet Style Transfer Model


Note that OpenVINO support for Apache MXNet is currently being deprecated and will be removed entirely in the future.

This article provides instructions on how to generate a model for style transfer, using the public MXNet neural style transfer sample.

Step 1: Download or clone the repository Zhaw’s Neural Style Transfer repository with an MXNet neural style transfer sample.

Step 2: Prepare the environment required to work with the cloned repository:


Python-tk installation is needed only for Linux. Python for Windows includes it by default.

  1. Install packages dependency.

    sudo apt-get install python-tk
  2. Install Python requirements:

    pip3 install --user mxnet
    pip3 install --user matplotlib
    pip3 install --user scikit-image

Step 3: Download the pre-trained VGG19 model and save it to the root directory of the cloned repository. The sample expects the model vgg19.params file to be in that directory.

Step 4: Modify source code files of style transfer sample from the cloned repository:

  1. Go to the fast_mrf_cnn subdirectory.

    cd ./fast_mrf_cnn
  2. Open the file and modify the decoder_symbol() function. You should see the following code there:

    def decoder_symbol():
        data = mx.sym.Variable('data')
        data = mx.sym.Convolution(data=data, num_filter=256, kernel=(3,3), pad=(1,1), stride=(1, 1), name='deco_conv1')

    Replace the code above with the following:

    def decoder_symbol_with_vgg(vgg_symbol):
        data = mx.sym.Convolution(data=vgg_symbol, num_filter=256, kernel=(3,3), pad=(1,1), stride=(1, 1), name='deco_conv1')
  3. Save and close the file.

  4. Open and edit the file. Go to the __init__() function in the Maker class:

    decoder = symbol.decoder_symbol()

    Modify it with the following code:

    decoder = symbol.decoder_symbol_with_vgg(vgg_symbol)
  5. To join the pre-trained weights with the decoder weights, make the following changes: After the code lines for loading the decoder weights:

    args = mx.nd.load('%s_decoder_args.nd'%model_prefix)
    auxs = mx.nd.load('%s_decoder_auxs.nd'%model_prefix)

    Add the following line:

  6. Use arg_dict instead of args as a parameter of the decoder.bind() function. Find the line below:

    self.deco_executor = decoder.bind(ctx=mx.gpu(), args=args, aux_states=auxs)

    Replace it with the following:

    self.deco_executor = decoder.bind(ctx=mx.cpu(), args=arg_dict, aux_states=auxs)
  7. Add the following code to the end of the generate() function in the Maker class to save the result model as a .json file:'{}-symbol.json'.format('vgg19'))'{}-symbol.json'.format('nst_vgg19'))
  8. Save and close the file.

Step 5: Follow the instructions from the file in the fast_mrf_cnn directory of the cloned repository and run the sample with a decoder model. For example, use the following code to run the sample with the pre-trained decoder weights from the models folder and output shape:

import make_image
maker = make_image.Maker('models/13', (1024, 768))
maker.generate('output.jpg', '../images/tubingen.jpg')

The models/13 string in the code above is composed of the following substrings:

  • models/ – path to the folder that contains .nd files with pre-trained styles weights.

  • 13 – prefix pointing to the default decoder for the repository, 13_decoder.


If an error prompts with “No module named cPickle”, try running the script from Step 5 in Python 2. After that return to Python 3 for the remaining steps.

Any style can be selected from collection of pre-trained weights. On the Chinese-language page, click the down arrow next to a size in megabytes. Then wait for an overlay box to appear, and click the blue button in it to download. The generate() function generates nst_vgg19-symbol.json and vgg19-symbol.json files for the specified shape. In the code, it is [1024 x 768] for a 4:3 ratio. You can specify another, for example, [224,224] for a square ratio.

Step 6: Run model conversion to generate an Intermediate Representation (IR):

  1. Create a new directory. For example:

    mkdir nst_model
  2. Copy the initial and generated model files to the created directory. For example, to copy the pre-trained decoder weights from the models folder to the nst_model directory, run the following commands:

    cp nst_vgg19-symbol.json nst_model
    cp vgg19-symbol.json nst_model
    cp ../vgg19.params nst_model/vgg19-0000.params
    cp models/13_decoder_args.nd nst_model
    cp models/13_decoder_auxs.nd nst_model


    Make sure that all the .params and .json files are in the same directory as the .nd files. Otherwise, the conversion process fails.

  3. Run model conversion for Apache MXNet. Use the --nd_prefix_name option to specify the decoder prefix and input_shape to specify input shapes in [N,C,W,H] order. For example:

    mo --input_symbol <path/to/nst_model>/nst_vgg19-symbol.json --framework mxnet --output_dir <path/to/output_dir> --input_shape [1,3,224,224] --nd_prefix_name 13_decoder --pretrained_model <path/to/nst_model>/vgg19-0000.params
  4. The IR is generated (.bin, .xml and .mapping files) in the specified output directory, and ready to be consumed by the OpenVINO Runtime.