backnotprop

private func stylizeImage(cgImage: CGImage, model: MLModel) -> CGImage {
    
    // size can change here if you want, remember to run right sizes in the fst evaluating script
    let input = StyleTransferInput(input: pixelBuffer(cgImage: cgImage, width: 883, height: 720))

    // model.prediction will run the style model on input image
    let outFeatures = try! model.prediction(from: input)
  
    // we get the image buffer after
    let output = outFeatures.featureValue(for: "add_37__0")!.imageBufferValue!

backnotprop

//
//  StyleTransferInput.swift
//  StyleTransfer
//

import CoreML

class StyleTransferInput : MLFeatureProvider {
    
    var input: CVPixelBuffer

backnotprop

private let models = [
    wave().model,
    udnie().model,
    rain_princess().model,
    la_muse().model
]

backnotprop

import coremltools

def convert_multiarray_output_to_image(spec, feature_name, is_bgr=False):  
    """  
    Convert an output multiarray to be represented as an image  
    This will modify the Model_pb spec passed in.  
    Example:  
        model = coremltools.models.MLModel('MyNeuralNetwork.mlmodel')  
        spec = model.get_spec()  
        convert_multiarray_output_to_image(spec,'imageOutput',is_bgr=False)  

backnotprop

import tfcoreml as tf_converter
tf_converter.convert(tf_model_path = 'output_graph.pb',
                     mlmodel_path = 'model_name.mlmodel',
                     output_feature_names = ['add_37:0'],
                     ## Note found this after running a conversion the first time
                     image_input_names = ['img_placeholder__0'])   

backnotprop

# tfcoreml src

# file1 : _interpret_shapes.py
#
# in the _SHAPE_TRANSLATOR_REGISTRY we need to add the Pow operation
_SHAPE_TRANSLATOR_REGISTRY = {
     ... previous keys ...
    # add this:
    'Pow': _identity,
}

backnotprop

# function ffwd, line 98
# https://github.com/lengstrom/fast-style-transfer/blob/master/evaluate.py#L98
if os.path.isdir(checkpoint_dir):
    ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
    if ckpt and ckpt.model_checkpoint_path:
        saver.restore(sess, ckpt.model_checkpoint_path)
        ########## add this for pre-trained models ###########
        frozen_graph_def = tf.graph_util.convert_variables_to_constants(sess,sess.graph_def,['add_37'])
        with open('output_graph.pb', 'wb') as f:
            f.write(frozen_graph_def.SerializeToString())

backnotprop

# function ffwd,line 93
# https://github.com/lengstrom/fast-style-transfer/blob/master/evaluate.py#L93
preds = transform.net(img_placeholder)
# !! Add This !!
print(preds)

backnotprop

/**
 * Cyclic Elimination Stage 3
 * - description needed
 */
function _cycleReduceStage() {
    let stable = false;
	// all or nothing phase
    while(!stable) {
	  // there is a start if any one person has more than 1 preference remaining
      let start = indexWithMultipleRemain();

backnotprop

/**
 * Elimination Stage 2
 * - removes all preferences in a list who can not be possibly matched
 */
function _eliminateStage() {
    _.forIn(_DB, (person,id) => {
      let keepLast = _.findIndex(person.choices, function(p) { return p.id == person.acceptedReceivedID; });
      for(let i = keepLast + 1; i < person.choices.length; i++) {
	      // each the rejected and rejecter can remove each other from choices list
        eliminateChoices(person, _DB[person.choices[i].id]);