From 3b96f61dcc2b18cd581defb2b56528889242d66f Mon Sep 17 00:00:00 2001
From: Manuel Costa <manuel@nativolabs.com>
Date: Wed, 28 Nov 2018 15:15:13 +0000
Subject: [PATCH] Create Convert2CoreML.py

---
 code/Convert2CoreML.py | 93 ++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 93 insertions(+)
 create mode 100644 code/Convert2CoreML.py

diff --git a/code/Convert2CoreML.py b/code/Convert2CoreML.py
new file mode 100644
index 0000000..34e7984
--- /dev/null
+++ b/code/Convert2CoreML.py
@@ -0,0 +1,93 @@
+#reference https://gist.github.com/viteinfinite/1901d9cb6d26c21967a08aa5534e05c2#file-convert-py
+#input and outputs are scaled according to https://github.com/nyoki-mtl/keras-facenet/blob/master/notebook/demo-webcam.ipynb
+#input is normalized per image through average and standard deviation 
+#output is done through L2 normalization
+
+import copy
+import coremltools
+from coremltools.proto import NeuralNetwork_pb2
+from coremltools.models.neural_network.quantization_utils import *
+import numpy
+from keras.models import Sequential
+from keras.layers import Dense
+from keras.layers import Dropout
+from keras.utils import np_utils
+from keras.models import load_model
+import os.path
+import sys
+
+
+#sys.path.append('/code/') # Import "code" from https://github.com/nyoki-mtl/keras-facenet 
+
+from inception_resnet_v1 import *
+
+#model source https://drive.google.com/drive/folders/1pwQ3H4aJ8a6yyJHZkTwtjcL4wYWQb7bn
+
+model = InceptionResNetV1(weights_path='facenet_keras_weights.h5')
+
+def convert_lambda(layer):
+    if layer.function == scaling:
+        params = NeuralNetwork_pb2.CustomLayerParams()
+        params.className = "scaling"
+        params.parameters["scale"].doubleValue = layer.arguments['scale']
+        return params
+    else:
+        return None
+
+coreml_model = coremltools.converters.keras.convert(
+    model,
+    input_names="image",
+    image_input_names="image",
+    output_names="output",
+    add_custom_layers=True,
+    model_precision='float16',
+    custom_conversion_functions={ "Lambda": convert_lambda })
+
+
+spec = coreml_model.get_spec()
+
+# get NN portion of the spec
+nn_spec = spec.neuralNetwork
+layers = nn_spec.layers # this is a list of all the layers
+layers_copy = copy.deepcopy(layers) # make a copy of the layers, these will be added back later
+del nn_spec.layers[:] # delete all the layers
+
+# add a scale layer now
+# since mlmodel is in protobuf format, we can add proto messages directly
+# To look at more examples on how to add other layers: see "builder.py" file in coremltools repo
+image_norm_layer = nn_spec.layers.add()
+image_norm_layer.name = 'image_norm_layer'
+image_norm_layer.input.append('image')
+image_norm_layer.output.append('image_normalized')
+
+params = image_norm_layer.mvn
+params.acrossChannels = False
+params.normalizeVariance = True
+params.epsilon = 1e-6
+
+# now add back the rest of the layers (which happens to be just one in this case: the crop layer)
+nn_spec.layers.extend(layers_copy)
+
+# need to also change the input of the crop layer to match the output of the scale layer
+nn_spec.layers[1].input[0] = 'image_normalized'
+
+nn_spec.layers[len(nn_spec.layers)-1].output[0] = 'embeddings_raw'
+
+embeddings_norm_layer = nn_spec.layers.add()
+embeddings_norm_layer.name = 'embeddings_norm_layer'
+embeddings_norm_layer.input.append('embeddings_raw')
+embeddings_norm_layer.output.append('output')
+
+params =  embeddings_norm_layer.l2normalize
+params.epsilon = 1e-6
+#nn_spec.layers[len(nn_spec.layers)-2].input[0] = 'embeddings'
+
+
+
+
+
+coreml_model = coremltools.models.MLModel(spec)
+
+spec = coreml_model.get_spec()
+
+coremltools.utils.save_spec(spec, 'facenet_keras_weights_coreml.mlmodel')