def predictor_from_cntk_model(modelFile, plotModel=False):
"""Loads a CNTK model and returns an ell.neural.NeuralNetworkPredictor"""
_logger = logger.get()
_logger.info("Loading...")
z = load_model(modelFile)
_logger.info("\nFinished loading.")
if plotModel:
filename = os.path.join(os.path.dirname(modelFile), os.path.basename(modelFile) + ".svg")
cntk_utilities.plot_model(z, filename)
_logger.info("Pre-processing...")
modelLayers = cntk_utilities.get_model_layers(z)
# Get the relevant CNTK layers that we will convert to ELL
layersToConvert = cntk_layers.get_filtered_layers_list(modelLayers)
_logger.info("\nFinished pre-processing.")
predictor = None
try:
# Create a list of ELL layers from the CNTK layers
ellLayers = cntk_layers.convert_cntk_layers_to_ell_layers(
layersToConvert)
# Create an ELL neural network predictor from the layers
predictor = ell.neural.NeuralNetworkPredictor(ellLayers)
except BaseException as exception:
_logger.error("Error occurred attempting to convert cntk layers to ELL layers: " + str(exception))
raise exception
return predictor
def __init__(self, layer):
if not layer.is_block:
raise ValueError(
"Error: Convolution layer node is not in block node")
self.op_name = 'Convolution'
# initialize weights and input characteristics
self.input_parameter = layer.arguments[0]
self.weights_parameter = utilities.find_parameter_by_name(
layer.parameters, 'W', 0)
self.bias_parameter = utilities.find_parameter_by_name(
layer.parameters, 'b', 1)
# Get the hyper-parameters for the convolution.
# They are on the convolution node inside this block.
convolution_nodes = depth_first_search(
layer.block_root,
lambda x: utilities.op_name_equals(x, 'Convolution'))
self.attributes = convolution_nodes[0].attributes
self.convolution_method = 0
self.input_shape = self.input_parameter.shape
super().__init__(layer)
nodes = utilities.get_model_layers(layer.block_root)
if utilities.is_softmax_activation(nodes):
self.additional_layer_text = 'softmax'
else:
activation_type = utilities.get_cntk_activation_name(nodes)
if activation_type:
self.additional_layer_text = activation_type
def process(self, ellLayers):
"""Appends the ELL equivalent of the current layer to ellLayers."""
# Note that a single CNTK Dense function block is equivalent to the following 3 ELL layers:
# - FullyConnectedLayer
# - BiasLayer
# - ActivationLayer. This layer is sometimes missing, depending on activation type.
#
# Therefore, make sure the output padding characteristics of the last layer reflect the next layer's
# padding requirements.
weightsParameter = utilities.find_parameter_by_name(
self.layer.parameters, 'W', 0)
biasParameter = utilities.find_parameter_by_name(
self.layer.parameters, 'b', 1)
weightsTensor = converters.get_tensor_from_cntk_dense_weight_parameter(
weightsParameter)
biasVector = converters.get_vector_from_cntk_trainable_parameter(
biasParameter)
# Create the ell.neural.LayerParameters for the various ELL layers
firstLayerParameters = ell.neural.LayerParameters(
self.layer.ell_inputShape, self.layer.ell_inputPaddingParameters, self.layer.ell_outputShapeMinusPadding,
ell.neural.NoPadding(), ell.nodes.PortType.smallReal)
middleLayerParameters = ell.neural.LayerParameters(
self.layer.ell_outputShapeMinusPadding, ell.neural.NoPadding(),
self.layer.ell_outputShapeMinusPadding, ell.neural.NoPadding(), ell.nodes.PortType.smallReal)
lastLayerParameters = ell.neural.LayerParameters(
self.layer.ell_outputShapeMinusPadding, ell.neural.NoPadding(),
self.layer.ell_outputShape, self.layer.ell_outputPaddingParameters, ell.nodes.PortType.smallReal)
layerParameters = firstLayerParameters
internalNodes = utilities.get_model_layers(self.layer.block_root)
activationType = utilities.get_ell_activation_type(internalNodes)
# Create the ELL fully connected layer
ellLayers.append(ell.neural.FullyConnectedLayer(
layerParameters, weightsTensor))
# Create the ELL bias layer
if (utilities.is_softmax_activation(internalNodes) or activationType is not None):
layerParameters = middleLayerParameters
else:
layerParameters = lastLayerParameters
ellLayers.append(ell.neural.BiasLayer(layerParameters, biasVector))
# Create the ELL activation layer
if (utilities.is_softmax_activation(internalNodes) or activationType is not None):
layerParameters = lastLayerParameters
# Special case: if this is softmax activation, create an ELL Softmax layer.
# Else, insert an ELL ActivationLayer
if (utilities.is_softmax_activation(internalNodes)):
ellLayers.append(ell.neural.SoftmaxLayer(layerParameters))
else:
if (activationType is not None):
ellLayers.append(ell.neural.ActivationLayer(
layerParameters, activationType))
def __init__(self, layer):
if not layer.is_block:
raise ValueError("Dense node is not a block node")
self.op_name = 'Dense'
super().__init__(layer)
internalNodes = utilities.get_model_layers(self.layer.block_root)
self.additional_layer_text = utilities.get_cntk_activation_name(internalNodes)
def clone_cntk_layer(self, feature):
"""Returns a clone of the CNTK layer for per-layer forward prop validation"""
weightsParameter = utilities.find_parameter_by_name(
self.layer.parameters, 'W', 0)
biasParameter = utilities.find_parameter_by_name(
self.layer.parameters, 'b', 1)
internalNodes = utilities.get_model_layers(self.layer.block_root)
activationType = utilities.get_cntk_activation_op(internalNodes)
includeBias = biasParameter is not None
layer = Dense(self.layer.shape, activation=activationType, bias=includeBias)(feature)
layer.parameters[0].value = weightsParameter.value
if includeBias:
layer.parameters[1].value = biasParameter.value
return layer
def clone_cntk_layer(self, feature):
"""Returns a clone of the CNTK layer for per-layer forward prop validation"""
nodes = utilities.get_model_layers(self.layer.block_root)
activation = utilities.get_cntk_activation_op(nodes)
weightsShape = self.weights_parameter.shape
pad = self.attributes['autoPadding'][0] or (
self.attributes['autoPadding'][1] and self.attributes['autoPadding'][2])
bias = (self.bias_parameter is not None)
layer = Convolution((weightsShape[2], weightsShape[3]), weightsShape[0],
pad=pad, activation=activation, bias=bias)(feature)
layer.parameters[0].value = self.weights_parameter.value
if bias:
layer.parameters[1].value = self.bias_parameter.value
return layer
def run(self):
self.report = open("report.md", "w")
self.report.write("# Comparison Results\n")
self.report.write("**model**: %s\n\n" % (self.model_file))
if self.image_file is not None:
self.image = self.load_image(self.image_file)
self.report.write("**image**: %s\n\n" % (self.image_file))
self.cntk_model = cntk.load_model(self.model_file)
modelLayers = cntk_utilities.get_model_layers(self.cntk_model)
# Get the relevant CNTK layers that we will convert to ELL
layersToConvert = cntk_layers.get_filtered_layers_list(modelLayers)
self.logger.info(
"----------------------------------------------------------------------------------"
)
if self.layers:
for layer in layersToConvert:
self.compare_layer(layer)
else:
self.compare_model(layersToConvert)
self.print_top_result()
self.report.close()
def process(self, ellLayers):
"""Helper to convert a convolutional layer to the ELL equivalent."""
# Note that a single CNTK Convolutional function block is equivalent to the following 3 ELL layers:
# - ConvolutionalLayer
# - BiasLayer. This layer is sometimes missing, depending on whether bias is included.
# - ActivationLayer. This layer is sometimes missing, depending on activation type.
#
# Therefore, make sure the output padding characteristics of the last layer reflect the next layer's
# padding requirements.
weightsTensor = converters.get_tensor_from_cntk_convolutional_weight_parameter(
self.weights_parameter)
internalNodes = utilities.get_model_layers(self.layer.block_root)
activationType = utilities.get_ell_activation_type(internalNodes)
isSoftmaxActivation = utilities.is_softmax_activation(internalNodes)
hasActivation = isSoftmaxActivation or activationType != None
hasBias = self.bias_parameter != None
# Create the ell.neural.LayerParameters for the various ELL layers
onlyLayerParameters = ell.neural.LayerParameters(
self.layer.ell_inputShape, self.layer.ell_inputPaddingParameters,
self.layer.ell_outputShape, self.layer.ell_outputPaddingParameters,
ell.nodes.PortType.smallReal)
firstLayerParameters = ell.neural.LayerParameters(
self.layer.ell_inputShape, self.layer.ell_inputPaddingParameters,
self.layer.ell_outputShapeMinusPadding, ell.neural.NoPadding(),
ell.nodes.PortType.smallReal)
middleLayerParameters = ell.neural.LayerParameters(
self.layer.ell_outputShapeMinusPadding,
ell.neural.NoPadding(), self.layer.ell_outputShapeMinusPadding,
ell.neural.NoPadding(), ell.nodes.PortType.smallReal)
lastLayerParameters = ell.neural.LayerParameters(
self.layer.ell_outputShapeMinusPadding, ell.neural.NoPadding(),
self.layer.ell_outputShape, self.layer.ell_outputPaddingParameters,
ell.nodes.PortType.smallReal)
# Choose the layer parameters for the convolutional layer. If there is
# bias or activation, then the convolution is the first of two or more,
# otherwise it is the only layer
if hasActivation or hasBias:
layerParameters = firstLayerParameters
else:
layerParameters = onlyLayerParameters
# Fill in the convolutional parameters
weightsShape = self.weights_parameter.shape
receptiveField = weightsShape[2]
stride = self.attributes['strides'][2]
filterBatchSize = layerParameters.outputShape.channels
convolutionalParameters = ell.neural.ConvolutionalParameters(
receptiveField, stride, self.convolution_method, filterBatchSize)
# Create the ELL convolutional layer
ellLayers.append(
ell.neural.ConvolutionalLayer(layerParameters,
convolutionalParameters,
weightsTensor))
# Create the ELL bias layer
if hasBias:
if hasActivation:
layerParameters = middleLayerParameters
else:
layerParameters = lastLayerParameters
biasVector = converters.get_vector_from_cntk_trainable_parameter(
self.bias_parameter)
ellLayers.append(ell.neural.BiasLayer(layerParameters, biasVector))
# Create the ELL activation layer
if hasActivation:
layerParameters = lastLayerParameters
# Special case: if this is softmax activation, create an ELL Softmax layer.
# Else, insert an ELL ActivationLayer
if (isSoftmaxActivation):
ellLayers.append(ell.neural.SoftmaxLayer(layerParameters))
else:
ellLayers.append(
ell.neural.ActivationLayer(layerParameters,
activationType))
