def test_convolution_layer(self):
"""Test a model with a single CNTK Convolution layer against the
equivalent ELL predictor. This verifies that the import functions
reshape and reorder values appropriately and that the equivalent ELL
layer produces comparable output
"""
# Create a Convolution CNTK layer with no bias or activation,
# auto-padding, stride of 1
convolutionLayer = Convolution((3, 3), 5, pad=(
True, True), strides=1, bias=False, init=0)
x = input((2, 3, 4)) # Input order for CNTK is channels, rows, columns
cntkModel = convolutionLayer(x)
# Create a test set of weights to use for both CNTK and ELL layers
# CNTK has these in filters, channels, rows, columns order
weightValues = np.arange(90, dtype=np.float_).reshape(5, 2, 3, 3)
# Set the weights
convolutionLayer.parameters[0].value = weightValues
# create an ELL Tensor from the cntk weights, which re-orders the
# weights and produces an appropriately dimensioned tensor
weightTensor = cntk_converters.\
get_float_tensor_from_cntk_convolutional_weight_parameter(
convolutionLayer.parameters[0])
# Create the equivalent ELL predictor
layerParameters = ell.LayerParameters(
# Input order for ELL is rows, columns, channels. Account for
# padding.
ell.TensorShape(3 + 2, 4 + 2, 2),
ell.ZeroPadding(1),
ell.TensorShape(3, 4, 5),
ell.NoPadding())
convolutionalParameters = ell.ConvolutionalParameters(3, 1, 0, 5)
layer = ell.FloatConvolutionalLayer(
layerParameters, convolutionalParameters, weightTensor)
predictor = ell.FloatNeuralNetworkPredictor([layer])
# Get the results for both
inputValues = np.arange(24, dtype=np.float32).reshape(2, 3, 4)
cntkResults = cntkModel(inputValues)
orderedCntkResults = cntk_converters.get_float_vector_from_cntk_array(
cntkResults)
orderedInputValues = cntk_converters.get_float_vector_from_cntk_array(
inputValues)
ellResults = predictor.Predict(orderedInputValues)
# Compare the results
np.testing.assert_array_equal(
orderedCntkResults, ellResults,
'results for Convolution layer do not match!')
# now run same over ELL compiled model
self.verify_compiled(
predictor, orderedInputValues, orderedCntkResults, "convolution",
"test")
def process_network(network, weightsData, convolutionOrder):
"""Returns an ell.FloatNeuralNetworkPredictor as a result of parsing the network layers"""
ellLayers = []
for layer in network:
if layer['type'] == 'net':
pass
elif layer['type'] == 'convolutional':
ellLayers += process_convolutional_layer(layer, weightsData,
convolutionOrder)
elif layer['type'] == 'connected':
ellLayers += process_fully_connected_layer(layer, weightsData)
elif layer['type'] == 'maxpool':
ellLayers.append(get_pooling_layer(layer, ell.PoolingType.max))
elif layer['type'] == 'avgpool':
ellLayers.append(get_pooling_layer(layer, ell.PoolingType.mean))
elif layer['type'] == 'softmax':
ellLayers.append(get_softmax_layer(layer))
else:
print("Skipping, ", layer['type'], "layer")
print()
if ellLayers:
# Darknet expects the input to be between 0 and 1, so prepend
# a scaling layer with a scale factor of 1/255
parameters = ellLayers[0].parameters
ellLayers = [get_first_scaling_layer(parameters)] + ellLayers
predictor = ell.FloatNeuralNetworkPredictor(ellLayers)
return predictor
def predictor_from_cntk_model(modelFile, plotModel=False):
"""Loads a CNTK model and returns an ell.NeuralNetworkPredictor"""
print("Loading...")
z = load_model(modelFile)
print("\nFinished loading.")
if plotModel:
filename = os.path.join(os.path.dirname(modelFile),
os.path.basename(modelFile) + ".png")
cntk_utilities.plot_model(z, filename)
print("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)
print("\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.FloatNeuralNetworkPredictor(ellLayers)
except BaseException as exception:
print("Error occurred attempting to convert cntk layers to ELL layers")
raise exception
return predictor
def get_predictor(self, layer):
ell_layers = []
# remove output_padding from because CNTK doesn't have output padding.
layer.layer.ell_outputPaddingParameters = ell.PaddingParameters(
ell.PaddingScheme.zeros, 0)
layer.layer.ell_outputShape = cntk_utilities.get_adjusted_shape(
layer.layer.output.shape, layer.layer.ell_outputPaddingParameters)
layer.process(ell_layers)
# Create an ELL neural network predictor from the relevant CNTK layers
return ell.FloatNeuralNetworkPredictor(ell_layers)
def test_prelu_activation_layer(self):
"""Test a model with a single CNTK PReLU activation layer against the
equivalent ELL predictor. This verifies that the import functions
reshape and reorder values appropriately and that the equivalent ELL
layer produces comparable output
"""
# Create a test set of alpha parameters to use for both CNTK and ELL
# layers
# Input order for CNTK is channels, rows, columns
alphaValues = np.linspace(
1, 2, num=16 * 10 * 10, dtype=np.float32).reshape(16, 10, 10)
# create an ELL Tensor from the alpha parameters, which re-orders and
# produces an appropriately dimensioned tensor
alphaTensor = cntk_converters.\
get_float_tensor_from_cntk_convolutional_weight_value_shape(
alphaValues, alphaValues.shape)
inputValues = np.linspace(
-5, 5, num=16 * 10 * 10, dtype=np.float32).reshape(16, 10, 10)
# Evaluate a PReLU CNTK layer
x = input((16, 10, 10))
p = parameter(shape=x.shape, init=alphaValues, name="prelu")
cntkModel = param_relu(p, x)
# Create the equivalent ELL predictor
layerParameters = ell.LayerParameters(
# Input order for ELL is rows, columns, channels
ell.TensorShape(10, 10, 16),
ell.NoPadding(),
ell.TensorShape(10, 10, 16),
ell.NoPadding())
layer = ell.FloatPReLUActivationLayer(layerParameters, alphaTensor)
predictor = ell.FloatNeuralNetworkPredictor([layer])
cntkResults = cntkModel(inputValues)
orderedCntkResults = cntk_converters.get_float_vector_from_cntk_array(
cntkResults)
orderedInputValues = cntk_converters.get_float_vector_from_cntk_array(
inputValues)
ellResults = predictor.Predict(orderedInputValues)
# Compare the results
np.testing.assert_array_equal(
orderedCntkResults, ellResults,
'results for PReLU Activation layer do not match!')
# now run same over ELL compiled model
self.verify_compiled(
predictor, orderedInputValues, orderedCntkResults,
"prelu_activation", "test")
def test_dense_layer(self):
"""Test a model with a single CNTK Dense layer against the equivalent
ELL predictor. This verifies that the import functions reshape and
reorder values appropriately and that the equivalent ELL layer
produces comparable output
"""
# Create a Dense CNTK layer with no bias or activation
denseLayer = Dense(5, bias=False)
x = input((2, 3, 4)) # Input order for CNTK is channels, rows, columns
cntkModel = denseLayer(x)
# Create a test set of weights to use for both CNTK and ELL layers
# CNTK has these in channels, rows, columns, [output shape] order
weightValues = np.arange(120, dtype=np.float_).reshape(2, 3, 4, 5)
# Set the weights
denseLayer.parameters[0].value = weightValues
# create an ELL Tensor from the cntk weights, which re-orders the
# weights and produces an appropriately dimensioned tensor
weightTensor = cntk_converters.\
get_float_tensor_from_cntk_dense_weight_parameter(
denseLayer.parameters[0])
# Create the equivalent ELL predictor
layerParameters = ell.LayerParameters(
# Input order for ELL is rows, columns, channels
ell.TensorShape(3, 4, 2),
ell.NoPadding(),
ell.TensorShape(1, 1, 5),
ell.NoPadding())
layer = ell.FloatFullyConnectedLayer(layerParameters, weightTensor)
predictor = ell.FloatNeuralNetworkPredictor([layer])
# Get the results for both
inputValues = np.arange(24, dtype=np.float32).reshape(2, 3, 4)
orderedInputValues = cntk_converters.get_float_vector_from_cntk_array(
inputValues)
cntkResults = cntkModel(inputValues)
orderedCntkResults = cntk_converters.get_float_vector_from_cntk_array(
cntkResults)
ellResults = predictor.Predict(orderedInputValues)
# Compare the results
np.testing.assert_array_equal(
orderedCntkResults, ellResults,
'results for Dense layer do not match!')
# now run same over ELL compiled model
self.verify_compiled(predictor, orderedInputValues, orderedCntkResults,
"dense", "test")
def compare_model(self, layers):
ellLayers = cntk_layers.convert_cntk_layers_to_ell_layers(layers)
# Create an ELL neural network predictor from the layers
predictor = ell.FloatNeuralNetworkPredictor(ellLayers)
shape = predictor.GetInputShape()
self.input_shape = (shape.channels, shape.rows, shape.columns
) # to CNTK (channel, rows, coumns) order
self.data = self.get_input_data()
if (len(self.cntk_model.arguments) > 1):
output = np.zeros(self.cntk_model.arguments[1].shape).astype(
np.float32)
predictions = self.cntk_model.eval({
self.cntk_model.arguments[0]: [self.data],
self.cntk_model.arguments[1]:
output
})
else:
predictions = self.cntk_model.eval(
{self.cntk_model.arguments[0]: [self.data]})
size = 0
output = None
if isinstance(predictions, dict):
for key in self.cntk_model.outputs:
shape = key.shape
if len(shape) > 0:
s = np.max(shape)
if (s > size):
size = s
output = predictions[key][0] / 100
else:
output = predictions[0]
self.verify_ell("Softmax", predictor, self.data, output)
self.data = output # make this the input to the next layer.
self.save_layer_outputs("Softmax")
def compare_model(self, layers):
ellLayers = cntk_layers.convert_cntk_layers_to_ell_layers(layers)
# Create an ELL neural network predictor from the layers
predictor = ell.FloatNeuralNetworkPredictor(ellLayers)
shape = predictor.GetInputShape()
# to CNTK (channel, rows, columns) order
self.input_shape = (shape.channels, shape.rows, shape.columns)
self.data = self.get_input_data()
if len(self.cntk_model.arguments) > 1:
output = np.zeros(self.cntk_model.arguments[1].shape,
dtype=np.float32)
predictions = self.cntk_model.eval({
self.cntk_model.arguments[0]: [self.data],
self.cntk_model.arguments[1]: [list(range(len(self.categories)))] })
else:
predictions = self.cntk_model.eval({
self.cntk_model.arguments[0]: [self.data]})
size = 0
cntk_output = None
if isinstance(predictions, dict):
for key in self.cntk_model.outputs:
shape = key.shape
if shape:
s = np.max(shape)
if s > size:
size = s
# CNTK models currently don't have softmax operations
# right now, so we work around it by including it
# explicitly
cntk_output = softmax(predictions[key][0]).eval()
else:
cntk_output = softmax(predictions).eval()
self.verify_ell("Softmax", predictor, self.data, cntk_output)
def compare_predictor_output(modelFile, labels, modelTestInput=None,
maxLayers=None):
"""Compares an ell.NeuralNetworkPredictor against its equivalent CNTK
model.
Parameters:
modelFile -- path to the CNTK model file
labels -- array of labels
modelTestInput -- input data in row, column, channel ordering
maxLayers -- integer to indicate how many layers to run before stopping.
Setting to None will run all layers and compare against the
original model.
"""
z = load_model(modelFile)
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, maxLayers)
if not layersToConvert:
raise RuntimeError("No layers are converted, nothing to test")
# Create a list of ELL layers from the relevant CNTK layers
print("\nCreating ELL predictor...")
ellLayers = cntk_layers.convert_cntk_layers_to_ell_layers(
layersToConvert)
# Create an ELL neural network predictor from the relevant CNTK layers
predictor = ell.FloatNeuralNetworkPredictor(ellLayers)
if not modelTestInput:
inputShape = predictor.GetInputShape()
modelTestInput = np.random.uniform(
low=0, high=255, size=(
inputShape.rows, inputShape.columns, inputShape.channels)
).astype(np.float_)
ellTestInput = modelTestInput.ravel() # rows, columns, channels
ellResults = predictor.Predict(ellTestInput)
# rows, columns, channels => channels, rows, columns
cntkTestInput = np.moveaxis(modelTestInput, -1, 0).astype(np.float32)
cntkTestInput = np.ascontiguousarray(cntkTestInput)
# Get the equivalent CNTK model
if not maxLayers:
print("\nRunning original CNTK model...")
_, out = z.forward(
{z.arguments[0]: [cntkTestInput],
z.arguments[1]: [list(range(len(labels)))]})
for output in z.outputs:
if (output.shape == (len(labels),)):
out = out[output]
cntkResults = softmax(out[0]).eval()
# For the full model, we compare prediction output instead of layers
np.testing.assert_array_almost_equal(
cntkResults, ellResults, 5,
'prediction outputs do not match! (for model ' + modelFile + ')')
else:
print("\nRunning partial CNTK model...")
if (layersToConvert[-1].layer.op_name == 'CrossEntropyWithSoftmax' and
len(layersToConvert) > 2):
# ugly hack for CrossEntropyWithSoftmax
zz = as_composite(layersToConvert[-2].layer)
zz = softmax(zz)
else:
zz = as_composite(layersToConvert[-1].layer)
zz = softmax(zz)
out = zz(cntkTestInput)
orderedCntkModelResults = cntk_converters.\
get_float_vector_from_cntk_array(out)
np.testing.assert_array_almost_equal(
orderedCntkModelResults, ellResults, 5,
('prediction outputs do not match! (for partial model ' +
modelFile + ')'))
def test_batch_normalization_layer(self):
"""Test a model with a single CNTK BatchNormalization layer against the
equivalent ELL predictor This verifies that the import functions
reshape and reorder values appropriately and that the equivalent ELL
layer produces comparable output
"""
# Create a test set of scales and biases to use for both CNTK and ELL
# layers
scaleValues = np.linspace(0.1, 0.5, num=16, dtype=np.float32)
scaleVector = cntk_converters.get_float_vector_from_cntk_array(
scaleValues)
biasValues = np.linspace(1, 2, num=16, dtype=np.float32)
biasVector = cntk_converters.get_float_vector_from_cntk_array(
biasValues)
meanValues = np.linspace(-0.5, 0.5, num=16, dtype=np.float32)
meanVector = cntk_converters.get_float_vector_from_cntk_array(
meanValues)
varianceValues = np.linspace(-1, 1, num=16, dtype=np.float32)
varianceVector = cntk_converters.get_float_vector_from_cntk_array(
varianceValues)
# Create a BatchNormalization CNTK layer
# CNTK's BatchNormalization layer does not support setting the running
# mean and variance, so we use a wrapper function around the
# batch_normalization op
batchNorm = BatchNormalizationTester(
init_scale=scaleValues, norm_shape=scaleValues.shape,
init_bias=biasValues, init_mean=meanValues,
init_variance=varianceValues)
# Input order for CNTK is channels, rows, columns
x = input((16, 10, 10))
cntkModel = batchNorm(x)
# Create the equivalent ELL predictor
layers = []
layerParameters = ell.LayerParameters(
# Input order for ELL is rows, columns, channels
ell.TensorShape(10, 10, 16),
ell.NoPadding(),
ell.TensorShape(10, 10, 16),
ell.NoPadding())
# CNTK BatchNorm = ELL's BatchNorm + Scaling + Bias
# 1e-5 is the default epsilon for CNTK's BatchNormalization Layer
epsilon = 1e-5
layers.append(ell.FloatBatchNormalizationLayer(
layerParameters, meanVector, varianceVector, epsilon,
ell.EpsilonSummand_variance))
layers.append(ell.FloatScalingLayer(layerParameters, scaleVector))
layers.append(ell.FloatBiasLayer(layerParameters, biasVector))
predictor = ell.FloatNeuralNetworkPredictor(layers)
inputValues = np.linspace(
-5, 5, num=16 * 10 * 10, dtype=np.float32).reshape(16, 10, 10)
cntkResults = cntkModel(inputValues)
orderedCntkResults = cntk_converters.get_float_vector_from_cntk_array(
# Note that cntk inserts an extra dimension of 1 in the front
cntkResults)
orderedInputValues = cntk_converters.get_float_vector_from_cntk_array(
inputValues)
ellResults = predictor.Predict(orderedInputValues)
# Compare the results (precision is 1 less decimal place from epsilon)
np.testing.assert_array_almost_equal(
orderedCntkResults, ellResults, 6,
'results for BatchNormalization layer do not match!')
# now run same over ELL compiled model
self.verify_compiled(
predictor, orderedInputValues, orderedCntkResults, "batch_norm",
"test", precision=6)
def test_binary_convolution_layer(self):
"""Test a model with a single CNTK Binary Convolution layer against the
equivalent ELL predictor. This verifies that the import functions
reshape and reorder values appropriately and that the equivalent ELL
layer produces comparable output
"""
# Create a test set of weights to use for both CNTK and ELL layers
# CNTK has these in filters, channels, rows, columns order
weightValues = np.random.uniform(
low=-5, high=5, size=(5, 2, 3, 3)).astype(np.float32)
# create an ELL Tensor from the cntk weights, which re-orders the
# weights and produces an appropriately dimensioned tensor
weightTensor = cntk_converters.\
get_float_tensor_from_cntk_convolutional_weight_value_shape(
weightValues, weightValues.shape)
# Create a Binary Convolution CNTK layer with no bias, no activation,
# stride 1
# Input order for CNTK is channels, rows, columns
x = input((2, 10, 10))
cntkModel = CustomSign(x)
cntkModel = BinaryConvolution(
(10, 10), num_filters=5, channels=2, init=weightValues,
pad=True, bias=False, init_bias=0, activation=False)(cntkModel)
# Create the equivalent ELL predictor
layerParameters = ell.LayerParameters(
# Input order for ELL is rows, columns, channels. Account for
# padding.
ell.TensorShape(10 + 2, 10 + 2, 2),
ell.ZeroPadding(1),
ell.TensorShape(10, 10, 5),
ell.NoPadding())
convolutionalParameters = ell.BinaryConvolutionalParameters(
3, 1, ell.BinaryConvolutionMethod.bitwise,
ell.BinaryWeightsScale.none)
layer = ell.FloatBinaryConvolutionalLayer(
layerParameters, convolutionalParameters, weightTensor)
predictor = ell.FloatNeuralNetworkPredictor([layer])
# Get the results for both
inputValues = np.random.uniform(
low=-50, high=50, size=(2, 10, 10)).astype(np.float32)
cntkResults = cntkModel(inputValues)
orderedCntkResults = cntk_converters.get_float_vector_from_cntk_array(
cntkResults)
orderedInputValues = cntk_converters.get_float_vector_from_cntk_array(
inputValues)
ellResults = predictor.Predict(orderedInputValues)
# Compare the results
np.testing.assert_array_equal(
orderedCntkResults, ellResults,
'results for Binary Convolution layer do not match!')
# now run same over ELL compiled model
self.verify_compiled(
predictor, orderedInputValues, orderedCntkResults,
"binary_convolution", "test")
def test_max_pooling_layer(self):
"""Test a model with a single CNTK MaxPooling layer against the
equivalent ELL predictor. This verifies that the import functions
reshape and reorder values appropriately and that the equivalent ELL
layer produces comparable output
"""
x = input((2, 15, 15))
count = 0
inputValues = np.random.uniform(
low=-5, high=5, size=(2, 15, 15)).astype(np.float32)
for pool_size, stride_size in product(range(2, 4), range(2, 3)):
count += 1
print("test pooling size ({0},{0}) and stride {1}".format(
pool_size, stride_size))
# Create a MaxPooling CNTK layer
poolingLayer = MaxPooling(
(pool_size, pool_size), pad=True, strides=stride_size)
# Input order for CNTK is channels, rows, columns
cntkModel = poolingLayer(x)
# Get the results for both
cntkResults = cntkModel(inputValues)[0]
outputShape = cntkResults.shape
padding = int((pool_size - 1) / 2)
rows = int(inputValues.shape[1] + 2*padding)
columns = int(inputValues.shape[2] + 2*padding)
channels = int(inputValues.shape[0])
# Create the equivalent ELL predictor
layerParameters = ell.LayerParameters(
# Input order for ELL is rows, columns, channels
ell.TensorShape(rows, columns, channels),
ell.MinPadding(padding),
ell.TensorShape(
outputShape[1], outputShape[2], outputShape[0]),
ell.NoPadding())
poolingParameters = ell.PoolingParameters(
pool_size, stride_size)
layer = ell.FloatPoolingLayer(
layerParameters, poolingParameters, ell.PoolingType.max)
predictor = ell.FloatNeuralNetworkPredictor([layer])
# Note that cntk inserts an extra dimension of 1 in the front
orderedCntkResults = cntk_converters.\
get_float_vector_from_cntk_array(cntkResults)
orderedInputValues = cntk_converters.\
get_float_vector_from_cntk_array(inputValues)
ellResults = predictor.Predict(orderedInputValues)
# Compare them
np.testing.assert_array_almost_equal(
orderedCntkResults, ellResults, 5,
('results for MaxPooling layer do not match! (poolsize = '
'{}, stride = {}').format(pool_size, stride_size))
# now run same over ELL compiled model
self.verify_compiled(
predictor, orderedInputValues, orderedCntkResults,
'max_pooling{}_{}'.format(pool_size, stride_size),
'test_' + str(count))