def process_batch_normalization_layer(layer, apply_padding, mean_vals,
variance_vals, scale_vals):
"""Returns ELL layers corresponding to a Darknet batch normalization layer"""
# Batch normalization in Darknet corresponds to BatchNormalizationLayer, ScalingLayer in ELL
layers = []
# Create BatchNormalizationLayer
layerParameters = create_layer_parameters(layer['outputShapeMinusPadding'],
0, ell.PaddingScheme.zeros,
layer['outputShapeMinusPadding'],
0, ell.PaddingScheme.zeros)
meanVector = ell.FloatVector(mean_vals.ravel())
varianceVector = ell.FloatVector(variance_vals.ravel())
layers.append(
ell.FloatBatchNormalizationLayer(layerParameters, meanVector,
varianceVector, 1e-6,
ell.EpsilonSummand_sqrtVariance))
# Create Scaling Layer
if (apply_padding):
layerParameters = create_layer_parameters(
layer['outputShapeMinusPadding'], 0, ell.PaddingScheme.zeros,
layer['outputShape'], layer['outputPadding'],
layer['outputPaddingScheme'])
else:
layerParameters = create_layer_parameters(
layer['outputShapeMinusPadding'], 0, ell.PaddingScheme.zeros,
layer['outputShapeMinusPadding'], 0, ell.PaddingScheme.zeros)
layers.append(ell.FloatScalingLayer(layerParameters, scale_vals.ravel()))
return layers
def get_float_vector_from_cntk_trainable_parameter(tensorParameter):
"""Returns an ell.FloatVector from a trainable parameter
Note that ELL's ordering is row, column, channel.
CNTK has them in filter, channel, row, column order.
"""
tensorShape = tensorParameter.shape
tensorValue = tensorParameter.value
orderedWeights = np.zeros(tensorValue.size, dtype=np.float)
i = 0
for columnValue in tensorValue:
orderedWeights[i] = columnValue
i += 1
return ell.FloatVector(orderedWeights)
def get_bias_layer(layer, apply_padding, bias_vals):
"""Return an ELL bias layer from a darknet layer"""
if (apply_padding):
layerParameters = create_layer_parameters(
layer['outputShapeMinusPadding'], 0, ell.PaddingScheme.zeros,
layer['outputShape'], layer['outputPadding'],
layer['outputPaddingScheme'])
else:
layerParameters = create_layer_parameters(
layer['outputShapeMinusPadding'], 0, ell.PaddingScheme.zeros,
layer['outputShapeMinusPadding'], 0, ell.PaddingScheme.zeros)
biasVector = ell.FloatVector(bias_vals.ravel())
return ell.FloatBiasLayer(layerParameters, biasVector)
def test_float():
testing = Testing()
# empty vector
e = ell.FloatVector()
np.testing.assert_equal(e.size(), 0)
# vector from list of floats
l = [1.1,2.2,3.3,4.4]
e = ell.FloatVector(l)
assert_compare_floats(e, l)
# vector from numpy array
a = np.array(range(10), dtype=np.float32)
e = ell.FloatVector(a)
np.testing.assert_equal(np.asarray(e), a)
# convert to numpy using array
b = np.array(e).ravel()
np.testing.assert_equal(a, b)
# copy_from numpy array
e = ell.FloatVector()
e.copy_from(a)
np.testing.assert_equal(np.asarray(e), a)
# convert data types
a = a.astype(np.float)
e = ell.FloatVector(a)
np.testing.assert_equal(np.asarray(e), a)
# enumerating array
for i in range(a.shape[0]):
x = a[i]
y = e[i]
np.testing.assert_equal(x, y)
# auto-ravel numpy arrays
a = np.ones((10,10), dtype=np.float32)
a *= range(10)
e = ell.FloatVector(a)
np.testing.assert_equal(np.asarray(e), a.ravel())
testing.ProcessTest("FloatVector test", True)
def get_float_vector_from_constant(constant, size):
# Workaround: For some reason, np.full is not returning a type that SWIG can parse. So just manually walk the array setting the scalar
array = np.zeros(size, dtype=np.float)
for i in range(array.size):
array[i] = constant
return ell.FloatVector(array)