def test_synMem2(self):
"""Test issue where nxcompiler runs out of synMem.
The problem is that whenever the number of syn bits is an exact
multiple of 64, nxcompiler requires an extra mem word. E.g. 128 bits
would count as 3 words.
We have implemented a check in the NxTF compiler that allocates an
extra mem word when it detects that the numBits are an exact multiple
of 64. So this test should pass.
"""
inputLayer = NxInputLayer((1, 1, 1))
x = NxConv2D(6, (1, 1),
name='conv1',
kernel_initializer='ones',
synapseEncoding='dense1',
useSharedSign=False)(inputLayer.input)
model = NxModel(inputLayer.input, x, saveOutput=False)
model.compileModel()
model.run(1)
model.disconnect()
layers = model.partitionOptimizer.getLayers()
assert layers[1].numSynMemWords == 2
def test_synMem(self):
"""Test issue where nxcompiler runs out of synMem.
The problem is that whenever the number of syn bits is an exact
multiple of 64, nxcompiler requires an extra mem word. E.g. 128 bits
would count as 3 words.
Normally, this test would fail due to an assertion in the nxcompiler
that checks for number of syn mem words to be smaller than 2**14.
We have implemented a check in the NxTF compiler that allocates an
extra mem word when it detects that the numBits are an exact multiple
of 64. So this test should pass.
"""
np.random.seed(123)
inputLayer = NxInputLayer((1, 1, 256))
x = NxConv2D(256, (1, 1), name='conv1')(inputLayer.input)
model = NxModel(inputLayer.input, x)
model.compileModel()
model.run(1)
model.disconnect()
def test_NxInputLayer(self):
"""
Test input layer in soft-reset mode.
"""
plot = False
verbose = False
resetMode = 'soft'
neuronSize = 2 if resetMode == 'soft' else 1
input_shape = (32, 32, 3)
inputSize = np.prod(input_shape)
inputLayer = NxInputLayer(input_shape,
probeSpikes=True,
vThMant=255,
resetMode=resetMode)
out = inputLayer.input
model = NxModel(out, out)
model.compile('adam', 'categorical_crossentropy', ['accuracy'])
model.summary()
layers = [inputLayer]
input_image = np.linspace(
0, 256, endpoint=False,
num=inputSize).reshape(input_shape).astype(int)
x_test = [input_image]
y_test = [0]
mapper = model.compileModel()
if verbose:
printLayerMappings(layers, mapper, synapses=True, inputAxons=True)
printLayers(layers)
layerProbes = []
numProbes = 32
for i, layer in enumerate(layers):
shape = layer.output_shape[1:]
# Define probes to read out currents.
vProbes = []
sProbes = []
uProbes = []
toProbe = numProbes if i == (len(layers) -
1) else numProbes * neuronSize
toProbe = min(toProbe,
int(np.asscalar(np.prod(shape))) * neuronSize)
for j in range(toProbe):
vProbes.append(layer[j].probe(ProbableStates.VOLTAGE))
sProbes.append(layer[j].probe(ProbableStates.ACTIVITY))
uProbes.append(layer[j].probe(ProbableStates.CURRENT))
layerProbes.append([uProbes, vProbes, sProbes])
# How many time steps to run each sample.
num_steps = 1000
# How many samples to test.
num_samples_to_test = 1
# Iterate over samples in testset.
for i, (input_image, target) in enumerate(zip(x_test, y_test)):
if i == num_samples_to_test:
break
if plot:
plt.hist(input_image.ravel())
plt.show()
# Set input bias currents.
for j, b in enumerate(np.ravel(input_image, 'F')):
inputLayer[j * neuronSize].biasMant = b
inputLayer[j * neuronSize].biasExp = 6
inputLayer[j * neuronSize].phase = 2
# Run model.
model.run(num_steps)
# Clean up.
data = [[extract(probe) for probe in lp] for lp in layerProbes]
spikesRates = []
for i, (layer, d) in enumerate(zip(layers, data)):
sData = d[2][:, (neuronSize - 1)::neuronSize]
spikecount = (sData // 127).sum(0)
spikesRates.append(spikecount / num_steps)
model.disconnect()
layer_activations = [x_test[0]]
if plot:
for activations, spikerate in zip(layer_activations, spikesRates):
plt.figure()
scale = np.max(activations)
spikesFlat = spikerate #spikerate.flatten()
plt.plot(
activations.flatten('F')[:len(spikesFlat)] / scale,
spikesFlat, '.')
plotLayerProbes(layers, data, neuronSize)
cor = np.corrcoef(np.ravel(spikesRates[-1]),
np.ravel(layer_activations[0],
'F')[:numProbes // 2])[0, 1]
self.assertGreater(cor, 0.99)
def test_random_cor(self):
"""
Tests the soft reset mode by comparing activations from an ANN
to spike-rate from the converted SNN. The input and weights
are randomly initialized.
"""
seed = 123
np.random.seed(seed)
plot = False
verbose = False
visualizePartitions = False
logger = None
resetMode = 'soft'
neuronSize = 2 if resetMode == 'soft' else 1
inputShape = (4, 4, 1)
inputScale = 255
inputImage = np.random.randint(int(inputScale * 0.25), int(inputScale),
inputShape)
maxNumSpikes = 100
numSteps = int(np.max(inputImage / 255) * maxNumSpikes)
inputLayer = NxInputLayer(batch_input_shape=(1, ) + inputShape,
vThMant=255,
visualizePartitions=visualizePartitions,
resetMode=resetMode,
probeSpikes=True)
out = inputLayer.input
layers = [inputLayer]
# Conv2D
kernelShape = (3, 3, 1)
kernelScale = 4
# No need to divide by thrGain because spike input receives equal gain.
vThMant = 2**9 - 1
kernel_init = partial(kernel_initializer, kernelScale=kernelScale)
numLayers = 1
for i in range(numLayers):
layer = NxConv2D(filters=kernelShape[-1],
kernel_size=kernelShape[:-1],
vThMant=vThMant,
kernel_initializer=kernel_init,
bias_initializer='ones',
validatePartitions=False,
probeSpikes=True,
activation='relu',
resetMode=resetMode)
layers.append(layer)
out = layer(out)
model = NxModel(inputLayer.input, out, logger=logger)
for layer in layers[1:]:
weights, biases = layer.get_weights()
weights, biases = to_integer(weights, biases, 8,
np.max(weights) // 2)
layer.set_weights([weights, biases])
mapper = model.compileModel()
if verbose:
printLayerMappings(layers, mapper, synapses=True, inputAxons=True)
printLayers(layers)
print(model.summary())
layerProbes = []
for layer in layers:
shape = layer.output_shape[1:]
# Define probes to read out currents.
vProbes = []
sProbes = []
uProbes = []
for i in range(int(np.asscalar(np.prod(shape))) * neuronSize):
vProbes.append(layer[i].probe(ProbableStates.VOLTAGE))
sProbes.append(layer[i].probe(ProbableStates.ACTIVITY))
uProbes.append(layer[i].probe(ProbableStates.CURRENT))
layerProbes.append([uProbes, vProbes, sProbes])
# Set bias currents
for i, b in enumerate(np.ravel(inputImage, 'F')):
inputLayer[i * neuronSize].biasMant = b
inputLayer[i * neuronSize].phase = 2
if verbose:
for layer in layers:
print(getCompartmentStates(layer, neuronSize))
model.run(numSteps)
if verbose:
for layer in layers:
print(getCompartmentStates(layer, neuronSize))
model.disconnect()
data = [[extract(probe) for probe in lp] for lp in layerProbes]
if plot:
plotLayerProbes(layers, data, neuronSize)
spikesRates = []
for i, (layer, d) in enumerate(zip(layers, data)):
sData = d[2][:, (neuronSize - 1)::neuronSize]
shape = layer.output_shape[1:]
spikecount = _data_to_img(sData // 127, shape)
spikesRates.append(spikecount / numSteps)
batchInputImage = np.expand_dims(inputImage, 0)
activations = model.predict(batchInputImage)[0]
if plot:
plt.figure()
plt.plot(inputImage.flatten(), spikesRates[0].flatten(), '.')
plt.show()
plt.figure()
plt.plot(activations.flatten(), spikesRates[-1].flatten(), '.')
plt.show()
plt.figure()
plt.imshow(normalize_image_dims(activations))
plt.show()
plt.figure()
plt.imshow(normalize_image_dims(spikesRates[-1]))
plt.show()
cor = np.corrcoef(np.ravel(spikesRates[-1]), np.ravel(activations))[0,
1]
self.assertGreater(cor, 0.99)
if verbose:
print(cor)
def test_Flatten(self):
"""Test correlation between ANN activations and SNN spikerates.
The network consists of a 3D input layer, followed by a flatten layer,
which has 1-to-1 connections to the output layer. The input pixel
values are set to ascending integers in Fortran style.
This test asserts that the spikerates in the output layer are close to
the ANN activations, by computing the Pearson correlation coefficient.
A perfect correlation cannot be expected due to quantization errors
when approximating ANN activations with discrete spikes. However,
correlations should be higher than 0.99.
"""
visualizePartitions = False
doPlot = False
# Height, width, depth
inputShape = (3, 4, 5)
numInputNeurons = int(np.asscalar(np.prod(inputShape)))
numOutputNeurons = numInputNeurons - 1
inputScale = 255
thrToInputRatio = 2**7
thrGain = 2**6
# No need to divide by thrGain because spike input receives equal gain.
vThMant = 1
vThMantInput = thrToInputRatio * inputScale // thrGain
maxNumSpikes = 100
numSteps = thrToInputRatio * maxNumSpikes
weights = np.eye(numInputNeurons, numOutputNeurons, dtype=int)
biases = np.zeros(numOutputNeurons, int)
nxInput = NxInputLayer(inputShape,
vThMant=vThMantInput,
visualizePartitions=visualizePartitions)
nxLayer = NxDense(numOutputNeurons,
weights=[weights, biases],
vThMant=vThMant,
validatePartitions=True,
probeSpikes=True)
nxModel = NxModel(nxInput.input, nxLayer(NxFlatten()(nxInput.input)))
nxModel.compileModel()
kerasInput = Input(inputShape)
kerasLayer = Dense(numOutputNeurons,
weights=[weights, biases])(Flatten()(kerasInput))
kerasModel = Model(kerasInput, kerasLayer)
# Define probes to read out currents.
sProbes = []
for i in range(numOutputNeurons):
sProbes.append(nxLayer[i].probe(ProbableStates.ACTIVITY))
# Set bias currents
inputImage = np.reshape(np.arange(numInputNeurons), inputShape, 'F')
inputImage = inputImage % 255
for i, b in enumerate(np.ravel(inputImage, 'F')):
nxInput[i].biasMant = b
nxInput[i].phase = 2
nxModel.run(numSteps)
nxModel.disconnect()
data = extract(sProbes)
spikecount = _data_to_img(data // 127, nxLayer.output_shape[1:])
spikerates = spikecount / numSteps * thrToInputRatio
batchInputImage = np.expand_dims(inputImage, 0)
activations = \
kerasModel.predict(batchInputImage)[0] / (vThMant * thrGain)
if doPlot:
plt.figure(3)
plt.imshow(normalize_image_dims(inputImage))
plt.show()
plt.figure(6)
plt.plot(activations.flatten(), spikerates.flatten(), '.')
plt.show()
plt.figure(7)
plt.imshow(normalize_image_dims(activations))
plt.show()
plt.figure(8)
plt.imshow(normalize_image_dims(spikerates))
plt.show()
corr = np.corrcoef(np.ravel(spikerates), np.ravel(activations))[0, 1]
self.assertAlmostEqual(corr,
1,
2,
msg="Correlation between ANN activations "
"and SNN spikerates is too low.")
def test_Conv2DBiases(self):
"""Test correlation between ANN activations and SNN spikerates.
The network consists of an input layer connected to a convolution
layer. The input to the network is a square gray-scale image of random
integers. The weights are also random integers. Biases are non-zero.
This test asserts that the spikerates in the output layer are close to
the ANN activations, by computing the Pearson correlation coefficient.
A perfect correlation cannot be expected due to quantization errors
when approximating ANN activations with discrete spikes. However,
correlations should be higher than 0.99.
"""
def bias_initializer(shape, dtype=None, biasScale=1):
"""Increasing integer bias initializer for Keras layer.
:param list | tuple | np.ndarray shape: Shape of biases.
:param str | type | None dtype: Data type of biases.
:param int biasScale: Scale factor applied to the biases.
:return: Bias tensor.
"""
return k.constant(np.arange(biasScale), dtype, shape)
numFilters = 16
kernelShape = (3, 3)
inputShape = (28, 28, 3)
vThMant = numFilters * 2**4
thrGain = 2**6
visualizePartitions = False
plotUV = False
numSteps = 500
bias_init = partial(bias_initializer, biasScale=numFilters)
layer = NxConv2D(filters=numFilters,
kernel_size=kernelShape,
vThMant=vThMant,
kernel_initializer='zeros',
bias_initializer=bias_init,
validatePartitions=True,
probeSpikes=True,
activation='relu',
strides=(2, 2))
inputLayer = NxInputLayer(batch_input_shape=(1, ) + inputShape,
vThMant=vThMant,
visualizePartitions=visualizePartitions)
model = NxModel(inputLayer.input, layer(inputLayer.input))
model.compileModel()
outputShape = layer.output_shape[1:]
# Define probes to read out currents.
vProbes = []
sProbes = []
for i in range(int(np.asscalar(np.prod(outputShape)))):
vProbes.append(layer[i].probe(ProbableStates.VOLTAGE))
sProbes.append(layer[i].probe(ProbableStates.ACTIVITY))
model.run(numSteps)
model.disconnect()
data = extract(sProbes)
spikecount = _data_to_img(data // 127, outputShape)
spikerates = spikecount / numSteps
batchInputImage = np.expand_dims(np.zeros(inputShape), 0)
activations = model.predict(batchInputImage)[0] / (vThMant * thrGain)
if plotUV:
plt.figure(1)
_plot_stimulus_response(vProbes, sProbes)
plt.show()
plt.figure(2)
plt.plot(activations.flatten(), spikerates.flatten(), '.')
plt.show()
plt.figure(3)
plt.imshow(normalize_image_dims(activations))
plt.show()
plt.figure(4)
plt.imshow(normalize_image_dims(spikerates))
plt.show()
corr = np.corrcoef(np.ravel(spikerates), np.ravel(activations))[0, 1]
self.assertAlmostEqual(corr,
1,
2,
msg="Correlation between ANN activations "
"and SNN spikerates is too low.")
def runCorrelationRandom(layer,
vThMant,
insertFlatten=False,
inputShape=None,
logger=None):
"""Run network to test correlation between ANN and SNN.
:param NxLayer | Layer layer: NxLayer to test.
:param int vThMant: Threshold of ``layer``; used to scale activations.
:param bool insertFlatten: Whether to flatten input before applying it to
``layer``.
:param np.ndarray | tuple | list inputShape: Shape of input to the network.
:param logging.Logger logger: Logger.
:return: Pearson correlation coefficient of ANN activations and SNN rates.
:rtype: float
"""
seed = 123
np.random.seed(seed)
visualizePartitions = False
plotUV = False
if inputShape is None:
inputShape = (7, 7, 1)
numInputNeurons = int(np.asscalar(np.prod(inputShape)))
inputScale = numInputNeurons - 1
thrToInputRatio = 2**7
thrGain = 2**0
vThMantInput = thrToInputRatio * inputScale // thrGain
maxNumSpikes = 100
numSteps = thrToInputRatio * maxNumSpikes
inputImage = np.random.randint(0, inputScale, inputShape)
inputLayer = NxInputLayer(batch_input_shape=(1, ) + inputShape,
vThMant=vThMantInput,
visualizePartitions=visualizePartitions)
out = layer(NxFlatten()(inputLayer.input)) \
if insertFlatten else layer(inputLayer.input)
model = NxModel(inputLayer.input, out, logger=logger)
model.compileModel()
outputShape = layer.output_shape[1:]
# Define probes to read out currents.
vProbes0 = []
for i in range(numInputNeurons):
vProbes0.append(inputLayer[i].probe(ProbableStates.VOLTAGE))
vProbes = []
sProbes = []
for i in range(int(np.asscalar(np.prod(outputShape)))):
vProbes.append(layer[i].probe(ProbableStates.VOLTAGE))
sProbes.append(layer[i].probe(ProbableStates.ACTIVITY))
# Set bias currents
for i, b in enumerate(np.ravel(inputImage, 'F')):
inputLayer[i].biasMant = b
inputLayer[i].phase = 2
model.run(numSteps)
model.disconnect()
data = extract(sProbes)
spikecount = _data_to_img(data // 127, outputShape)
spikerates = spikecount / numSteps * thrToInputRatio
batchInputImage = np.expand_dims(inputImage, 0)
activations = model.predict(batchInputImage)[0] / (vThMant * thrGain)
if plotUV:
plt.figure(1)
_plot_stimulus_response(vProbes0, [])
plt.show()
plt.figure(2)
_plot_stimulus_response(vProbes, sProbes)
plt.show()
plt.figure(3)
plt.imshow(normalize_image_dims(inputImage))
plt.show()
plt.figure(6)
plt.plot(activations.flatten(), spikerates.flatten(), '.')
plt.show()
plt.figure(7)
plt.imshow(normalize_image_dims(activations))
plt.show()
plt.figure(8)
plt.imshow(normalize_image_dims(spikerates))
plt.show()
return np.corrcoef(np.ravel(spikerates), np.ravel(activations))[0, 1]
def runModelFromConfig(argsInput, argsHidden, argsOutput, n=None):
"""A helper function used by test_correlationUniform.
Given a model configuration, this function creates a NxModel, checks
that the model architecture is valid, then partitions, maps and runs the
network with a constant white input image.
The function probes voltage and current states, which can be plotted.
Spikecountss are derived from the current trace.and are used to assert
that the SNN activity equals the ANN activations.
:param InputParams argsInput: Parameters for input layer.
:param LayerParams argsHidden: Parameters for hidden layer.
:param LayerParams argsOutput: Parameters for output layer.
:param int n: The id of the current config. Only used for printing.
"""
# Define stimulus and runtime settings.
# #####################################
inputScale = 1
vThMantInput = 1
vThMant = 1
numSteps = (vThMantInput << 6) // inputScale + 1 + 1 + 1 + 1
biasExp = 0
visualizePartitions = False
plotUV = False
# Build model.
# ############
inputShape = (argsInput.imgHeight, argsInput.imgWidth, argsInput.imgDepth)
batchInputShape = (1, ) + inputShape
inputLayer = NxInputLayer(batch_input_shape=batchInputShape,
biasExp=biasExp,
vThMant=vThMantInput,
visualizePartitions=visualizePartitions)
inputLayer._maxNumCompartments = argsInput.maxNumCxPerCore
hiddenLayer = NxConv2D(argsHidden.kernelDepth,
(argsHidden.kernelHeight, argsHidden.kernelWidth),
strides=(argsHidden.strideY, argsHidden.strideX),
padding=argsHidden.padding,
vThMant=vThMant,
synapseEncoding=argsHidden.encoding,
kernel_initializer='ones',
visualizePartitions=visualizePartitions,
validatePartitions=True)
hiddenLayer._maxNumCompartments = argsHidden.maxNumCxPerCore
outputLayer = NxConv2D(argsOutput.kernelDepth,
(argsOutput.kernelHeight, argsOutput.kernelWidth),
strides=(argsOutput.strideY, argsOutput.strideX),
padding=argsOutput.padding,
vThMant=vThMant,
synapseEncoding=argsOutput.encoding,
kernel_initializer='ones',
validatePartitions=True)
outputLayer._maxNumCompartments = argsOutput.maxNumCxPerCore
hiddenShape = hiddenLayer.compute_output_shape(batchInputShape)
outputShape = outputLayer.compute_output_shape(hiddenShape)
hiddenShape = hiddenShape[1:]
outputShape = outputShape[1:]
if np.any(np.array(hiddenShape) <= 0) or \
np.any(np.array(outputShape) <= 0):
print("Configuration resulted in negative layer shape; skip.")
print("Input layer shape: {}".format(inputShape))
print("Hidden layer shape: {}".format(hiddenShape))
print("Output layer shape: {}".format(outputShape))
return
# Create NxModel from input to output layer.
snn = NxModel(inputLayer.input,
outputLayer(hiddenLayer(inputLayer.input)),
numCandidatesToCompute=1)
snn.compileModel()
# Create plain Keras model from input to hidden layer so we
# can read out hidden layer activations.
model1 = Model(inputLayer.input, hiddenLayer(inputLayer.input))
hiddenInput = Input(hiddenShape)
model2 = Model(hiddenInput, outputLayer(hiddenInput))
# Define probes to read out currents.
# ###################################
u = ProbableStates.CURRENT
v = ProbableStates.VOLTAGE
uProbes1 = []
vProbes1 = []
for i in range(int(np.asscalar(np.prod(hiddenShape)))):
uProbes1.append(hiddenLayer[i].probe(u))
if plotUV:
vProbes1.append(hiddenLayer[i].probe(v))
uProbes2 = []
vProbes2 = []
for i in range(int(np.asscalar(np.prod(outputShape)))):
uProbes2.append(outputLayer[i].probe(u))
if plotUV:
vProbes2.append(outputLayer[i].probe(v))
# Apply input image via bias currents.
# ####################################
inputImage = inputScale * np.ones(inputShape, int)
for i, biasMant in enumerate(np.ravel(inputImage, 'F')):
inputLayer[i].biasMant = biasMant
inputLayer[i].phase = 2
# Run model.
# ##########
snn.run(numSteps)
snn.disconnect()
# Get SNN spike counts.
data1 = extract(uProbes1)
spikecount1 = _data_to_img(data1 // (vThMant * 2**6), hiddenShape)
data2 = extract(uProbes2)
spikecount2 = _data_to_img(data2 // (vThMant * 2**6), outputShape)
# Get ANN activations as target reference.
batchInputImage = np.expand_dims(inputImage, 0)
activations1 = model1.predict(batchInputImage)
activations2 = model2.predict(activations1 > 0)
# Remove batch dim.
activations1 = activations1[0].astype(int)
activations2 = activations2[0].astype(int)
# Plotting.
# #########
if plotUV:
plt.figure()
_plot_stimulus_response(vProbes1, uProbes1)
plt.show()
plt.figure()
_plot_stimulus_response(vProbes2, uProbes2)
plt.show()
# Validation.
# ###########
def getErrorMessage(spikecounts, activations):
"""Generate an error message for failed test.
The error message contains a code snippet to reproduce failure.
:param np.ndarray spikecounts: SNN spike counts of layer.
:param np.ndarray activations: ANN activations of layer.
:return: Error message.
:rtype: str
"""
return ("SNN spikecounts not equal to ANN activations.\n" +
"Spikecounts: \n{}\n\n".format(spikecounts) +
"Activations: \n{}\n\n".format(activations) +
"Use the following code to reproduce the error.\n\n" +
"def test_correlation{}():\n".format('' if n is None else n) +
"\targsInput = {}\n".format(argsInput) +
"\targsHidden = {}\n".format(argsHidden) +
"\targsOutput = {}\n".format(argsOutput) +
"\trunModelFromConfig(argsInput, argsHidden, argsOutput)\n\n")
assert np.array_equal(spikecount1, activations1), \
getErrorMessage(spikecount1, activations1)
assert np.array_equal(spikecount2, activations2), \
getErrorMessage(spikecount2, activations2)