def testPredictionMultipleCategories(self):
""" Test the distribution of predictions.
Here, we intend the classifier to learn the associations:
[1,3,5] => bucketIdx 0 & 1
[2,4,6] => bucketIdx 2 & 3
The classifier should get the distribution almost right given enough
repetitions and a small learning rate
"""
c = Classifier(0.001)
SDR1 = SDR(10)
SDR1.sparse = [1, 3, 5]
SDR2 = SDR(10)
SDR2.sparse = [2, 4, 6]
random.seed(42)
for _ in range(5000):
c.learn(pattern=SDR1, classification=[0, 1])
c.learn(pattern=SDR2, classification=[2, 3])
result1 = c.infer(pattern=SDR1)
self.assertAlmostEqual(result1[0], 0.5, places=1)
self.assertAlmostEqual(result1[1], 0.5, places=1)
result2 = c.infer(pattern=SDR2)
self.assertAlmostEqual(result2[2], 0.5, places=1)
self.assertAlmostEqual(result2[3], 0.5, places=1)
def main(parameters=default_parameters, argv=None, verbose=True):
# Load data.
train_labels, train_images, test_labels, test_images = load_ds(
'mnist_784', 10000, shape=[28, 28]) # HTM: ~95.6%
#train_labels, train_images, test_labels, test_images = load_ds('Fashion-MNIST', 10000, shape=[28,28]) # HTM baseline: ~83%
training_data = list(zip(train_images, train_labels))
test_data = list(zip(test_images, test_labels))
random.shuffle(training_data)
# Setup the AI.
enc = SDR(train_images[0].shape)
sp = SpatialPooler(
inputDimensions=enc.dimensions,
columnDimensions=parameters['columnDimensions'],
potentialRadius=parameters['potentialRadius'],
potentialPct=parameters['potentialPct'],
globalInhibition=True,
localAreaDensity=parameters['localAreaDensity'],
stimulusThreshold=int(round(parameters['stimulusThreshold'])),
synPermInactiveDec=parameters['synPermInactiveDec'],
synPermActiveInc=parameters['synPermActiveInc'],
synPermConnected=parameters['synPermConnected'],
minPctOverlapDutyCycle=parameters['minPctOverlapDutyCycle'],
dutyCyclePeriod=int(round(parameters['dutyCyclePeriod'])),
boostStrength=parameters['boostStrength'],
seed=
0, # this is important, 0="random" seed which changes on each invocation
spVerbosity=99,
wrapAround=False)
columns = SDR(sp.getColumnDimensions())
columns_stats = Metrics(columns, 99999999)
sdrc = Classifier()
# Training Loop
for i in range(len(train_images)):
img, lbl = training_data[i]
encode(img, enc)
sp.compute(enc, True, columns)
sdrc.learn(
columns, lbl
) #TODO SDRClassifier could accept string as a label, currently must be int
print(str(sp))
print(str(columns_stats))
# Testing Loop
score = 0
for img, lbl in test_data:
encode(img, enc)
sp.compute(enc, False, columns)
if lbl == np.argmax(sdrc.infer(columns)):
score += 1
score = score / len(test_data)
print('Score:', 100 * score, '%')
return score
def testSingleValue(self):
"""Send same value 10 times and expect high likelihood for prediction."""
classifier = Classifier(alpha=0.5)
# Enough times to perform inference and learn associations
inp = SDR(10)
inp.randomize(.2)
for recordNum in range(10):
classifier.learn(inp, 2)
retval = classifier.infer(inp)
self.assertGreater(retval[2], 0.9)
def testSerialization4(self):
# This test verifies that saveToFile() and loadFromFile with BINARY on Classifier are accessible from Python.
inputData = SDR( 1000 ).randomize( 0.02 )
categories = { 'A': 0, 'B': 1, 'C': 2, 'D': 3 }
c1 = Classifier()
c1.learn(inputData, categories['B'] )
file = "Classifier_test_save.BINARY"
c1.saveToFile(file)
c2 = Classifier()
c2.loadFromFile(file)
result2 = c2.infer( inputData )
self.assertTrue(numpy.argmax( result2 ) == categories['B'])
os.remove(file)
def testPredictionDistributionOverlap(self):
""" Test the distribution of predictions with overlapping input SDRs
Here, we intend the classifier to learn the associations:
SDR1 => bucketIdx 0 (30%)
=> bucketIdx 1 (30%)
=> bucketIdx 2 (40%)
SDR2 => bucketIdx 1 (50%)
=> bucketIdx 3 (50%)
SDR1 and SDR2 has 10% overlaps (2 bits out of 20)
The classifier should get the distribution almost right despite the overlap
"""
c = Classifier(0.0005)
# generate 2 SDRs with 2 shared bits
SDR1 = SDR(100)
SDR2 = SDR(100)
SDR1.randomize(.20)
SDR2.setSDR(SDR1)
SDR2.addNoise(.9)
random.seed(42)
for _ in range(5000):
randomNumber = random.random()
if randomNumber < 0.3:
bucketIdx = 0
elif randomNumber < 0.6:
bucketIdx = 1
else:
bucketIdx = 2
c.learn(SDR1, bucketIdx)
randomNumber = random.random()
if randomNumber < 0.5:
bucketIdx = 1
else:
bucketIdx = 3
c.learn(SDR2, bucketIdx)
result1 = c.infer(SDR1)
self.assertAlmostEqual(result1[0], 0.3, places=1)
self.assertAlmostEqual(result1[1], 0.3, places=1)
self.assertAlmostEqual(result1[2], 0.4, places=1)
result2 = c.infer(SDR2)
self.assertAlmostEqual(result2[1], 0.5, places=1)
self.assertAlmostEqual(result2[3], 0.5, places=1)
def testPredictionDistribution(self):
""" Test the distribution of predictions.
Here, we intend the classifier to learn the associations:
[1,3,5] => bucketIdx 0 (30%)
=> bucketIdx 1 (30%)
=> bucketIdx 2 (40%)
[2,4,6] => bucketIdx 1 (50%)
=> bucketIdx 3 (50%)
The classifier should get the distribution almost right given enough
repetitions and a small learning rate
"""
c = Classifier(alpha=0.001)
SDR1 = SDR(10)
SDR1.sparse = [1, 3, 5]
SDR2 = SDR(10)
SDR2.sparse = [2, 4, 6]
random.seed(42)
for _ in range(5000):
randomNumber = random.random()
if randomNumber < 0.3:
bucketIdx = 0
elif randomNumber < 0.6:
bucketIdx = 1
else:
bucketIdx = 2
c.learn(pattern=SDR1, classification=bucketIdx)
randomNumber = random.random()
if randomNumber < 0.5:
bucketIdx = 1
else:
bucketIdx = 3
c.learn(pattern=SDR2, classification=bucketIdx)
result1 = c.infer(pattern=SDR1)
self.assertAlmostEqual(result1[0], 0.3, places=1)
self.assertAlmostEqual(result1[1], 0.3, places=1)
self.assertAlmostEqual(result1[2], 0.4, places=1)
result2 = c.infer(pattern=SDR2)
self.assertAlmostEqual(result2[1], 0.5, places=1)
self.assertAlmostEqual(result2[3], 0.5, places=1)
def trainNumDecoder(encoder, mini, maxi, noise):
clsr = Classifier()
# loop from smallest to largest number 10 times
for y in range(10):
for x in range(mini, maxi + 1):
# encode current number
encoded = encoder.encode(x)
#corruptSDR(encoded, noise)
# associate encoding to 'class'
# but classes here are numbers.
clsr.learn(encoded, x)
# test every number to see if we learned it
for x in range(mini, maxi + 1):
encoded = encoder.encode(x)
out = decode(clsr, encoded)
if out != x:
print("error in decode training:", out, "->", x)
return clsr
def testExampleUsage(self):
# Make a random SDR and associate it with a category.
inputData = SDR(1000).randomize(0.02)
categories = {'A': 0, 'B': 1, 'C': 2, 'D': 3}
clsr = Classifier()
clsr.learn(inputData, categories['B'])
assert (numpy.argmax(clsr.infer(inputData)) == categories['B'])
# Estimate a scalar value. The Classifier only accepts categories, so
# put real valued inputs into bins (AKA buckets) by subtracting the
# minimum value and dividing by a resolution.
scalar = 567.8
minimum = 500
resolution = 10
clsr.learn(inputData, int((scalar - minimum) / resolution))
assert (numpy.argmax(clsr.infer(inputData)) * resolution +
minimum == 560)
# Predict 1 and 2 time steps into the future.
# Make a sequence of 4 random SDRs, each SDR has 1000 bits and 2% sparsity.
sequence = [SDR(1000).randomize(0.02) for i in range(4)]
# Make category labels for the sequence.
labels = [4, 5, 6, 7]
# Make a Predictor and train it.
pred = Predictor([1, 2])
pred.learn(0, sequence[0], labels[0])
pred.learn(1, sequence[1], labels[1])
pred.learn(2, sequence[2], labels[2])
pred.learn(3, sequence[3], labels[3])
# Give the predictor partial information, and make predictions
# about the future.
pred.reset()
A = pred.infer(0, sequence[0])
assert (numpy.argmax(A[1]) == labels[1])
assert (numpy.argmax(A[2]) == labels[2])
B = pred.infer(1, sequence[1])
assert (numpy.argmax(B[1]) == labels[2])
assert (numpy.argmax(B[2]) == labels[3])
def testOverlapPattern(self):
classifier = Classifier(alpha=10.0)
inp = SDR(10)
inp.randomize(.2)
classifier.learn(pattern=inp, classification=9)
classifier.learn(pattern=inp, classification=9)
inp.addNoise(.5)
retval = classifier.infer(pattern=inp)
# Since overlap - should be previous with high likelihood
self.assertGreater(retval[9], 0.9)
classifier.learn(pattern=inp, classification=2)
classifier.learn(pattern=inp, classification=2)
# Second example: now new value should be more probable than old
retval = classifier.infer(pattern=inp)
self.assertGreater(retval[2], retval[9])
def testSerialization1(self):
# This test verifies that pickle works for pickle of a Classifier
SDR1 = SDR(15); SDR1.sparse = [1, 5, 9]
SDR2 = SDR(15); SDR2.sparse = [0, 6, 9, 11]
SDR3 = SDR(15); SDR3.sparse = [6, 9]
SDR4 = SDR(15); SDR4.sparse = [1, 5, 9]
c1 = Classifier()
c1.learn(pattern=SDR1, classification=4)
c1.learn(pattern=SDR2, classification=5)
c1.learn(pattern=SDR3, classification=5)
c1.learn(pattern=SDR4, classification=4)
c1.learn(pattern=SDR4, classification=4)
serialized = pickle.dumps(c1)
c2 = pickle.loads(serialized)
result1 = c1.infer(SDR1)
result2 = c2.infer(SDR1)
#print(" testSerialization1 result: %.6f, %.6f, %.6f, %.6f, %.6f, %.6f "%( result1[0], result1[1], result1[2], result1[3], result1[4], result1[5]));
self.assertEqual(len(result1), 6)
self.assertAlmostEqual(result1[0], 0.166344, places=5)
self.assertAlmostEqual(result1[1], 0.166344, places=5)
self.assertAlmostEqual(result1[2], 0.166344, places=5)
self.assertAlmostEqual(result1[3], 0.166344, places=5)
self.assertAlmostEqual(result1[4], 0.167847, places=5)
self.assertAlmostEqual(result1[5], 0.166777, places=5)
self.assertEqual(len(result1), len(result2))
for i in range(len(result1)):
self.assertAlmostEqual(result1[i], result2[i], places=5)
def testSerialization3(self):
# This test verifies that saveToFile() and loadFromFile() on Classifier are accessable from Python.
SDR1 = SDR(15); SDR1.sparse = [1, 5, 9]
SDR2 = SDR(15); SDR2.sparse = [0, 6, 9, 11]
SDR3 = SDR(15); SDR3.sparse = [6, 9]
SDR4 = SDR(15); SDR4.sparse = [1, 5, 9]
c1 = Classifier()
c1.learn(pattern=SDR1, classification=4)
c1.learn(pattern=SDR2, classification=5)
c1.learn(pattern=SDR3, classification=5)
c1.learn(pattern=SDR4, classification=4)
# The Predictor now has some data in it, try serialization.
file = "Classifier_test_save.XML"
c1.saveToFile(file, "XML")
c2 = Classifier()
c2.loadFromFile(file, "XML")
os.remove(file)
result1 = c1.infer(SDR1)
result2 = c2.infer(SDR1)
self.assertEqual(len(result1), len(result2))
for i in range(len(result1)):
self.assertAlmostEqual(result1[i], result2[i], places=5)
def testPredictionDistributionContinuousLearning(self):
""" Test continuous learning
First, we intend the classifier to learn the associations:
SDR1 => bucketIdx 0 (30%)
=> bucketIdx 1 (30%)
=> bucketIdx 2 (40%)
SDR2 => bucketIdx 1 (50%)
=> bucketIdx 3 (50%)
After 20000 iterations, we change the association to
SDR1 => bucketIdx 0 (30%)
=> bucketIdx 1 (20%)
=> bucketIdx 3 (40%)
No further training for SDR2
The classifier should adapt continuously and learn new associations for
SDR1, but at the same time remember the old association for SDR2
"""
c = Classifier(0.001)
SDR1 = SDR(10)
SDR1.sparse = [1, 3, 5]
SDR2 = SDR(10)
SDR2.sparse = [2, 4, 6]
random.seed(42)
for _ in range(10000):
randomNumber = random.random()
if randomNumber < 0.3:
bucketIdx = 0
elif randomNumber < 0.6:
bucketIdx = 1
else:
bucketIdx = 2
c.learn(SDR1, bucketIdx)
randomNumber = random.random()
if randomNumber < 0.5:
bucketIdx = 1
else:
bucketIdx = 3
c.learn(SDR2, bucketIdx)
result1 = c.infer(SDR1)
self.assertAlmostEqual(result1[0], 0.3, places=1)
self.assertAlmostEqual(result1[1], 0.3, places=1)
self.assertAlmostEqual(result1[2], 0.4, places=1)
result2 = c.infer(SDR2)
self.assertAlmostEqual(result2[1], 0.5, places=1)
self.assertAlmostEqual(result2[3], 0.5, places=1)
for _ in range(20000):
randomNumber = random.random()
if randomNumber < 0.3:
bucketIdx = 0
elif randomNumber < 0.6:
bucketIdx = 1
else:
bucketIdx = 3
c.learn(SDR1, bucketIdx)
result1new = c.infer(SDR1)
self.assertAlmostEqual(result1new[0], 0.3, places=1)
self.assertAlmostEqual(result1new[1], 0.3, places=1)
self.assertAlmostEqual(result1new[3], 0.4, places=1)
result2new = c.infer(SDR2)
self.assertSequenceEqual(list(result2), list(result2new))
def main(parameters=default_parameters, argv=None, verbose=True):
parser = argparse.ArgumentParser()
parser.add_argument('--data_dir',
type=str,
default=os.path.join(os.path.dirname(__file__), '..',
'..', '..', 'build', 'ThirdParty',
'mnist_data', 'mnist-src'))
args = parser.parse_args(args=argv)
# Load data.
train_labels, train_images, test_labels, test_images = load_mnist(
args.data_dir)
training_data = list(zip(train_images, train_labels))
test_data = list(zip(test_images, test_labels))
random.shuffle(training_data)
random.shuffle(test_data)
# Setup the AI.
enc = SDR((train_images[0].shape))
sp = SpatialPooler(
inputDimensions=enc.dimensions,
columnDimensions=parameters['columnDimensions'],
potentialRadius=parameters['potentialRadius'],
potentialPct=parameters['potentialPct'],
globalInhibition=True,
localAreaDensity=parameters['localAreaDensity'],
stimulusThreshold=int(round(parameters['stimulusThreshold'])),
synPermInactiveDec=parameters['synPermInactiveDec'],
synPermActiveInc=parameters['synPermActiveInc'],
synPermConnected=parameters['synPermConnected'],
minPctOverlapDutyCycle=parameters['minPctOverlapDutyCycle'],
dutyCyclePeriod=int(round(parameters['dutyCyclePeriod'])),
boostStrength=parameters['boostStrength'],
seed=0,
spVerbosity=99,
wrapAround=False)
columns = SDR(sp.getColumnDimensions())
columns_stats = Metrics(columns, 99999999)
sdrc = Classifier()
# Training Loop
for i in range(len(train_images)):
img, lbl = random.choice(training_data)
enc.dense = img >= np.mean(img) # Convert greyscale image to binary.
sp.compute(enc, True, columns)
sdrc.learn(columns, lbl)
print(str(sp))
print(str(columns_stats))
# Testing Loop
score = 0
for img, lbl in test_data:
enc.dense = img >= np.mean(img) # Convert greyscale image to binary.
sp.compute(enc, False, columns)
if lbl == np.argmax(sdrc.infer(columns)):
score += 1
score = score / len(test_data)
print('Score:', 100 * score, '%')
return score