def __init__(
self,
configs=(), # block configuration
num_l=2, # number of labels
min_val=0.0, # minimum input value
max_val=1.0, # maximum input value
num_i=1024, # number of input statelets
num_ai=128, # number of active input statelets
num_s=512, # number of statelets
num_as=8, # number of active statelets
pct_pool=0.8, # pooling percentage
pct_conn=0.5, # initially connected percentage
pct_learn=0.3): # learn percentage
PERM_THR = 20
PERM_INC = 2
PERM_DEC = 1
# seed the random number generator
# bb.seed(0) # TODO: fix seeding
self.st = ScalarTransformer(min_val, max_val, num_i, num_ai)
self.pc = PatternClassifier(num_l, num_s, num_as, PERM_THR, PERM_INC,
PERM_DEC, pct_pool, pct_conn, pct_learn)
self.pc.input.add_child(self.st.output, 0)
self.pc.init()
def __init__(
self,
configs=(), # block configuration
labels=(0, 1), # labels
min_val=-1.0, # ScalarEncoder minimum input value
max_val=1.0, # ScalarEncoder maximum input value
num_i=1024, # ScalarEncoder number of statelets
num_ai=128, # ScalarEncoder number of active statelets
num_s=32, # PatternClassifier number of statelets
num_as=8, # PatternClassifier number of active statelets
pct_pool=0.8, # PatternClassifier pool percentage
pct_conn=0.5, # PatternClassifier initial connection percentage
pct_learn=0.25): # PatternClassifier learn percentage
# seed the random number generator
bb.seed(0)
# build blocks from config descriptions if given
blocks = get_blocks(configs)
self.encoders = blocks["encoders"]
self.pc = blocks["pattern_classifier"]
if len(self.encoders) == 0:
self.encoders.append(ScalarEncoder(min_val, max_val, num_i,
num_ai))
if self.pc == None:
num_l = len(labels)
self.pc = PatternClassifier(labels, num_s, num_as, 20, 2, 1,
pct_pool, pct_conn, pct_learn)
for encoder in self.encoders:
self.pc.input.add_child(encoder.output)
class Classifier():
def __init__(
self,
configs=(), # block configuration
num_l=2, # number of labels
min_val=0.0, # minimum input value
max_val=1.0, # maximum input value
num_i=1024, # number of input statelets
num_ai=128, # number of active input statelets
num_s=512, # number of statelets
num_as=8, # number of active statelets
pct_pool=0.8, # pooling percentage
pct_conn=0.5, # initially connected percentage
pct_learn=0.3): # learn percentage
PERM_THR = 20
PERM_INC = 2
PERM_DEC = 1
# seed the random number generator
# bb.seed(0) # TODO: fix seeding
self.st = ScalarTransformer(min_val, max_val, num_i, num_ai)
self.pc = PatternClassifier(num_l, num_s, num_as, PERM_THR, PERM_INC,
PERM_DEC, pct_pool, pct_conn, pct_learn)
self.pc.input.add_child(self.st.output, 0)
self.pc.init()
#def save(self, path='./', name='classifier'):
# self.pc.save(path + name + "_pc.bin")
#def load(self, path='./', name='classifier'):
# self.pc.load(path + name + "_pc.bin")
def fit(self, value=0.0, label=0):
self.st.set_value(value)
self.pc.set_label(label)
self.st.feedforward()
self.pc.feedforward(learn=True)
return self.pc.get_probabilities()
def predict(self, value=0.0):
self.st.set_value(value)
self.st.feedforward()
self.pc.feedforward(learn=False)
return self.pc.get_probabilities()
def fit(self, X, y):
"""Fit the model using X as training data and y as target values
Parameters
----------
X : {array-like, sparse matrix}
Training data. If array or matrix, shape [num_samples, num_features],
y : {array-like, sparse matrix}
Target values of shape = [num_samples] or [num_samples, num_outputs]
Returns
-------
y_new : array, shape (num_samples, num_outputs)
Classified data
"""
X, y = check_X_y(X, y, multi_output=True)
#X, y = check_X_y(X, y)
if y.ndim == 1 or y.ndim == 2 and y.shape[1] == 1:
if y.ndim != 1:
warnings.warn(
"A column-vector y was passed when a 1d array "
"was expected. Please change the shape of y to "
"(num_samples, ), for example using ravel().",
DataConversionWarning,
stacklevel=2)
self.outputs_2d_ = False
y = y.reshape((-1, 1))
else:
self.outputs_2d_ = True
check_classification_targets(y)
self.classes_ = []
self.class_indices_ = []
self._y = np.empty(y.shape, dtype=np.int)
for k in range(self._y.shape[1]):
classes, self._y[:, k] = np.unique(y[:, k], return_inverse=True)
self.classes_.append(classes)
#class_names, self.class_indices_ = np.unique(self.classes_, return_inverse=True)
#self.labels = np.array([str(val) for val in self.class_indices_])
if not self.outputs_2d_:
self.classes_ = self.classes_[0]
self._y = self._y.ravel()
# class for undefined and unseen values
if self.use_undefined_class:
undef_class = np.max(self.classes_) + 1
self.classes_ = np.append(self.classes_, undef_class)
#print("with undefined class:", self.classes_)
# update labels in DPC config
self.dpc_config['num_l'] = len(self.classes_)
#print("classes_:", self.classes_)
#print("self._y:", self._y)
# instantiate the HyperGrid Transform
self.gridEncoder = HyperGridTransform(**self.hgt_config)
# fit the HyperGrid transform to the data
X_new = self.gridEncoder.fit_transform(X)
#print("HyperGrid Parameters")
#print(self.gridEncoder.subspace_periods)
#print(self.gridEncoder.subspace_vectors)
# get the number of bits being used for transformed output
self.num_bits = self.gridEncoder.num_bits
self.num_act_bits = self.gridEncoder.num_act_bits
#print("BlankBlock:")
#print("num_bits:", self.num_bits)
#print("num_act_bits:", self.num_act_bits)
# Blank Block to hold the hypergrid output
self.blankBlock = BlankBlock(num_s=self.num_bits)
# Create PatternClassifier block
self.dpc = PatternClassifier(**self.dpc_config)
#print("PatternClassifier:")
#self.dpc.print_parameters()
# connect blocks together
self.dpc.input.add_child(self.blankBlock.output, 0)
# Train Network
probs = self._fit(X_new, self._y)
#print("data:", X_new)
#print("labels:", self._y)
#print("training:", probs)
return self
class BBClassifier:
def __init__(
self,
# Training Arguments
num_epochs=3,
use_undefined_class=False,
# Distributed Pattern Classifier Arguments
num_l=2, # number of labels
num_s=512, # number of statelets
num_as=8, # number of active statelets
pct_pool=0.8, # percent pooled
pct_conn=0.8, # percent initially connected
pct_learn=0.3, # percent learn
seed=0,
# HyperGrid Transform Arguments
num_bins=4,
num_acts=1,
num_grids=64,
num_subspace_dims=1,
origin=None,
num_input_dims=None,
max_period=2.0,
min_period=0.05,
use_orthogonal_bases=False,
use_normal_dist_bases=False,
use_standard_bases=False,
set_bases=None,
set_periods=None,
use_random_uniform_periods=False,
use_evenly_spaced_periods=False,
random_state=None):
"""Classify N-dimensional inputs using Hypergrid Transform and Distributed Pattern Classifier
Parameters
----------
num_epochs: integer
Number of training epochs
use_undefined_class: boolean
Whether to reserve a class for test samples that have no training data
num_s: integer
Number of class detectors to allocate for the distributed pattern classifier
num_as: integer
Number of active class detectors per time step for the distributed pattern classifier
pct_pool: float, Between 0.0 and 1.0
Percentage of random bits an individual detector has potential access to.
pct_conn: float, Between 0.0 and pct_pool
Percentage of random bits an individual detector is currently connected to.
pct_learn: float, Between 0.0 and 1.0
Percentage of bits to update when training occurs.
num_bins: integer
Number of bins to create for each grid.
num_acts: integer
Number of contiguous bins to activate along each subspace dimension for each grid.
num_grids: integer
Number of grids to generate.
num_subspace_dims: integer
Dimensionality of subspaces to map input to
origin: array-like, shape {num_features}
Point of origin in input space for embedding a sample into the grids.
max_period: float
Maximum bound on grid period
min_period: float
Minimum bound on grid period
use_orthogonal_bases: boolean
Generate random orthogonal basis vectors for each subspace
use_normal_dist_bases: boolean
Generate normal distribution of basis vectors, points sampled on a sphere
use_standard_bases: boolean
Use randomly selected standard basis vectors for each grid
set_bases: array-like, shape {num_grids, num_subspace_dims, num_features}
Use manually specified subspace basis vectors for each grid
set_periods: array-like, shape {num_grids, num_subspace_dims}
Use manually specified periods for each grid and its subspaces
use_random_uniform_periods: boolean
Use random periods for subspace grids
use_evenly_spaced_periods: boolean
Use evenly spaced periods for subspace grids over the interval min_period to max_period
random_state: integer
Seed for random number generators
"""
self.num_epochs = num_epochs
self.use_undefined_class = use_undefined_class
self._y = []
self.classes_ = np.array([])
self.outputs_2d_ = False
self.dpc_config = dict(num_l=num_l,
num_s=num_s,
num_as=num_as,
perm_thr=20,
perm_inc=2,
perm_dec=1,
pct_pool=pct_pool,
pct_conn=pct_conn,
pct_learn=pct_learn,
num_t=2,
seed=seed)
self.hgt_config = dict(
num_grids=num_grids,
num_bins=num_bins,
num_acts=num_acts,
num_subspace_dims=num_subspace_dims,
origin=origin,
num_input_dims=num_input_dims,
max_period=max_period,
min_period=min_period,
use_normal_dist_bases=use_normal_dist_bases,
use_standard_bases=use_standard_bases,
use_orthogonal_bases=use_orthogonal_bases,
use_evenly_spaced_periods=use_evenly_spaced_periods,
use_random_uniform_periods=use_random_uniform_periods,
set_bases=set_bases,
set_periods=set_periods,
random_state=random_state,
flatten_output=True)
def __del__(self):
pass
def _generate_config(self, num_labels=2):
# connect BlankBlock to DPC
# optionally add PatternPooler
pass
def reset(self):
pass
def fit(self, X, y):
"""Fit the model using X as training data and y as target values
Parameters
----------
X : {array-like, sparse matrix}
Training data. If array or matrix, shape [num_samples, num_features],
y : {array-like, sparse matrix}
Target values of shape = [num_samples] or [num_samples, num_outputs]
Returns
-------
y_new : array, shape (num_samples, num_outputs)
Classified data
"""
X, y = check_X_y(X, y, multi_output=True)
#X, y = check_X_y(X, y)
if y.ndim == 1 or y.ndim == 2 and y.shape[1] == 1:
if y.ndim != 1:
warnings.warn(
"A column-vector y was passed when a 1d array "
"was expected. Please change the shape of y to "
"(num_samples, ), for example using ravel().",
DataConversionWarning,
stacklevel=2)
self.outputs_2d_ = False
y = y.reshape((-1, 1))
else:
self.outputs_2d_ = True
check_classification_targets(y)
self.classes_ = []
self.class_indices_ = []
self._y = np.empty(y.shape, dtype=np.int)
for k in range(self._y.shape[1]):
classes, self._y[:, k] = np.unique(y[:, k], return_inverse=True)
self.classes_.append(classes)
#class_names, self.class_indices_ = np.unique(self.classes_, return_inverse=True)
#self.labels = np.array([str(val) for val in self.class_indices_])
if not self.outputs_2d_:
self.classes_ = self.classes_[0]
self._y = self._y.ravel()
# class for undefined and unseen values
if self.use_undefined_class:
undef_class = np.max(self.classes_) + 1
self.classes_ = np.append(self.classes_, undef_class)
#print("with undefined class:", self.classes_)
# update labels in DPC config
self.dpc_config['num_l'] = len(self.classes_)
#print("classes_:", self.classes_)
#print("self._y:", self._y)
# instantiate the HyperGrid Transform
self.gridEncoder = HyperGridTransform(**self.hgt_config)
# fit the HyperGrid transform to the data
X_new = self.gridEncoder.fit_transform(X)
#print("HyperGrid Parameters")
#print(self.gridEncoder.subspace_periods)
#print(self.gridEncoder.subspace_vectors)
# get the number of bits being used for transformed output
self.num_bits = self.gridEncoder.num_bits
self.num_act_bits = self.gridEncoder.num_act_bits
#print("BlankBlock:")
#print("num_bits:", self.num_bits)
#print("num_act_bits:", self.num_act_bits)
# Blank Block to hold the hypergrid output
self.blankBlock = BlankBlock(num_s=self.num_bits)
# Create PatternClassifier block
self.dpc = PatternClassifier(**self.dpc_config)
#print("PatternClassifier:")
#self.dpc.print_parameters()
# connect blocks together
self.dpc.input.add_child(self.blankBlock.output, 0)
# Train Network
probs = self._fit(X_new, self._y)
#print("data:", X_new)
#print("labels:", self._y)
#print("training:", probs)
return self
def predict(self, X):
"""Predict the class labels for the provided data
Parameters
----------
X : array-like, shape (num_query, num_features)
Test samples.
Returns
-------
y : array of shape [num_samples] or [num_samples, num_outputs]
Class labels for each data sample.
"""
X = check_array(X)
num_samples, num_features = X.shape
if self._y.ndim == 1 or self._y.ndim == 2 and self._y.shape[1] == 1:
num_outputs = 1
else:
num_outputs = self._y.shape[1]
# transform data
X_new = self.gridEncoder.transform(X)
probabilities = self._predict(X_new)
classes_ = self.classes_
if not self.outputs_2d_:
classes_ = [self.classes_]
#classes_ = self.labels
#if not self.outputs_2d_:
# classes_ = [self.labels]
y_pred = np.empty((num_samples, num_outputs), dtype=classes_[0].dtype)
# print("y_pred:", type(y_pred), y_pred.shape)
# print("y_pred:", y_pred)
for k in range(num_samples):
py = probabilities[k, :]
y_best = np.argmax(py)
# print("y_best =", y_best)
y_pred[k, :] = y_best
if not self.outputs_2d_:
y_pred = y_pred.ravel()
# print("y_pred:", type(y_pred), y_pred.shape)
# print("y_pred:", y_pred)
return y_pred
def predict_proba(self, X):
"""Return probability estimates for the test data X.
Parameters
----------
X : array-like, shape (num_query, num_features), \
or (num_query, num_indexed) if metric == 'precomputed'
Test samples.
Returns
-------
p : array of shape = [num_samples, num_classes], or a list of num_outputs
of such arrays if num_outputs > 1.
The class probabilities of the input samples. Classes are ordered
by lexicographic order.
"""
X = check_array(X)
num_samples, num_features = X.shape
if self._y.ndim == 1 or self._y.ndim == 2 and self._y.shape[1] == 1:
num_outputs = 1
else:
num_outputs = self._y.shape[1]
# transform data
X_new = self.gridEncoder.transform(X)
return self._predict(X_new)
def score(self, X, y, sample_weight=None):
"""Returns the mean accuracy on the given test data and labels.
In multi-label classification, this is the subset accuracy
which is a harsh metric since you require for each sample that
each label set be correctly predicted.
Parameters
----------
X : array-like, shape = (num_samples, num_features)
Test samples.
y : array-like, shape = (num_samples) or (num_samples, num_outputs)
True labels for X.
sample_weight : array-like, shape = [num_samples], optional
Sample weights.
Returns
-------
score : float
Mean accuracy of self.predict(X) wrt. y.
"""
from sklearn.metrics import accuracy_score
return accuracy_score(y, self.predict(X), sample_weight=sample_weight)
def _fit(self, X, y):
probabilities = []
# train pattern classifier
for i in range(self.num_epochs):
epoch_probs = []
# Train Network
#t0 = time.time()
for k in range(y.shape[0]):
input = X[k, :]
target = y[k]
self.blankBlock.output.bits = input
self.blankBlock.feedforward()
self.dpc.set_label(target)
self.dpc.feedforward(learn=True)
curr_prob = self.dpc.get_probabilities()
epoch_probs.append(curr_prob)
#t1 = time.time()
#print("train epoch time = %fs with size %d" % ((t1 - t0), y.shape[0]))
probabilities.append(epoch_probs)
return np.asarray(probabilities)
def _predict(self, X):
probabilities = []
# num_points = 1000
num_points = X.shape[0]
#t0 = time.time()
for k in range(X.shape[0]):
input = X[k, :]
self.blankBlock.output.bits = input
self.blankBlock.feedforward()
self.dpc.feedforward(learn=False)
curr_prob = self.dpc.get_probabilities()
probabilities.append(curr_prob)
#t1 = time.time()
#print("%d points, time = %fs" % (num_points, (t1 - t0)))
return np.asarray(probabilities)
hog_fd = hog(x_train[0],
orientations=orientations,
pixels_per_cell=pixels_per_cell,
cells_per_block=cells_per_block,
visualize=False,
multichannel=False,
feature_vector=True)
# setup BrainBlocks architecture
blankblock = BlankBlock(num_s=len(hog_fd))
classifier = PatternClassifier(
labels=(0,1,2,3,4,5,6,7,8,9),
num_s=num_s,
num_as=9,
perm_thr=20,
perm_inc=2,
perm_dec=1,
pct_pool=0.8,
pct_conn=1.0,
pct_learn=0.25)
classifier.input.add_child(blankblock.output)
# train BrainBlocks classifier
bb_train_time = 0
print("Training...", flush=True)
for _ in range(num_epochs):
for i in range(num_trains):
hog_fd = hog(x_train[i],
orientations=orientations,
num_as=8) # number of active statelets
sl = SequenceLearner(
num_spc=10, # number of statelets per column
num_dps=10, # number of coincidence detectors per statelet
num_rpd=12, # number of receptors per coincidence detector
d_thresh=6, # coincidence detector threshold
perm_thr=1, # receptor permanence threshold
perm_inc=1, # receptor permanence increment
perm_dec=0) # receptor permanence decrement
pc = PatternClassifier(
labels=labels, # user-defined labels
num_s=640, # number of statelets
num_as=8, # number of active statelets
perm_thr=20, # receptor permanence threshold
perm_inc=2, # receptor permanence increment
perm_dec=1, # receptor permanence decrement
pct_pool=0.8, # pooling percentage
pct_conn=0.8, # initially connected percentage
pct_learn=0.25) # learn percentage 0.25
sl.input.add_child(e.output)
pc.input.add_child(sl.output)
aed = AbnormalEventDetector(5, 5)
print('val scr lbl prob ae output_active_statelets')
for i in range(len(values)):
e.compute(value=values[i])
sl.compute(learn=True)
num_trains = len(x_train)
num_tests = len(x_test)
pixel_thresh = 128 # from 0 to 255
# setup BrainBlocks classifier architecture
input_block = BlankBlock(num_s=784)
labels = (0, 1, 2, 3, 4, 5, 6, 7, 8, 9)
num_statelets = 1000
classifier = PatternClassifier(labels=labels,
num_s=num_statelets,
num_as=10,
perm_thr=20,
perm_inc=2,
perm_dec=1,
pct_pool=0.1,
pct_conn=1.0,
pct_learn=0.25)
classifier.input.add_child(input_block.output)
# train BrainBLocks classifier
print("Training...", flush=True)
t0 = time.time()
for i in range(num_trains):
bitimage = binarize_image(x_train[i], pixel_thresh)
input_block.output.bits = flatten_image(bitimage)
classifier.compute(y_train[i], learn=True)
def test_pattern_classifier():
e = SymbolsEncoder(
max_symbols=8, # maximum number of symbols
num_s=1024) # number of statelets
pc = PatternClassifier(
labels=(0, 1), # user-defined labels
num_s=512, # number of statelets
num_as=8, # number of active statelets
perm_thr=20, # receptor permanence threshold
perm_inc=2, # receptor permanence increment
perm_dec=1, # receptor permanence decrement
pct_pool=0.8, # pooling percentage
pct_conn=0.5, # initially connected percentage
pct_learn=0.25) # learn percentage
pc.input.add_child(e.output)
for _ in range(10):
e.compute(value=0)
pc.compute(label=0, learn=True)
e.compute(value=1)
pc.compute(label=1, learn=True)
e.compute(value=0)
pc.compute(learn=False)
actual_probs = np.array(pc.get_probabilities())
expect_probs = np.array([1.0, 0.0])
np.testing.assert_array_equal(actual_probs, expect_probs)
e.compute(value=1)
pc.compute(learn=False)
actual_probs = np.array(pc.get_probabilities())
expect_probs = np.array([0.0, 1.0])
np.testing.assert_array_equal(actual_probs, expect_probs)
run_t0 = time.time()
# Define scenario parameters
num_epochs = 1
num_trains = len(x_train)
num_tests = len(x_test)
pixel_thresh = 128 # from 0 to 255
num_s = 8000
# Setup BrainBlocks architecture
blankblock = BlankBlock(num_s=784)
classifier = PatternClassifier(num_l=10,
num_s=num_s,
num_as=9,
perm_thr=20,
perm_inc=2,
perm_dec=1,
pct_pool=0.8,
pct_conn=1.0,
pct_learn=0.3)
classifier.input.add_child(blankblock.output, 0)
# Train BrainBlocks classifier
bb_train_time = 0
print("Training...", flush=True)
for _ in range(num_epochs):
for i in range(num_trains):
bitimage = binarize(x_train[i], pixel_thresh)
blankblock.output.bits = flatten(bitimage)
# retrieve the integer classes from above
int_classes = [k for k in range(len(le.classes_))]
# define blocks
se = ScalarEncoder(
min_val=-1.0, # minimum input value
max_val=1.0, # maximum input value
num_s=1024, # number of statelets
num_as=128) # number of active statelets
pp = PatternClassifier(
labels=int_classes, # user-defined labels
num_s=512, # number of statelets
num_as=8, # number of active statelets
perm_thr=20, # receptor permanence threshold
perm_inc=2, # receptor permanence increment
perm_dec=1, # receptor permanence decrement
pct_pool=0.8, # pooling percentage
pct_conn=0.5, # initially connected percentage
pct_learn=0.25) # learn percentage
# connect blocks
pp.input.add_child(se.output)
# fit
for i in range(len(x_trains)):
se.compute(x_trains[i])
pp.compute(y_trains_ints[i], learn=True)
# predict
probs = []
class Classifier():
def __init__(
self,
configs=(), # block configuration
labels=(0, 1), # labels
min_val=-1.0, # ScalarEncoder minimum input value
max_val=1.0, # ScalarEncoder maximum input value
num_i=1024, # ScalarEncoder number of statelets
num_ai=128, # ScalarEncoder number of active statelets
num_s=32, # PatternClassifier number of statelets
num_as=8, # PatternClassifier number of active statelets
pct_pool=0.8, # PatternClassifier pool percentage
pct_conn=0.5, # PatternClassifier initial connection percentage
pct_learn=0.25): # PatternClassifier learn percentage
# seed the random number generator
bb.seed(0)
# build blocks from config descriptions if given
blocks = get_blocks(configs)
self.encoders = blocks["encoders"]
self.pc = blocks["pattern_classifier"]
if len(self.encoders) == 0:
self.encoders.append(ScalarEncoder(min_val, max_val, num_i,
num_ai))
if self.pc == None:
num_l = len(labels)
self.pc = PatternClassifier(labels, num_s, num_as, 20, 2, 1,
pct_pool, pct_conn, pct_learn)
for encoder in self.encoders:
self.pc.input.add_child(encoder.output)
def print_parameters(self):
for encoder in self.encoders:
encoder.print_parameters()
self.pc.print_parameters()
def save_memories(self, path='./', name='classifier'):
self.pc.save_memories(path + name + "_pc.bin")
def load_memories(self, path='./', name='classifier'):
self.pc.load_memories(path + name + "_pc.bin")
def fit(self, inputs=(), labels=()):
probs = []
num_steps = 0xFFFFFFFF
num_measurands = len(inputs)
num_encoders = len(self.encoders)
if num_measurands != num_encoders:
print("Warning: compute() num_measurands != num_encoders")
return probs
for input in inputs:
len_input = len(input)
if len_input < num_steps:
num_steps = len_input
for s in range(num_steps):
for e in range(num_encoders):
self.encoders[e].compute(inputs[e][s])
self.pc.compute(labels[s], learn=True)
probs.append(self.pc.get_probabilities())
return probs
def predict(self, inputs=()):
probs = []
num_steps = 0xFFFFFFFF
num_measurands = len(inputs)
num_encoders = len(self.encoders)
if num_measurands != num_encoders:
print("Warning: compute() num_measurands != num_encoders")
return probs
for input in inputs:
len_input = len(input)
if len_input < num_steps:
num_steps = len_input
for s in range(num_steps):
for e in range(num_encoders):
self.encoders[e].compute(inputs[e][s])
self.pc.compute(0, learn=False)
probs.append(self.pc.get_probabilities())
return probs
def test_pattern_classifier():
e = SymbolsEncoder(max_symbols=8, num_s=1024)
pc = PatternClassifier(labels=(0, 1),
num_s=512,
num_as=8,
perm_thr=20,
perm_inc=2,
perm_dec=1,
pct_pool=0.8,
pct_conn=0.5,
pct_learn=0.25)
pc.input.add_child(e.output)
for _ in range(10):
e.compute(0)
pc.compute(0, True)
e.compute(1)
pc.compute(1, True)
e.compute(0)
pc.compute(0, False)
actual_probs = np.array(pc.get_probabilities())
expect_probs = np.array([1.0, 0.0])
np.testing.assert_array_equal(actual_probs, expect_probs)
e.compute(1)
pc.compute(1, False)
actual_probs = np.array(pc.get_probabilities())
expect_probs = np.array([0.0, 1.0])
np.testing.assert_array_equal(actual_probs, expect_probs)
le = preprocessing.LabelEncoder()
le.fit(y_trains)
y_trains_ints = le.transform(y_trains)
# Setup blocks
st = ScalarTransformer(
min_val=-1.0, # minimum input value
max_val=1.0, # maximum input value
num_s=1024, # number of statelets
num_as=128) # number of active statelets
pp = PatternClassifier(
num_l=2, # number of labels
num_s=512, # number of statelets
num_as=8, # number of active statelets
perm_thr=20, # receptor permanence threshold
perm_inc=2, # receptor permanence increment
perm_dec=1, # receptor permanence decrement
pct_pool=0.8, # percent pooled
pct_conn=0.5, # percent initially connected
pct_learn=0.3) # percent learn
# Connect blocks
pp.input.add_child(st.output, 0)
# Fit
for i in range(len(x_trains)):
st.set_value(x_trains[i])
pp.set_label(y_trains_ints[i])
st.feedforward()
pp.feedforward(learn=True)
num_as=8) # number of active statelets
sl = SequenceLearner(
num_spc=10, # number of statelets per column
num_dps=10, # number of coincidence detectors per statelet
num_rpd=12, # number of receptors per coincidence detector
d_thresh=6, # coincidence detector threshold
perm_thr=1, # receptor permanence threshold
perm_inc=1, # receptor permanence increment
perm_dec=0) # receptor permanence decrement
pc = PatternClassifier(
num_l=20, # number of labels
num_s=640, # number of statelets
num_as=8, # number of active statelets
perm_thr=20, # receptor permanence threshold
perm_inc=2, # receptor permanence increment
perm_dec=1, # receptor permanence decrement
pct_pool=0.8, # pooling percentage
pct_conn=0.8, # initially connected percentage
pct_learn=0.25) # learn percentage 0.25
sl.input.add_child(st.output)
pc.input.add_child(sl.output)
aed = AbnormalEventDetector(5, 5)
print('val scr lbl prob ae output_active_statelets')
for i in range(len(values)):
st.set_value(values[i])