def random_state(rows, cols, connections, random_state, ccore_flag):
sample = read_sample(SIMPLE_SAMPLES.SAMPLE_SIMPLE1)
params = som_parameters()
params.random_state = random_state
network_1 = som(rows, cols, connections, ccore=ccore_flag)
steps_1 = network_1.train(sample, 100, True)
network_2 = som(rows, cols, connections, ccore=ccore_flag)
steps_2 = network_2.train(sample, 100, True)
assert steps_1 == steps_2
assert network_1.weights == network_2.weights
assert network_1.capture_objects == network_2.capture_objects
assert network_1.awards == network_2.awards
def templateTestAwardNeurons(file, rows, cols, time, expected_result, autostop, ccore_flag, parameters = None, **kwargs):
store_load = kwargs.get('store_load', False)
types = [type_conn.func_neighbor, type_conn.grid_eight, type_conn.grid_four, type_conn.honeycomb]
sample = read_sample(file)
if (parameters is None):
parameters = som_parameters()
for stucture in types:
network = som(rows, cols, stucture, parameters, ccore = ccore_flag)
if store_load:
dump_network = pickle.dumps(network)
network = pickle.loads(dump_network)
network.train(sample, time, autostop)
winners = network.get_winner_number()
assert winners == len(expected_result)
if sorted(network.awards) != expected_result:
network.show_network(awards = True)
assert sorted(network.awards) == expected_result
total_capture_points = 0
for points in network.capture_objects:
total_capture_points += len(points)
assert total_capture_points == sum(expected_result)
del network
def templateTestAwardNeurons(file, rows, cols, time, expected_result, autostop, ccore_flag, parameters = None, **kwargs):
store_load = kwargs.get('store_load', False)
types = [type_conn.func_neighbor, type_conn.grid_eight, type_conn.grid_four, type_conn.honeycomb]
sample = read_sample(file)
if (parameters is None):
parameters = som_parameters()
for stucture in types:
network = som(rows, cols, stucture, parameters, ccore = ccore_flag)
if store_load:
dump_network = pickle.dumps(network)
network = pickle.loads(dump_network)
network.train(sample, time, autostop)
winners = network.get_winner_number()
assert winners == len(expected_result)
if sorted(network.awards) != expected_result:
network.show_network(awards = True)
assert sorted(network.awards) == expected_result
total_capture_points = 0
for points in network.capture_objects:
total_capture_points += len(points)
assert total_capture_points == sum(expected_result)
del network
def train(self):
samples = [];
print("Digit images preprocessing...");
for index_digit in range(0, 10, 1):
list_file_digit_sample = IMAGE_DIGIT_SAMPLES.GET_LIST_IMAGE_SAMPLES(index_digit);
for file_name in list_file_digit_sample:
data = read_image(file_name);
image_pattern = rgb2gray(data);
for index_pixel in range(len(image_pattern)):
if (image_pattern[index_pixel] < 128):
image_pattern[index_pixel] = 1;
else:
image_pattern[index_pixel] = 0;
samples += [ image_pattern ];
print("SOM initialization...");
self.__network = som(2, 5, type_conn.grid_four, None, True);
print("SOM training...");
self.__network.train(samples, 300);
print("SOM is ready...");
def template_self_organization(file,
rows,
cols,
time,
structure,
init_type=None,
init_radius=None,
init_rate=None,
umatrix=False,
pmatrix=False,
awards=False):
parameters = som_parameters()
if (init_type is not None):
parameters.init_type = init_type
if (init_radius is not None):
parameters.init_radius = init_radius
if (init_rate is not None):
parameters.init_learn_rate = init_rate
sample = read_sample(file)
network = som(rows, cols, structure, parameters, True)
network.train(sample, time)
network.show_network(False, dataset=False)
if (umatrix is True):
network.show_distance_matrix()
if (pmatrix is True):
network.show_density_matrix()
if (awards is True):
network.show_winner_matrix()
def __init__(self, data, rows, cols, radius):
"""!
@brief Constructor of the double layer oscillatory network SYNC-SOM.
@param[in] data (list): Input data that is presented as list of points (objects), each point should be represented by list or tuple.
@param[in] rows (uint): Rows of neurons (number of neurons in column) in the input layer (self-organized feature map).
@param[in] cols (uint): Columns of neurons (number of neurons in row) in the input later (self-organized feature map).
@param[in] radius (double): Connectivity radius between objects that defines connection between oscillators in the second layer.
"""
self._data = data
self._radius = radius * radius
self._som = som(rows, cols, conn_type=type_conn.grid_four, ccore=False)
# The first (input) later - SOM layer.
self._som_osc_table = list()
self._sync = None
# The second (output) layer - Sync layer.
self._struct = None
# Structure of connections between oscillators in the second layer - Sync layer.
# For convenience
self._analyser = None
def train(self):
samples = [];
print("Digit images preprocessing...");
for index_digit in range(0, 10, 1):
list_file_digit_sample = IMAGE_DIGIT_SAMPLES.GET_LIST_IMAGE_SAMPLES(index_digit);
for file_name in list_file_digit_sample:
data = read_image(file_name);
image_pattern = rgb2gray(data);
for index_pixel in range(len(image_pattern)):
if (image_pattern[index_pixel] < 128):
image_pattern[index_pixel] = 1;
else:
image_pattern[index_pixel] = 0;
samples += [ image_pattern ];
print("SOM initialization...");
self.__network = som(2, 5, type_conn.grid_four, None, True);
print("SOM training...");
self.__network.train(samples, 300);
print("SOM is ready...");
def templateTestAwardNeurons(self,
file,
rows,
cols,
time,
expected_result,
autostop=False,
ccore_flag=False,
parameters=None):
types = [
type_conn.func_neighbor, type_conn.grid_eight, type_conn.grid_four,
type_conn.honeycomb
]
sample = read_sample(file)
if (parameters is None):
parameters = som_parameters()
for stucture in types:
network = som(rows, cols, stucture, parameters, ccore=ccore_flag)
network.train(sample, time, autostop)
if (sorted(network.awards) != expected_result):
network.show_network(awards=True)
print(sorted(network.awards))
assert sorted(network.awards) == expected_result
total_capture_points = 0
for points in network.capture_objects:
total_capture_points += len(points)
assert total_capture_points == sum(expected_result)
del network
def testSomVisualizationByCore(self):
sample = read_sample(SIMPLE_SAMPLES.SAMPLE_SIMPLE4)
parameters = som_parameters()
network = som(5, 5, type_conn.grid_eight, parameters, ccore = True)
network.train(sample, 100, True)
network.show_network()
network.show_winner_matrix()
network.show_distance_matrix()
network.show_density_matrix()
def testSomVisualizationByCore(self):
sample = read_sample(SIMPLE_SAMPLES.SAMPLE_SIMPLE4)
parameters = som_parameters()
network = som(5, 5, type_conn.grid_eight, parameters, ccore=True)
network.train(sample, 100, True)
network.show_network()
network.show_winner_matrix()
network.show_distance_matrix()
network.show_density_matrix()
def process(self):
"""!
@brief Performs cluster analysis by competition between neurons of SOM.
@remark Results of clustering can be obtained using corresponding get methods.
@see get_clusters()
"""
self.__network = som(1, self.__amount_clusters, type_conn.grid_four, None, self.__ccore);
self.__network.train(self.__data_pointer, self.__epouch, True);
def process(self):
"""!
@brief Performs cluster analysis by competition between neurons of SOM.
@remark Results of clustering can be obtained using corresponding get methods.
@see get_clusters()
"""
self.__network = som(1, self.__amount_clusters, type_conn.grid_four, None, self.__ccore);
self.__network.train(self.__data_pointer, self.__epouch, True);
def __init__(self, data, rows, cols):
"""!
@brief Constructor of the double layer oscillatory network SYNC-SOM.
@param[in] data (list): Input data that is presented as list of points (objects), each point should be represented by list or tuple.
@param[in] rows (uint): Rows of neurons (number of neurons in column) in the input layer (self-organized feature map).
@param[in] cols (uint): Columns of neurons (number of neurons in row) in the input later (self-organized feature map).
"""
self._data = data;
self._som = som(rows, cols, data, 100, conn_type = type_conn.grid_four);
self._som_osc_table = list();
def __init__(self, data, rows, cols):
"""!
@brief Constructor of the double layer oscillatory network SYNC-SOM.
@param[in] data (list): Input data that is presented as list of points (objects), each point should be represented by list or tuple.
@param[in] rows (uint): Rows of neurons (number of neurons in column) in the input layer (self-organized feature map).
@param[in] cols (uint): Columns of neurons (number of neurons in row) in the input later (self-organized feature map).
"""
self._data = data
self._som = som(rows, cols, data, 100, conn_type=type_conn.grid_four)
self._som_osc_table = list()
def process(self):
"""!
@brief Performs cluster analysis by competition between neurons of SOM.
@return (somsc) Returns itself (SOM Simple Clustering instance).
@see get_clusters()
"""
self.__network = som(1, self.__amount_clusters, type_conn.grid_four, None, self.__ccore)
self.__network.train(self.__data_pointer, self.__epouch, True)
return self
def testDoubleTrain(self):
sample = read_sample(SIMPLE_SAMPLES.SAMPLE_SIMPLE1);
parameters = som_parameters();
network = som(2, 2, type_conn.grid_eight, parameters, ccore = False);
network.train(sample, 100, False);
network.train(sample, 100, False);
assert sum(network.awards) == len(sample);
total_capture_points = 0;
for points in network.capture_objects:
total_capture_points += len(points);
assert total_capture_points == len(sample);
def testDoubleTrain(self):
sample = read_sample(SIMPLE_SAMPLES.SAMPLE_SIMPLE1);
parameters = som_parameters();
network = som(2, 2, type_conn.grid_eight, parameters, ccore = False);
network.train(sample, 100, False);
network.train(sample, 100, False);
assert sum(network.awards) == len(sample);
total_capture_points = 0;
for points in network.capture_objects:
total_capture_points += len(points);
assert total_capture_points == len(sample);
def templateTestSimulate(connections, ccore_flag):
sample = read_sample(SIMPLE_SAMPLES.SAMPLE_SIMPLE1);
network = som(1, 2, connections, ccore = ccore_flag);
network.train(sample, 100);
expected_winners = [0, 1];
for i in range(len(sample)):
index_winner = network.simulate(sample[i]);
if ( (i == 0) and (index_winner == 1)):
expected_winners = [1, 0];
if (i < 5):
assert expected_winners[0] == index_winner;
else:
assert expected_winners[1] == index_winner;
def templateTestAwardNeurons(self, file, rows, cols, time, expected_result, autostop = False, ccore_flag = False):
types = [type_conn.func_neighbor, type_conn.grid_eight, type_conn.grid_four, type_conn.honeycomb];
sample = read_sample(file);
for stucture in types:
network = som(rows, cols, sample, time, stucture, ccore = ccore_flag);
network.train(autostop);
assert sorted(network.awards) == expected_result;
total_capture_points = 0;
for points in network.capture_objects:
total_capture_points += len(points);
assert total_capture_points == sum(expected_result);
del network;
def process(self):
"""!
@brief Performs cluster analysis by competition between neurons in self-organized map.
@return (somsc) Returns itself (SOM Simple Clustering instance).
@see get_clusters()
"""
params = som_parameters()
params.random_state = self.__random_state
self.__network = som(1, self.__amount_clusters, type_conn.grid_four,
params, self.__ccore)
self.__network.train(self.__data_pointer, self.__epouch, True)
return self
def templateTestWinners(ccore_flag):
types = [type_conn.func_neighbor, type_conn.grid_eight, type_conn.grid_four, type_conn.honeycomb]
sample = read_sample(SIMPLE_SAMPLES.SAMPLE_SIMPLE3)
for stucture in types:
network = som(5, 5, stucture, ccore = ccore_flag)
network.train(sample, 100)
assert sum(network.awards) == 60
points = list()
for i in range(network.size):
if network.awards[i] > 0:
points += network.capture_objects[i]
assert len(points) == len(sample)
points = sorted(points)
for i in range(len(points)):
assert points[i] == i
def templateTestWinners(ccore_flag):
types = [type_conn.func_neighbor, type_conn.grid_eight, type_conn.grid_four, type_conn.honeycomb]
sample = read_sample(SIMPLE_SAMPLES.SAMPLE_SIMPLE3)
for stucture in types:
network = som(5, 5, stucture, ccore = ccore_flag)
network.train(sample, 100)
assert sum(network.awards) == 60
points = list()
for i in range(network.size):
if network.awards[i] > 0:
points += network.capture_objects[i]
assert len(points) == len(sample)
points = sorted(points)
for i in range(len(points)):
assert points[i] == i
def templateTestSimulate(connections, ccore_flag, **kwargs):
store_load = kwargs.get('store_load', False)
sample = read_sample(SIMPLE_SAMPLES.SAMPLE_SIMPLE1)
network = som(1, 2, connections, ccore = ccore_flag)
network.train(sample, 100)
if store_load:
dump_network = pickle.dumps(network)
network = pickle.loads(dump_network)
expected_winners = [0, 1]
for i in range(len(sample)):
index_winner = network.simulate(sample[i])
if (i == 0) and (index_winner == 1):
expected_winners = [1, 0]
if i < 5:
assert expected_winners[0] == index_winner
else:
assert expected_winners[1] == index_winner
def templateTestSimulate(connections, ccore_flag, **kwargs):
store_load = kwargs.get('store_load', False)
sample = read_sample(SIMPLE_SAMPLES.SAMPLE_SIMPLE1)
network = som(1, 2, connections, ccore = ccore_flag)
network.train(sample, 100)
if store_load:
dump_network = pickle.dumps(network)
network = pickle.loads(dump_network)
expected_winners = [0, 1]
for i in range(len(sample)):
index_winner = network.simulate(sample[i])
if (i == 0) and (index_winner == 1):
expected_winners = [1, 0]
if i < 5:
assert expected_winners[0] == index_winner
else:
assert expected_winners[1] == index_winner
def __init__(self, data, rows, cols, radius):
"""!
@brief Constructor of the double layer oscillatory network SYNC-SOM.
@param[in] data (list): Input data that is presented as list of points (objects), each point should be represented by list or tuple.
@param[in] rows (uint): Rows of neurons (number of neurons in column) in the input layer (self-organized feature map).
@param[in] cols (uint): Columns of neurons (number of neurons in row) in the input later (self-organized feature map).
@param[in] radius (double): Connectivity radius between objects that defines connection between oscillators in the second layer.
"""
self._data = data;
self._radius = radius * radius;
self._som = som(rows, cols, conn_type = type_conn.grid_four, ccore = False); # The first (input) later - SOM layer.
self._som_osc_table = list();
self._sync = None; # The second (output) layer - Sync layer.
self._struct = None; # Structure of connections between oscillators in the second layer - Sync layer.
# For convenience
self._analyser = None;
def template_self_organization(file, rows, cols, time, structure, init_type = None, init_radius = None, init_rate = None, umatrix = False, pmatrix = False, awards = False):
parameters = som_parameters();
if (init_type is not None):
parameters.init_type = init_type;
if (init_radius is not None):
parameters.init_radius = init_radius;
if (init_rate is not None):
parameters.init_learn_rate = init_rate;
sample = read_sample(file);
network = som(rows, cols, structure, parameters, True);
network.train(sample, time);
network.show_network(False, dataset = False);
if (umatrix is True):
network.show_distance_matrix();
if (pmatrix is True):
network.show_density_matrix();
if (awards is True):
network.show_winner_matrix();
img_org.item(y, x, 2)])
print('selected_pixel_bgr :\n', bgr_val)
## 5. SOM
sample = bgr_val
# create SOM parameters
parameters = som_parameters() # initialization of initial neuron weights, radius, rateOfLearning, autostop
# create self-organized feature map with size 7x7
rows = 6 # five rows
cols = 6 # five columns
structure = type_conn.honeycomb # each neuron has max. four neighbors.
network = som(rows, cols, structure, parameters)
# train network on sample during 100 epochs.
network.train(sample, 50)
# simulate trained network using randomly modified point from input dataset.
index_point = random.randint(0, len(sample) - 1)
point = sample[index_point] # obtain randomly point from data
point[0] += random.random() * 0.2 # change randomly X-coordinate
point[1] += random.random() * 0.2 # change randomly Y-coordinate
index_winner = network.simulate(point)
# check what are objects from input data are much close to randomly modified.
index_similar_objects = network.capture_objects[index_winner]