def do_spatial_learning(self, patterns, ignore_empty=False):
"""Learn spatial patterns.
Args:
patterns: a list of nodes' outputs from the previous layer
or the sensor
ignore_empty: a boolean indicating whether to ignore empty patterns
"""
# prepare patterns -- merge child outputs if necessary
merged_patterns = []
flatten = self.nodes[0].sp.algorithm is not 'product'
for i in range(np.prod(self.size)):
node_patterns = itemgetter(*self.links[i])(patterns)
if flatten:
node_patterns = np.asarray(utils.flatten_list(node_patterns))
else:
node_patterns = np.asarray(node_patterns)
merged_patterns.append(node_patterns)
if self.node_cloning:
# one master node receives all the input
for p in merged_patterns:
# ignore empty patterns
if self._is_first_level() and ignore_empty:
if np.min(p) == np.max(p):
continue
self.nodes[0].do_spatial_learning(p)
else:
# each node gets its own input
for n, p in zip(self.nodes, merged_patterns):
# ignore empty patterns
if self._is_first_level() and ignore_empty:
if np.min(p) == np.max(p):
continue
n.do_spatial_learning(p)
def do_spatial_learning(self, patterns, ignore_empty=False):
"""Learn spatial patterns.
Args:
patterns: a list of nodes' outputs from the previous layer
or the sensor
ignore_empty: a boolean indicating whether to ignore empty patterns
"""
# prepare patterns -- merge child outputs if necessary
merged_patterns = []
flatten = self.nodes[0].sp.algorithm is not 'product'
for i in range(np.prod(self.size)):
node_patterns = itemgetter(*self.links[i])(patterns)
if flatten:
node_patterns = np.asarray(utils.flatten_list(node_patterns))
else:
node_patterns = np.asarray(node_patterns)
merged_patterns.append(node_patterns)
if self.node_cloning:
# one master node receives all the input
for p in merged_patterns:
# ignore empty patterns
if self._is_first_level() and ignore_empty:
if np.min(p) == np.max(p):
continue
self.nodes[0].do_spatial_learning(p)
else:
# each node gets its own input
for n, p in zip(self.nodes, merged_patterns):
# ignore empty patterns
if self._is_first_level() and ignore_empty:
if np.min(p) == np.max(p):
continue
n.do_spatial_learning(p)
def testFlatten(self):
nested_list = [[1, 2, 3], [4, 5, 6]]
correct_list = [1, 2, 3, 4, 5, 6]
flatten_list = utils.flatten_list(nested_list)
self.assertEqual(len(flatten_list), 6)
self.assertListEqual(flatten_list, correct_list)
def do_temporal_learning(self, patterns, is_reset=False):
for i, n in enumerate(self.nodes):
if is_reset:
n.do_temporal_learning(None, is_reset=True)
else:
node_patterns = itemgetter(*self.links[i])(patterns)
node_patterns = np.asarray(utils.flatten_list(node_patterns))
n.do_temporal_learning(node_patterns)
def do_temporal_learning(self, patterns, is_reset=False):
for i, n in enumerate(self.nodes):
if is_reset:
n.do_temporal_learning(None, is_reset=True)
else:
node_patterns = itemgetter(*self.links[i])(patterns)
node_patterns = np.asarray(utils.flatten_list(node_patterns))
n.do_temporal_learning(node_patterns)
def testFlatten(self):
nested_list = [[1, 2, 3], [4, 5, 6]]
correct_list = [1, 2, 3, 4, 5, 6]
flatten_list = utils.flatten_list(nested_list)
self.assertEqual(len(flatten_list), 6)
self.assertListEqual(flatten_list, correct_list)
def infer(self, patterns, merge_output=False):
"""Run infererence in each node in the level with the corresponding
pattern from the previous level (input image).
Args:
patterns: a list of patterns from the level below or from the sensor
merge_output: whether to return concatenated inference vector or a
list of inference vectors for each pattern
Returns:
an inference vector of a list of inference vectors
"""
output = [] # nodes' outputs in the same order as in self.nodes
for i, n in enumerate(self.nodes):
node_patterns = itemgetter(*self.links[i])(patterns)
node_patterns = np.asarray(utils.flatten_list(node_patterns))
output.append(n.infer(node_patterns))
if merge_output:
# from [[output1], [output2]] makes [output1, output2]
return np.asarray(utils.flatten_list(output))
else:
return np.asarray(output)
def infer(self, patterns, merge_output=False):
"""Run infererence in each node in the level with the corresponding
pattern from the previous level (input image).
Args:
patterns: a list of patterns from the level below or from the sensor
merge_output: whether to return concatenated inference vector or a
list of inference vectors for each pattern
Returns:
an inference vector of a list of inference vectors
"""
output = [] # nodes' outputs in the same order as in self.nodes
for i, n in enumerate(self.nodes):
node_patterns = itemgetter(*self.links[i])(patterns)
node_patterns = np.asarray(utils.flatten_list(node_patterns))
output.append(n.infer(node_patterns))
if merge_output:
# from [[output1], [output2]] makes [output1, output2]
return np.asarray(utils.flatten_list(output))
else:
return np.asarray(output)
