Fixed bug when training SOM with toroidal topology

neural_map/_neural_map.py

@@ -126,16 +126,16 @@ def __init__(self,
  _check_inputs.value_type(hexagonal, bool)
  _check_inputs.value_type(toroidal, bool)
  _check_inputs.value_type(seed, int)
- self._metric = metric
- self._kwargs = kwargs
+ self.metric = metric
+ self.kwargs = kwargs
 
  if isinstance(metric, str):
  def distance(first_array, second_array):
- return cdist(first_array, second_array, self._metric, **self._kwargs)
+ return cdist(first_array, second_array, self.metric, **self.kwargs)
 
  else:
  def distance(first_array, second_array):
- return self._metric(first_array, second_array, **self._kwargs)
+ return self.metric(first_array, second_array, **self.kwargs)
 
  self.distance = distance
  self.distance(array([[0., 1.]]), array([[1., 2.], [3., 4.]]))
@@ -157,16 +157,16 @@ def distance(first_array, second_array):
 
  self.weights = weights
  self.activation_map = zeros((self.columns, self.rows))
- self._adjacent_nodes_relative_positions = [[(1, 0), (0, 1), (-1, 0), (0, -1)],
-  [(1, 0), (0, 1), (-1, 0), (0, -1)]]
+ self.adjacent_nodes_relative_positions = [[(1, 0), (0, 1), (-1, 0), (0, -1)],
+ [(1, 0), (0, 1), (-1, 0), (0, -1)]]
  self.positions = transpose(meshgrid(arange(self.columns), arange(self.rows)),
  axes=[2, 1, 0]).astype(float)
 
  if self.hexagonal:
  self.positions[:, 0::2, 0] += 0.5
  self.positions[..., 1] *= (3 ** 0.5) * 0.5
  self.height *= (3 ** 0.5) * 0.5
- self._adjacent_nodes_relative_positions = [
+ self.adjacent_nodes_relative_positions = [
  [(1, 0), (1, 1), (0, 1), (-1, 0), (0, -1), (1, -1)],
  [(1, 0), (0, 1), (-1, 1), (-1, 0), (-1, -1), (0, -1)]
  ]
@@ -507,7 +507,7 @@ def train(self,
  # rotate update matrix to match its center with the bmu position
  update_matrix = update_matrix_over_center[tuple([
  (update_matrix_indices[0] - (offset[0] + offset_correction)) % self.columns,
- (update_matrix_indices[1] - offset[1][:, None]) % self.rows
+ (update_matrix_indices[1] - offset[1]) % self.rows # hola
  ])]
 
  # computes displacement for all the relative positions of neighbourhood nodes
@@ -655,14 +655,14 @@ def get_unified_distance_matrix(self):
 
  adjacency_matrix = zeros((self.columns * self.rows, self.columns * self.rows)) * nan
  fill_diagonal(adjacency_matrix, 0.)
- adjacency_count = len(self._adjacent_nodes_relative_positions[0])
+ adjacency_count = len(self.adjacent_nodes_relative_positions[0])
  unified_distance_matrix = zeros((self.columns, self.rows, 1 + adjacency_count))
 
  for x_index in range(self.columns):
  for y_index in range(self.rows):
  adjacent_nodes = 0
 
- for k, (i, j) in enumerate(self._adjacent_nodes_relative_positions[y_index % 2]):
+ for k, (i, j) in enumerate(self.adjacent_nodes_relative_positions[y_index % 2]):
  if self.toroidal:
  neighbour = array([[self.weights[
  (x_index + i + self.columns) % self.columns,
@@ -824,8 +824,8 @@ def k_medoids(self, n_clusters=4):
  _check_inputs.value_type(n_clusters, int)
  _check_inputs.positive(n_clusters)
 
- if isinstance(self._metric, str):
- metric = self._metric
+ if isinstance(self.metric, str):
+ metric = self.metric
 
  else:
  def custom_metric(first_set, second_set):
@@ -930,7 +930,7 @@ def evaluate(self, data):
  s_bmu = unravel_index(argmin(activation_map), activation_map.shape)
  error = 1
 
- for (i, j) in self._adjacent_nodes_relative_positions[f_bmu[1] % 2]:
+ for (i, j) in self.adjacent_nodes_relative_positions[f_bmu[1] % 2]:
  if (self.toroidal and s_bmu[0] == (f_bmu[0] + i + self.columns) % self.columns
  and s_bmu[1] == (f_bmu[1] + j + self.rows) % self.rows):
  error = 0
@@ -957,7 +957,7 @@ def get_dict(self):
  return {
  **{
  'variables': self.variables,
- 'metric': self._metric,
+ 'metric': self.metric,
  'columns': self.columns,
  'rows': self.rows,
  'hexagonal': self.hexagonal,
@@ -966,7 +966,7 @@ def get_dict(self):
  'weights': self.weights,
  'relative_positions': self.relative_positions
  },
- **self._kwargs
+ **self.kwargs
  }
 
  def get_connections_and_reverse(self, min_cluster_size=3):

setup.py

@@ -5,7 +5,7 @@
 
 setuptools.setup(
  name='neural_map',
- version="0.0.3",
+ version="0.0.4",
  author="Franco José Bobadilla",
  author_email='1709673@ucc.edu.ar',
  description='NeuralMap is a data analysis tool based on Self-Organizing Maps',