def train_unsupervised_hierarchy(X1, X2, n_epochs=30):
g1 = gwr_h_unimodal(n_layers=2, window_size=[3, 3], gwr_pars=pars)
g2 = gwr_h_unimodal(n_layers=2, window_size=[3, 3], gwr_pars=pars)
g3 = gwr(**pars)
X1_ = g1.train(X1, n_epochs=n_epochs)
X2_ = g2.train(X2, n_epochs=n_epochs)
X = np.hstack((X1_, X2_))
g3.train(X, n_epochs=n_epochs)
return g1, g2, g3
def __init__(self, n_layers = 2, window_size = [3, 1], lab_ratio =1 / 2,
gwr_pars = []):
'''
NOTE: The data X is used in training without window for layer 1,
then window_size[0] is used to get the trajectory from the first
to the second layer. If you have only two layers, window_size[1]
is the output trajectory already windowed.
'''
self.n_layers = n_layers
self.window_size = window_size
self.lab_ratio = lab_ratio
if isinstance(gwr_pars, list):
self.H = []
for i in range(self.n_layers):
try:
par = gwr_pars[i]
self.H.append(gwr(**par))
except IndexError:
self.H.append(gwr())
elif isinstance(gwr_pars, dict):
self.H = [gwr(**gwr_pars) for _ in range(self.n_layers)]
import numpy as np
import networkx as nx
import matplotlib.pyplot as plt
from neuralgas.oss_gwr import gwr
gwr = gwr(tau_b=0.3, tau_n=0.1, kappa=1.05)
gwr.G = nx.Graph()
gwr.G.add_node(0, attr_dict={'pos': np.array([1, 1]), 'fir': 1})
gwr.G.add_node(1, attr_dict={'pos': np.array([1, 1]), 'fir': 1})
gwr.G.add_edge(0, 1)
fir, fir_n = [1], [1]
for i in range(10):
gwr._update_firing(0)
fir.append(gwr.G.node[0]['fir'])
fir_n.append(gwr.G.node[1]['fir'])
f, ax = plt.subplots(1, 1)
ax.set_title('Synapse habituation')
ax.plot(fir, label='Best matching synapse')
ax.plot(fir_n, label='Neighbor synapse')
ax.legend()
import numpy as np import sklearn.datasets import matplotlib.pyplot as plt from neuralgas.oss_gwr import gwr ''' Growing two 2D neural gases to learn the distribution of Gaussian data with different standard deviations. The visualization shows only the positions of the neurons (and not the connections). ''' iris = sklearn.datasets.load_iris() X1 = np.random.normal(loc=2, scale=1.0, size=[1000, 2]) X2 = np.random.normal(loc=2, scale=2, size=[1000, 2]) g1 = gwr(act_thr=0.75, random_state=None, max_size=30) g1.train(X1, n_epochs=3) Xg1 = g1.get_positions() g2 = gwr(act_thr=0.75, random_state=None, max_size=30) g2.train(X2, n_epochs=3) Xg2 = g2.get_positions() marks = 8 f, ax = plt.subplots(2, 2) ax[0, 0].scatter(X1[:, 0], X1[:, 1], s=marks) ax[1, 0].scatter(X2[:, 0], X2[:, 1], s=marks) ymin, ymax = ax[1, 0].get_ylim() xmin, xmax = ax[1, 0].get_xlim() ax[0, 1].scatter(Xg1[:, 0], Xg1[:, 1], s=marks)
from __future__ import division import sklearn.datasets import matplotlib.pyplot as plt from neuralgas.oss_gwr import gwr iris = sklearn.datasets.load_iris() X = iris.data y = iris.target g1 = gwr(act_thr=0.75, random_state=12345) g1.train(X[:, :2], n_epochs=1) Xg1 = g1.get_positions() g1.train(X[:, :2], n_epochs=19, warm_start=True) Xg1_ = g1.get_positions() g2 = gwr(act_thr=0.75, random_state=1234) g2.train(X[:, 2:], n_epochs=1) Xg2 = g2.get_positions() g2.train(X[:, 2:], n_epochs=19, warm_start=True) Xg2_ = g2.get_positions() marks = 15 f, ax = plt.subplots(2, 3) ax[0, 0].scatter(X[:, 0], X[:, 1], s=marks) ymin, ymax = ax[0, 0].get_ylim() xmin, xmax = ax[0, 0].get_xlim() ax[0, 1].scatter(Xg1[:, 0], Xg1[:, 1], s=marks) ax[0, 2].scatter(Xg1_[:, 0], Xg1_[:, 1], s=marks) for i in range(1, 3):
X = iris.data
X1 = X[:, :2]
X2 = X[:, 2:]
y = iris.target
g1 = gwr_h_unimodal(n_layers=2,
window_size=[1, 2],
gwr_pars=[{
'act_thr': 0.7
}, {
'act_thr': 0.7
}])
g2 = gwr_h_unimodal(n_layers=2,
window_size=[1, 2],
gwr_pars=[{
'act_thr': 0.7
}, {
'act_thr': 0.7
}])
g3 = gwr()
XX1 = g1.train(X1)
XX2 = g2.train(X2)
X3 = np.hstack((XX1, XX2))
g3.train(X3)
Xf = _propagate_trajectories(X3, network=g3, ws=2)
# then you can take a oss_gwr