def create_gng(max_nodes, n_inputs, step=0.2, n_start_nodes=2, max_edge_age=50): return algorithms.GrowingNeuralGas( n_inputs=n_inputs, n_start_nodes=n_start_nodes, shuffle_data=True, verbose=False, step=step, neighbour_step=0.005, max_edge_age=max_edge_age, max_nodes=max_nodes, n_iter_before_neuron_added=100, after_split_error_decay_rate=0.5, error_decay_rate=0.995, min_distance_for_update=0.01, )
# Run GNG
from neupy import algorithms, utils
utils.reproducible()
#gng with 3 features, because the data is 3-D now
gng = algorithms.GrowingNeuralGas(
n_inputs=4,
n_start_nodes=2,
shuffle_data=True,
verbose=False,
step=0.1,
neighbour_step=0.001,
max_edge_age=50,
# Before this was 1000
max_nodes=10000,
# To get 5000 notes, this was 2
# To get 554 nodes, this was 20
# To get 112 nodes, this was 100
n_iter_before_neuron_added=20,
after_split_error_decay_rate=0.5,
error_decay_rate=0.995,
min_distance_for_update=0.2,
)
def run_gng(data, i):
# Training will slow down overtime and we increase number
# of data samples for training
#n = int(0.5 * gng.n_iter_before_neuron_added * (1 + i // 100))
sample_vectors = np.array([p.sv for p in patches])
sv_len = len(sample_vectors[0])
utils.reproducible()
gng = algorithms.GrowingNeuralGas(
n_inputs=sv_len,
n_start_nodes=2,
shuffle_data=True,
verbose=False,
step=0.1,
neighbour_step=0.001,
max_edge_age=70,
max_nodes=250,
n_iter_before_neuron_added=200,
after_split_error_decay_rate = 0.5,
error_decay_rate=0.995,
min_distance_for_update=0.2,
)
gng.train(sample_vectors, epochs=40)
nodes = []
for node in gng.graph.nodes:
nodes.append(node.weight[0])