def test_forwardFork(self):
g = gt.Graph()
g.add_vertex(3)
g.add_edge(g.vertex(0), g.vertex(1))
g.add_edge(g.vertex(0), g.vertex(2))
n = Net(1, 2, g)
# Init activation functions
for v in n.g.vertices():
n.activation[v] = Identity()
# All weights and bias == 0
assert_allclose(n.forward([1.0]), [0.0, 0.0])
# Init weights
for e in n.g.edges():
n.weightProp[e] = 1.0
assert_allclose(n.forward([0.0]), [0.0, 0.0])
assert_allclose(n.forward([1.0]), [1.0, 1.0])
assert_allclose(n.forward([12.0]), [12.0, 12.0])
assert_allclose(n.forward([-12.0]), [-12.0, -12.0])
# Init bias
for v in n.g.vertices():
n.biasProp[v] = 1.0
assert_allclose(n.forward([0.0]), [1.0, 1.0])
assert_allclose(n.forward([1.0]), [2.0, 2.0])
def mlp(sizes, weightGenerator=np.random.random, biasGenerator=np.random.random, activationFunction=LogSigmoid(-1, 1)):
r"""Generate Multilayer Perceptron (MLP) and return as a Net
:param sizes: Size of each layer
:type sizes: list of integers
:param functor weightGenerator: Functor for generating weights
:param functor biasGenerator: Functor for generating bias values
:param activationFunction: Activation function - the same for all neurons
:type activationFunction: Activation function
:returns: Net object with the requested architecture
:rtype: :class:`~gtnn.network.Net`
"""
n = Net(sizes[0], sizes[-1])
layerId = n.addVertexProperty("layerId", "short")
lastLayer = []
presentLayer = []
# Create all layers
for layerIdx, size in enumerate(sizes):
layerProp = n.addLayer() # TODO: test layer
for i in range(size):
v = n.g.add_vertex()
layerProp[v] = True
layerId[v] = layerIdx
n.activation[v] = activationFunction
n.biasProp[v] = biasGenerator()
presentLayer.append(v)
for l in lastLayer:
e = n.g.add_edge(l, v)
n.weightProp[e] = weightGenerator()
lastLayer = list(presentLayer)
presentLayer = list()
# Create a subgraph
n.prepare()
return n
def test_backwardFork(self):
g = gt.Graph()
g.add_vertex(3)
g.add_edge(g.vertex(0), g.vertex(1))
g.add_edge(g.vertex(0), g.vertex(2))
n = Net(1, 2, g)
# Init activation functions
for v in n.g.vertices():
n.activation[v] = Identity()
# All weights == 0
assert_allclose(n.errorProp.a, [0, 0, 0])
n.backward([1, 1])
assert_allclose(n.errorProp.a, [0, 1, 1])
n.backward([-100, -100])
assert_allclose(n.errorProp.a, [0, -100, -100])
# Init weights
for e in n.g.edges():
n.weightProp[e] = 1.0
# Tests
n.backward([1, 1])
assert_allclose(n.errorProp.a, [2, 1, 1])
n.backward([30, 30])
assert_allclose(n.errorProp.a, [60, 30, 30])
n.backward([0, 0])
assert_allclose(n.errorProp.a, [0, 0, 0])
n.backward([-10, -8])
assert_allclose(n.errorProp.a, [-18, -8, -10])
n.backward([3, 57])
assert_allclose(n.errorProp.a, [60, 57, 3])
def test_construc(self):
g = gt.Graph()
g.add_vertex(4)
g.add_edge(g.vertex(0), g.vertex(1))
g.add_edge(g.vertex(0), g.vertex(2))
g.add_edge(g.vertex(1), g.vertex(3))
g.add_edge(g.vertex(2), g.vertex(3))
g.add_vertex(1)
g.add_edge(g.vertex(4), g.vertex(0))
n = Net(1, 1, g)
for e in g.edges():
n.weightProp[e] = 1
for v in g.vertices():
n.activation[v] = LogSigmoid(-1, 1)
n.prepare()
n.backward([1])
n.forward([1])
n.backward([2])
