def addLayers(self, neurons, activation_fun):
self.epoch = 0
log.logNN.info("neurons= " + str(neurons))
self.nHidden = len(neurons)
self.layers = []
self.v = []
self.act_fun = []
for i in range(self.nHidden + 1):
if activation_fun[i] == 'relu':
self.act_fun.append(lambda x, der: af.ReLU(x, der))
elif activation_fun[i] == 'sigmoid':
self.act_fun.append(lambda x, der: af.sigmoid(x, der))
elif activation_fun[i] == 'linear':
self.act_fun.append(lambda x, der: af.linear(x, der))
elif activation_fun[i] == 'tanh':
self.act_fun.append(lambda x, der: af.tanh(x, der))
elif activation_fun[i] == 'leakyrelu':
self.act_fun.append(lambda x, der: af.LReLU(x, der))
for i in range(self.nHidden):
n = neurons[i]
Wh = np.random.randn(N_FEATURES if i == 0 else neurons[i - 1],
n) * math.sqrt(2.0 / self.numEx)
bWh = np.random.randn(1, n) * math.sqrt(2.0 / self.numEx)
self.layers.append([Wh, bWh])
self.v.append([0, 0])
Wo = np.random.randn(neurons[-1], N_CLASSES) * math.sqrt(
2.0 / self.numEx)
bWo = np.random.randn(1, N_CLASSES) * math.sqrt(2.0 / self.numEx)
self.layers.append([Wo, bWo])
self.v.append([0, 0])
def addLayers(self, neurons, activation_fun, weights=None):
self.epoch = 0
log.logNN.info("neurons= " + str(neurons))
self.nHidden = len(neurons)
self.layers = []
self.v = [[0, 0] for _ in range(self.nHidden + 1)]
act_fun_factory = {
"relu": lambda x, der: af.ReLU(x, der),
"sigmoid": lambda x, der: af.sigmoid(x, der),
"linear": lambda x, der: af.linear(x, der),
"tanh": lambda x, der: af.tanh(x, der),
"leakyrelu": lambda x, der: af.LReLU(x, der)
}
self.act_fun = [act_fun_factory[f] for f in activation_fun]
if weights == None:
weights_hidden_shapes = list(
zip([N_FEATURES] + neurons[:-1], neurons))
weights_hidden = [
np.random.randn(row, col) * math.sqrt(2.0 / self.numEx)
for row, col in weights_hidden_shapes
]
#bias_hidden = [np.random.randn(1, n) * math.sqrt(2.0 / self.numEx) for n in neurons]
#weights_hidden = [np.random.normal(scale=0.05, size=(row, col)) for row, col in weights_hidden_shapes]
#bias_hidden = [np.random.normal(scale=0.05, size=(1, n)) for n in neurons]
bias_hidden = [np.ones((1, n)) * 0.0001 for n in neurons]
self.layers = [[w, b]
for w, b in list(zip(weights_hidden, bias_hidden))]
Wo = np.random.randn(neurons[-1], N_CLASSES) * math.sqrt(
2.0 / self.numEx)
#Wo = np.random.normal(scale=0.05,size=(neurons[-1], N_CLASSES))
# bWo = np.random.randn(1, N_CLASSES) * math.sqrt(2.0 / self.numEx)
bWo = np.ones((1, N_CLASSES)) * 0.0001
self.layers += [[Wo, bWo]]
else:
self.layers = weights
def test_sigmoid_min(self):
self.assertAlmostEqual(af.sigmoid(-1000), 0)
def test_sigmoid_max(self):
self.assertAlmostEqual(af.sigmoid(1000), 1)
def test_sigmoid_zero(self):
self.assertEqual(af.sigmoid(0), 0.5)