def test_logistic_derivative(transformer_factory):
# bprop is on the output
inputs = np.array([0, 1, -2], dtype=np.float).reshape((3, 1))
f = 1.0 / (1.0 + np.exp(-inputs))
outputs = f * (1.0 - f)
compare_tensors(Logistic(),
inputs, outputs, deriv=True, tol=1e-7)
class LogisticPair(ActivationPair):
neon_activation = Logistic()
tolerance = 1e-7
def reference_value(self, x):
return 1.0 / (1.0 + np.exp(-x))
def reference_derivative(self, x):
f = self.reference_value(x)
return f * (1.0 - f)
def make_network(scope1=None, scope2=None):
# 2 layer network, each layer has its own scope
x = ng.placeholder(axes) # inputs
t = ng.placeholder(ng.make_axes([ng.make_axis(length=1),
N])) # targets
with Layer.variable_scope(scope1):
layer1 = Affine(ConstantInit(val=Wlin1),
nout=nout1,
bias_init=ConstantInit(val=Wbias1),
activation=Rectlin(),
batch_norm=False)
with Layer.variable_scope(scope2):
layer2 = Affine(ConstantInit(val=Wlin2),
nout=1,
bias_init=ConstantInit(val=Wbias2),
activation=Logistic(),
batch_norm=False)
seq = Sequential([layer1, layer2])
p_t = seq(x)
t_cast = ng.cast_axes(t, p_t.axes) # TODO: how can this be avoided?
loss = ng.cross_entropy_binary(p_t, t_cast)
return seq, x, t, loss
def test_logistic(transformer_factory):
inputs = np.array([0, 1, -2]).reshape((3, 1))
outputs = 1.0 / (1.0 + np.exp(-inputs)).reshape((3, 1))
compare_tensors(Logistic(), inputs, outputs, tol=1e-7)