def test_deriv(self):
vars = np.arange(-100, 100, 2.5)
test_fcns = [
TanSig(),
PureLin(),
LogSig(),
HardLim(),
HardLims(),
SatLin(),
SatLins(),
SatLinPrm(1, 0, 1),
SatLinPrm(1, -1, 1)
]
def diff(f, x, h=1E-6):
x1 = np.array([x - h])
x2 = np.array([x])
r = (f(x2)[0] - f(x1)[0]) / h
return r
for test_fcn in test_fcns:
for var in vars:
m = diff(test_fcn, var)
v = test_fcn(np.array([var]))
t = test_fcn.deriv(np.array([var]), v)[0]
self.assertAlmostEqual(m, t, 5)
def test_logsig(self):
test_fcn = LogSig()
vars = [-2.5, -0.5, 0.0, 0.1, 3.0]
from math import exp
model = lambda x: 1 / (1 + exp(-x))
m_res = map(model, vars)
t_res = test_fcn(np.array(vars)).tolist()
for m, t in zip(m_res, t_res):
self.assertEqual(m, t)
def test_props(self):
test_fcns = [
TanSig(),
PureLin(),
LogSig(),
HardLim(),
HardLims(),
Competitive(),
SatLin(),
SatLins()
]
vars = [-1e50, -2.5, -0.5, 0.0, 0.1, 3.0, 1e50]
for test_fcn in test_fcns:
self.assertEqual(test_fcn.out_minmax[1] >= test_fcn.out_minmax[0],
True)
self.assertEqual(test_fcn.inp_active[1] >= test_fcn.inp_active[0],
True)
for v in vars:
r = test_fcn(np.array([v]))
self.assertEqual(
test_fcn.out_minmax[0] <= r <= test_fcn.out_minmax[1],
True)