def test_single_state_func(self):
"""
Just to see if the functions manage to run at all
"""
y = self.solution[1::, 2]
# test out whether the single state function 'ok'
sir_obj = SquareLoss(self.theta, self.ode, self.x0, self.t[0],
self.t[1::], y, 'R')
sir_obj.cost()
sir_obj.gradient()
sir_obj.hessian()
def test_single_state_func(self):
"""
Just to see if the functions manage to run at all
"""
y = self.solution[1::, 2]
# test out whether the single state function 'ok'
sir_obj = SquareLoss(self.theta, self.ode, self.x0, self.t[0],
self.t[1::], y, 'R')
sir_obj.cost()
sir_obj.gradient()
sir_obj.hessian()
def test_FH_Square_vector_weight(self):
# now the weight is a vector
w = np.random.rand(29, 2)
obj = SquareLoss(self.theta, self.ode, self.x0, self.t[0], self.t[1::],
self.solution[1::, :], ['V', 'R'], w)
s = ((self.r * np.array(w))**2).sum()
self.assertTrue(np.allclose(obj.cost(), s))
def test_FH_Square(self):
# initial values
x0 = [-1.0, 1.0]
# params
paramEval = [('a', 0.2), ('b', 0.2), ('c', 3.0)]
# the time points for our observations
t = numpy.linspace(0, 20, 30).astype('float64')
ode = common_models.FitzHugh().setParameters(
paramEval).setInitialValue(x0, t[0])
# Standard. Find the solution which we will be used as "observations later"
solution, output = ode.integrate(t[1::], full_output=True)
# initial guess
theta = [0.5, 0.5, 0.5]
#objFH = squareLoss(theta,ode,x0,t0,t,solution[1::,1],'R')
objFH = SquareLoss(theta, ode, x0, t[0], t[1::], solution[1::, :],
['V', 'R'])
r = objFH.residual()
# weight for each component
w = [2.0, 3.0]
s1 = 0
for i in range(2):
s1 += ((r[:, i] * w[i])**2).sum()
objFH1 = SquareLoss(theta, ode, x0, t[0], t[1::], solution[1::, :],
['V', 'R'], w)
# now the weight is a vector
w = numpy.random.rand(29, 2)
objFH2 = SquareLoss(theta, ode, x0, t[0], t[1::], solution[1::, :],
['V', 'R'], w)
s2 = ((r * numpy.array(w))**2).sum()
if abs(objFH1.cost() - s1) >= 1e-2:
raise Exception("Failed!")
if abs(objFH2.cost() - s2) >= 1e-2:
raise Exception("Failed!")
def test_FH_Square_scalar_weight(self):
# weight for each component
w = [2.0, 3.0]
s = 0
for i in range(2):
s += ((self.r[:, i] * w[i])**2).sum()
obj = SquareLoss(self.theta, self.ode, self.x0, self.t[0], self.t[1::],
self.solution[1::, :], ['V', 'R'], w)
self.assertTrue(np.allclose(obj.cost(), s))
def test_SIR_Estimate_SquareLoss(self):
# define the model and parameters
ode = common_models.SIR({'beta':0.5,'gamma':1.0/3.0})
# the initial state, normalized to zero one
x0 = [1, 1.27e-6, 0]
# set the time sequence that we would like to observe
t = numpy.linspace(0, 150, 100)
# Standard. Find the solution.
solution = scipy.integrate.odeint(ode.ode, x0, t)
# y = copy.copy(solution[:,1:3])
# initial value
theta = [0.2, 0.2]
# test out whether the single state function 'ok'
objSIR = SquareLoss(theta, ode, x0, t[0], t[1::],
solution[1::,2], 'R')
objSIR.cost()
objSIR.gradient()
objSIR.hessian()
# now we go on the real shit
objSIR = SquareLoss(theta, ode, x0, t[0], t[1::],
solution[1::,1:3], ['I','R'])
# constraints
EPSILON = numpy.sqrt(numpy.finfo(numpy.float).eps)
boxBounds = [(EPSILON, 5), (EPSILON, 5)]
resQP = scipy.optimize.minimize(fun=objSIR.cost,
jac=objSIR.sensitivity,
x0=theta,
method='SLSQP',
bounds=boxBounds)
target = numpy.array([0.5, 1.0/3.0])
if numpy.any(abs(resQP['x']-target) >= 1e-2):
raise Exception("Failed!")
def test_SIR_Estimate_SquareLoss(self):
# define the model and parameters
ode = common_models.SIR({'beta': 0.5, 'gamma': 1.0 / 3.0})
# the initial state, normalized to zero one
x0 = [1, 1.27e-6, 0]
# set the time sequence that we would like to observe
t = numpy.linspace(0, 150, 100)
# Standard. Find the solution.
solution = scipy.integrate.odeint(ode.ode, x0, t)
# y = copy.copy(solution[:,1:3])
# initial value
theta = [0.2, 0.2]
# test out whether the single state function 'ok'
objSIR = SquareLoss(theta, ode, x0, t[0], t[1::], solution[1::, 2],
'R')
objSIR.cost()
objSIR.gradient()
objSIR.hessian()
# now we go on the real shit
objSIR = SquareLoss(theta, ode, x0, t[0], t[1::], solution[1::, 1:3],
['I', 'R'])
# constraints
EPSILON = numpy.sqrt(numpy.finfo(numpy.float).eps)
boxBounds = [(EPSILON, 5), (EPSILON, 5)]
resQP = scipy.optimize.minimize(fun=objSIR.cost,
jac=objSIR.sensitivity,
x0=theta,
method='SLSQP',
bounds=boxBounds)
target = numpy.array([0.5, 1.0 / 3.0])
if numpy.any(abs(resQP['x'] - target) >= 1e-2):
raise Exception("Failed!")
def test_FH_Square(self):
# initial values
x0 = [-1.0, 1.0]
# params
paramEval = [('a', 0.2), ('b', 0.2),('c', 3.0)]
# the time points for our observations
t = numpy.linspace(0, 20, 30).astype('float64')
ode = common_models.FitzHugh().setParameters(paramEval).setInitialValue(x0, t[0])
# Standard. Find the solution which we will be used as "observations later"
solution, output = ode.integrate(t[1::], full_output=True)
# initial guess
theta = [0.5, 0.5, 0.5]
#objFH = squareLoss(theta,ode,x0,t0,t,solution[1::,1],'R')
objFH = SquareLoss(theta, ode, x0, t[0], t[1::], solution[1::,:], ['V','R'])
r = objFH.residual()
# weight for each component
w = [2.0, 3.0]
s1 = 0
for i in range(2): s1 += ((r[:,i]*w[i])**2).sum()
objFH1 = SquareLoss(theta, ode, x0, t[0], t[1::], solution[1::,:],
['V','R'], w)
# now the weight is a vector
w = numpy.random.rand(29, 2)
objFH2 = SquareLoss(theta, ode, x0, t[0], t[1::], solution[1::,:],
['V','R'], w)
s2 = ((r * numpy.array(w))**2).sum()
if abs(objFH1.cost() - s1) >= 1e-2:
raise Exception("Failed!")
if abs(objFH2.cost() - s2) >= 1e-2:
raise Exception("Failed!")
disp=3, full_output=True)
xhat, output = ip(objFH.cost,
objFH.gradient,
x0=theta,
lb=lb, ub=ub,
G=None, h=None,
A=None, b=None,
maxiter=200,
method='bar',
disp=3, full_output=True)
x = numpy.array(theta)
oldFx = objFH.cost(x)
g = objFH.gradient(x)
deltaX = 0.5 * numpy.linalg.solve(objFH.jtj(x),-g)
objFH.cost(x + deltaX)
step, fc, gc, fx, old_fval, new_slope = scipy.optimize.line_search(objFH.cost,
objFH.gradient,
numpy.array(theta),
deltaX,
g,
oldFx
)
numpy.array(x) + step * deltaX
from pygotools.optutils import lineSearch
lineFunc = lineSearch(1, x, deltaX, objFH.cost)
