def eqProbs(p, U):
A = len(p)
pC = adouble(np.zeros((A, )))
pF = adouble(np.zeros((A, )))
for i in xrange(A):
c = cStar(p[i], U, i)
g1 = g_jo(c, U, i)
pC[i] = g1
pF[i] = h_jo(c, U, i) / g1
return np.hstack((p, pC, pF)).reshape((-1, 3), order='F')
def tape(fID, gID, lID, x0, beta):
t0 = time.time()
# trace objective function
adolc.trace_on(fID)
ax = adolc.adouble(x0)
adolc.independent(ax)
ay = action(ax, beta)
#ay = actTDtest(ax, beta)
adolc.dependent(ay)
adolc.trace_off()
# trace constraint function
adolc.trace_on(gID)
ax = adolc.adouble(x0)
adolc.independent(ax)
ay = eval_g(ax)
adolc.dependent(ay)
adolc.trace_off()
# trace lagrangian function
adolc.trace_on(lID)
# xtmp = [x0, lambdas, obj_factor]
# xtmp = np.hstack(x0,np.ones(ncon),1.)
ax = adolc.adouble(x0)
alagrange = adolc.adouble(np.ones(ncon))
aobj_factor = adolc.adouble(1.)
adolc.independent(ax)
adolc.independent(alagrange)
adolc.independent(aobj_factor)
ay = eval_lagrangian(ax, alagrange, aobj_factor, beta)
adolc.dependent(ay)
adolc.trace_off()
t1 = time.time()
print "tape time = ", t1-t0
eval_jac_g_adolc = Eval_jac_g(x0)
#eval_h_adolc = Eval_h(x0, np.ones(ncon), 1.)
eval_h_adolc = Eval_h_dense(x0, np.ones(ncon), 1.)
t2 = time.time()
print "Hess time = ", t2-t1
#
return eval_jac_g_adolc, eval_h_adolc
def adjoint_solve(self, data, weight_sens=False):
# Evaluate the jacobian of residuals w.r.t. states and augmentation field
ad.trace_on(1)
ad_states = ad.adouble(self.states)
ad_beta = ad.adouble(self.beta)
ad.independent(ad_states)
ad.independent(ad_beta)
ad_res = self.evalResidual(ad_states, ad_beta)
ad.dependent(ad_res)
ad.trace_off()
jacres = ad.jacobian(1, np.hstack((self.states, self.beta)))
Rq = jacres[:,0:np.shape(self.states)[0]]
Rb = jacres[:,np.shape(self.states)[0]:]
# Obtain the jacobian of objective function w.r.t. states and augmentation field
Jq, Jb = self.getObjJac(data)
# Solve the discrete adjoint system to obtain sensitivity
psi = np.linalg.solve(Rq.T,Jq)
sens = Jb - np.matmul(Rb.T,psi)
# Obtain the sensitivity of the objective function w.r.t. NN weights
if weight_sens==True:
d_weights = nn.nn.nn_get_weights_sens(np.asfortranarray(self.nn_params["network"]),
self.nn_params["act_fn"],
self.nn_params["loss_fn"],
self.nn_params["opt"],
np.asfortranarray(self.nn_params["weights"]),
np.asfortranarray(self.features),
1,
np.shape(self.beta)[0],
np.asfortranarray(sens),
np.asfortranarray(self.nn_params["opt_params_array"]))
return d_weights
else:
return sens
def pCostpW_adolc(self, W, p_target):
tag = self.pCostpW_tag
if not self.pCostpW_traced:
aW = adouble(W.flatten(order='F'))
ap = adouble(p_target)
adolc.trace_on(tag)
adolc.independent(aW)
aW3 = np.reshape(aW, W.shape, order='F')
acost = cost.inverse_pressure_design(aW3, ap)
adolc.dependent(acost)
adolc.trace_off()
return adolc.gradient(self.pCostpW_tag, W.flatten(order='F'))
def __init__(self,func,x):
""" Initializes the tape for a function which allows
the user to then quickly evaluate the function,
gradient, and hessian at arbitrary points
func is assumed to be a function from a vector array
to a single number.
Also updates the global CUR_ADOLC_TAPE_NUMBER which
keeps track of the tapes that have been used by
ADOLC
Parameters
----------
func: function
function to be evaluated
x: ndarray
ndarray of correct shape and type for the
function
"""
# access the tape number counter
global CUR_ADOLC_TAPE_NUMBER
# initialize the ADOLC tape for the function
adolc.trace_on(CUR_ADOLC_TAPE_NUMBER)
x = adolc.adouble(x)
adolc.independent(x)
y = func(x)
adolc.dependent(y)
adolc.trace_off()
self.tape_number = CUR_ADOLC_TAPE_NUMBER
CUR_ADOLC_TAPE_NUMBER += 1
def condassign(b, c, d):
"""
Returns the result of (b>0)? c : d
Use instead of if statements to force taping both c and d
expressions.
"""
if type(b) == ad._adolc.adub:
ad_a = ad.adouble(0.)
ad.condassign(ad_a, ad.adouble(b), ad.adouble(c), ad.adouble(d))
return ad_a
else:
if b > 0.:
return c
else:
return d
def condassign(b, c, d):
"""
Returns the result of (b>0)? c : d
Use instead of if statements to force taping both c and d
expressions.
"""
if type(b) == ad._adolc.adub:
ad_a = ad.adouble(0.)
ad.condassign(ad_a, ad.adouble(b), ad.adouble(c), ad.adouble(d))
return ad_a
else:
if b > 0.:
return c
else:
return d
def calc_residual_jacobian(self, q, dq=1e-25):
n = np.size(q)
q_c = q.copy()
q = ad.adouble(q)
tag = 0
ad.trace_on(tag)
ad.independent(q)
R = self.calc_residual(q)
ad.dependent(R)
ad.trace_off()
options = np.array([0, 0, 0, 0], dtype=int)
q = q_c
dRdq = calc_jacobian(q, tag=tag, shape=(n, n))
#pat = ad.sparse.jac_pat(tag, q_c, options)
#if 1:
# result = ad.colpack.sparse_jac_no_repeat(tag, q, options)
#else:
# result = ad.colpack.sparse_jac_repeat(tag, q, nnz, ridx, cidx, values)
#nnz = result[0]
#ridx = result[1]
#cidx = result[2]
#values = result[3]
#dRdq = sp.csr_matrix((values, (ridx, cidx)), shape=(n, n))
#dRdq = np.zeros([n, n], dtype=q.dtype)
#for i in range(n):
# q[i] = q[i] + 1j*dq
# R = self.calc_residual(q)
# dRdq[:,i] = np.imag(R[:])/dq
# q[i] = q[i] - 1j*dq
return dRdq
def calc_residual(self, q, dtype=None):
R = np.zeros_like(q)
if dtype == ad.adouble:
R = ad.adouble(R)
n = self.n
R[0::3], R[1::3], R[2::3] = self.calc_momentum_residual(q), self.calc_k_residual(q), self.calc_omega_residual(q)
#R = komegaf.calc_residual(self.y.astype(np.float64), q.astype(np.float64), np.float64(self.dp), self.beta.astype(np.float64))
return R
def PyAdolc_dvLJ(x):
adolc.trace_on(0)
ad_x = adolc.adouble(np.zeros(np.shape(np.ravel(x)),dtype=float))
adolc.independent(ad_x)
ad_y = Adolc_vLJ(ad_x)
adolc.dependent(ad_y)
adolc.trace_off()
return adolc.gradient(0, np.ravel(x))
def _trace_con(self, x):
if self._con_trace_id is None and self.m > 0:
self._con_trace_id = self._get_trace_id() + 1
adolc.trace_on(self._con_trace_id)
x = adolc.adouble(x)
adolc.independent(x)
y = self.cons(x)
adolc.dependent(y)
adolc.trace_off()
def _trace_obj(self, x):
if self._obj_trace_id is None:
self._obj_trace_id = self._get_trace_id()
adolc.trace_on(self._obj_trace_id)
x = adolc.adouble(x)
adolc.independent(x)
y = self.obj(x)
adolc.dependent(y)
adolc.trace_off()
def pRpW_adolc(self, sim, W, area):
tag = self.pRpW_tag
if not self.pRpW_traced:
aW = adouble(W.flatten(order='F'))
aarea = adouble(area)
adolc.trace_on(tag)
adolc.independent(aW)
aW3 = np.reshape(aW, W.shape, order='F')
aresidual = q1d.evaluate_residual(sim, aW3, area)
aresidual.flatten(order='F')
adolc.dependent(aresidual)
adolc.trace_off()
return adolc.jacobian(self.pRpW_tag, W.flatten(order='F'))
def ppRpWpW_adolc(self, sim, W, area):
tag = self.ppRpWpW_tag
if not self.ppRpWpW_traced:
aW = adouble(W.flatten(order='F'))
aarea = adouble(area)
adolc.trace_on(tag)
adolc.independent(aW)
aW3 = np.reshape(aW, W.shape, order='F')
pRpW = self.pRpW_adolc(sim, aW3, area)
pRpW.flatten(order='F')
adolc.dependent(pRpW)
adolc.trace_off()
return adolc.hessian(self.ppRpWpW_tag, W.flatten(order='F'))
def const(p, U):
A = len(p)
psi = adouble(np.zeros((A, )))
for i in xrange(A):
c = cStar(p[i], U, i)
g1 = g_jo(c, U, i)
j = h_jo(c, U, i) / g1
psi[i] = (p[i] - f_jo(j, U, i))
return psi
def _trace_lag(self, x, z):
self._trace_obj(x)
self._trace_con(x)
self._trace_cons_pos(x)
unconstrained = self.m == 0 and self.nbounds == 0
if self._lag_trace_id is None:
if unconstrained:
self._lag_trace_id = self._obj_trace_id
return
self._lag_trace_id = self._get_trace_id() + 3
adolc.trace_on(self._lag_trace_id)
x = adolc.adouble(x)
z = adolc.adouble(z)
adolc.independent(x)
adolc.independent(z)
l = self.lag(x, z)
adolc.dependent(l)
adolc.trace_off()
def __init__(self, f, x, test = 'f'):
self.f = f
self.x = x
adolc.trace_on(0)
ax = adolc.adouble(x)
adolc.independent(ax)
y = f(ax)
adolc.dependent(y)
adolc.trace_off()
def __init__(self, f, x, test='f'):
self.f = f
self.x = x
adolc.trace_on(0)
ax = adolc.adouble(x)
adolc.independent(ax)
y = f(ax)
adolc.dependent(y)
adolc.trace_off()
def _create_topography_active_interpolate(self,
tag=0,
separate_faults=False):
import adolc
# Sanitise kwargs
assert isinstance(tag, int)
assert tag >= 0
for control in self.active_controls:
assert control in self.all_controls
# Initialise tape
adolc.trace_on(tag)
# Read parameters and mark active variables as independent
msg = "INIT: Subfault {:d}: shear modulus {:4.1e} Pa, seismic moment is {:4.1e}"
for i, subfault in enumerate(self.subfaults):
for control in self.all_controls:
if control in self.active_controls:
subfault.__setattr__(
control,
adolc.adouble(self.control_parameters[control][i]))
adolc.independent(subfault.__getattribute__(control))
else:
subfault.__setattr__(control,
self.control_parameters[control][i])
if self.debug and self.debug_mode == 'full':
try:
print(msg.format(i, subfault.mu, subfault.Mo().val))
except Exception:
print(msg.format(i, subfault.mu, subfault.Mo()))
# Create the topography, thereby calling Okada
self.print_debug("SETUP: Creating topography using Okada model...")
self.fault.create_dtopography(verbose=self.debug
and self.debug_mode == 'full',
active=True)
# Compute quantity of interest
self.J_subfaults = [0.0 for j in range(self.N)]
data = self._data_to_interpolate
for j in range(self.N):
for i in range(self.N):
self.J_subfaults[j] += (data[i, j] -
self.fault.dtopo.dZ_a[i, j])**2
self.J_subfaults[j] /= self.N**2
self.J = sum(self.J_subfaults)
# Mark dependence
if separate_faults:
for j in range(self.N):
adolc.dependent(self.J_subfaults[j])
else:
adolc.dependent(self.J)
adolc.trace_off()
def run(N,D,dt,beta,x0):
if x0=='start':
#init = 20.0*np.random.random_sample((N,D))-10.0
#np.savetxt('initpaths.txt',init)
init = np.loadtxt('initpaths.txt')
else:
init = x0.reshape(N,D)
#init = np.loadtxt("data_D{0}_dt{1}_noP.txt".format(D,dt))[:N,:]
if init.shape != (N,D):
raise "x is wrong dims!"
epsg = 1e-8
epsf = 1e-8
epsx = 1e-8
maxits = 10000
### Use ADOL-C to generate trace, used for derivatives and
### evaluations
start = time.time()
adolc.trace_on(adolcID)
ax = adolc.adouble(init.flatten())
adolc.independent(ax)
af = action(ax, beta)
adolc.dependent(af)
adolc.trace_off()
print "taped =", time.time()-start,"s"
#Test Gradient numerically
# flat = init.flatten()
# grad = np.zeros_like(flat)
# cost1 = alglib_func(flat,grad,1)
# grad2 = grad.copy()
# for i in range(len(flat)):
# perturb = flat.copy()
# perturb[i] = perturb[i]+0.00001
# cost2 =alglib_func(perturb,grad2,1)
# numgrad = (cost2-cost1)/0.00001
# print numgrad/grad[i]
m = af.shape[0]
state = xalglib.minlmcreatevj( m ,list(init.flatten()))
xalglib.minlmsetcond(state,epsg, epsf,epsx,maxits)
xalglib.minlmoptimize_vj(state, lm_func, lm_jac)
final, rep = xalglib.minlbfgsresults(state)
print "optimized: ", time.time()-start, "s"
print "Exit flag = ", rep.terminationtype, rep.iterationscount
print "Action = ", action(final,beta)
return rep.iterationscount, action(final,beta), final
def explicit_euler(x0,f,ts,p,q):
N = size(ts)
if isinstance(p[0],adolc._adolc.adouble):
x = array([adolc.adouble(0) for m in range(Nm)])
else:
x = zeros(N)
x[0] = x0
for n in range(1,N):
h = ts[n] - ts[n-1]
x[n]= x[n-1] + h*f(ts[n-1],x[n-1],p,q)
return x
def _trace_con(self, x):
if self._con_trace_id is None and self.m > 0:
self._con_trace_id = self._get_trace_id() + 1
adolc.trace_on(self._con_trace_id)
x = adolc.adouble(x)
adolc.independent(x)
y = self.cons(x)
adolc.dependent(y)
adolc.trace_off()
def explicit_euler(x0, f, ts, p, q):
N = size(ts)
if isinstance(p[0], adolc.adouble):
x = array([adolc.adouble(0) for m in range(Nm)])
else:
x = zeros(N)
x[0] = x0
for n in range(1, N):
h = ts[n] - ts[n - 1]
x[n] = x[n - 1] + h * f(ts[n - 1], x[n - 1], p, q)
return x
def _trace_obj(self, x):
if self._obj_trace_id is None:
self._obj_trace_id = self._get_trace_id()
adolc.trace_on(self._obj_trace_id)
x = adolc.adouble(x)
adolc.independent(x)
y = self.obj(x)
adolc.dependent(y)
adolc.trace_off()
def func(tag):
n=50
x=goffin_init(n)
tag=1
adolc.trace_on(tag)
ax = adolc.independent(adolc.adouble(x))
ay = goffin(ax,n)
adolc.dependent(ay)
trace_off()
y=ay.val
return y
def getObjJac(self, data):
ad.trace_on(1)
ad_states = ad.adouble(self.states)
ad_beta = ad.adouble(self.beta)
ad.independent(ad_states)
ad.independent(ad_beta)
ad_obj = self.getObjRaw(ad_states, data, ad_beta)
ad.dependent(ad_obj)
ad.trace_off()
jacobj = ad.jacobian(1, np.hstack((self.states, self.beta)))
Jq = jacobj[:,0:np.shape(self.states)[0]]
Jb = jacobj[:,np.shape(self.states)[0]:]
return Jq[0,:], Jb[0,:]
def explicit_euler(x0,f,ts,p,q): Nm = size(ts) Nx = size(x0) if isinstance(p[0],adolc.adouble): x = array([[adolc.adouble(0) for n in range(Nx)] for m in range(Nm)]) else: x = zeros((Nm,Nx)) x[0,:] = x0[:] for m in range(1,Nm): h = ts[m] - ts[m-1] x[m,:]= x[m-1,:] + h*f(ts[m-1],x[m-1,:],p,q) return x
def trace(self, dims):
# trace function
t = numpy.zeros(1)
x = numpy.zeros(dims['x'])
f = numpy.zeros(dims['x'])
p = numpy.zeros(dims['p'])
u = numpy.zeros(dims['u'])
adolc.trace_on(123)
at = adolc.adouble(t)
ax = adolc.adouble(x)
af = adolc.adouble(f)
ap = adolc.adouble(p)
au = adolc.adouble(u)
adolc.independent(at)
adolc.independent(ax)
adolc.independent(ap)
adolc.independent(au)
self.ffcn(t, ax, af, ap, au)
adolc.dependent(af)
adolc.trace_off()
self.traced = True
def calc_delJ_delbeta(self, q, beta):
n = np.size(beta)
beta_c = beta.copy()
beta = ad.adouble(beta)
tag = 1
ad.trace_on(tag)
ad.independent(beta)
F = self.objective.objective(q, beta)
#print ad.value(F)
ad.dependent(F)
ad.trace_off()
beta = beta_c
dJdbeta = calc_jacobian(beta, tag=tag, sparse=False)
return dJdbeta
def calc_delJ_delq(self, q, beta):
n = np.size(q)
q_c = q.copy()
q = ad.adouble(q)
tag = 2
ad.trace_on(tag)
ad.independent(q)
F = self.objective.objective(q, beta)
ad.dependent(F)
ad.trace_off()
q = q_c
dJdq = calc_jacobian(q, tag=tag, sparse=False)
return dJdq
def calc_delR_delbeta(self, q):
nb = np.size(self.beta)
n = np.size(q)
beta_c = self.beta.copy()
self.beta = ad.adouble(self.beta)
tag = 3
ad.trace_on(tag)
ad.independent(self.beta)
R = self.calc_residual(q, dtype=ad.adouble)
ad.dependent(R)
ad.trace_off()
self.beta = beta_c
dRdbeta = calc_jacobian(self.beta, tag=tag, sparse=False)
return dRdbeta
def PyAdolc_vLJ_Optimize(x):
N = len(x)
adolc.trace_on(0)
ad_x = adolc.adouble(np.zeros(np.shape(np.ravel(x)),dtype=float))
adolc.independent(ad_x)
ad_y = Adolc_vLJ(ad_x)
adolc.dependent(ad_y)
adolc.trace_off()
Adolc_BFGSres = optimize.minimize(PyAdolc_vLJ_for_Optimize, np.ravel(x), \
method='L-BFGS-B', \
jac = PyAdolc_dvLJ_for_Optimize, \
options={'disp': False})
return np.reshape(Adolc_BFGSres.x, (N,D))
def tape_A(self, xtrace):
"""
Tape the objective function.
"""
print('Taping action evaluation...')
tstart = time.time()
"""
trace objective function
"""
adolc.trace_on(self.adolcID)
# set the active independent variables
ax = adolc.adouble(xtrace)
adolc.independent(ax)
# set the dependent variable (or vector of dependent variables)
af = self.A(ax)
adolc.dependent(af)
adolc.trace_off()
#IPOPT needs the lagrangian functions to be traced
if self.method == 'IPOPT':
"""
trace lagrangian unconstrained
"""
#This adolc numbering could cause problems in the future
self.lagrangeID = self.adolcID + 1000
adolc.trace_on(self.lagrangeID)
ax = adolc.adouble(xtrace)
aobj_factor = adolc.adouble(1.)
adolc.independent(ax)
adolc.independent(aobj_factor)
ay = self.eval_lagrangian(ax, aobj_factor)
adolc.trace_off()
self.taped = True
print('Done!')
print('Time = {0} s\n'.format(time.time() - tstart))
def __init__(self, f, x, scaler=False):
if not isinstance(x, list):
x_lst = [x]
else:
x_lst = x
self.id = next(self.__class__._ids)
self.scaler = scaler
# x = np.random.uniform(x_L, x_U)
adolc.trace_on(self.id)
a_x_lst = [adolc.adouble(v) for v in x_lst]
for ax in a_x_lst:
adolc.independent(ax)
ay = f(*a_x_lst)
adolc.dependent(ay)
adolc.trace_off()
def qr(in_A):
"""
QR decomposition of A
Q,R = qr(A)
"""
# input checks
Ndim = numpy.ndim(in_A)
assert Ndim == 2
N,M = numpy.shape(in_A)
assert N==M
# prepare R and QT
R = in_A.copy()
if isinstance(in_A[0,0], adolc._adolc.adouble):
QT = numpy.array([[adolc.adouble(0) for c in range(N)] for r in range(N) ])
else:
QT = numpy.zeros((N,N))
for n in range(N):
QT[n,n] += 1
# main algorithm
for n in range(N):
for m in range(n+1,N):
a = R[n,n]
b = R[m,n]
r = numpy.sqrt(a**2 + b**2)
c = a/r
s = b/r
for k in range(N):
Rnk = R[n,k]
R[n,k] = c*Rnk + s*R[m,k]
R[m,k] =-s*Rnk + c*R[m,k];
QTnk = QT[n,k]
QT[n,k] = c*QTnk + s*QT[m,k]
QT[m,k] =-s*QTnk + c*QT[m,k];
#print 'QT:\n',QT
#print 'R:\n',R
#print '-------------'
return QT.T,R
def evalJacobian(self, states, beta_inv):
# Evaluate jacobian of residuals w.r.t. states
ad.trace_on(1)
ad_states = ad.adouble(states)
ad.independent(ad_states)
ad_res = self.evalResidual(ad_states, beta_inv)
ad.dependent(ad_res)
ad.trace_off()
return ad.jacobian(1, states)
def _create_dtopography_active(self, verbose=False):
import adolc
num_subfaults = len(self.subfaults)
tic = perf_counter()
msg = "created topography for subfault {:d}/{:d} ({:.1f} seconds)"
dz = np.zeros(self.dtopo.X.shape)
dz = adolc.adouble(dz)
for k, subfault in enumerate(self.subfaults):
subfault.okada()
dz += subfault.dtopo.dZ[0, :, :].reshape(dz.shape)
if k % 10 == 0 and verbose:
print(msg.format(k+1, num_subfaults, perf_counter() - tic))
tic = perf_counter()
self.dtopo.dZ_a = dz
self.dtopo.dZ = np.array([dzi.val for dzi in np.ravel(dz)]).reshape((1, ) + dz.shape)
return self.dtopo
def __init__(self):
dwl.OptimizationModel.__init__(self)
self.setDimensionOfState(4)
self.setDimensionOfConstraints(2)
self.setNumberOfNonzeroJacobian(8)
self.setNumberOfNonzeroHessian(10)
# trace objective function
x0 = np.array([1.0, 5.0, 5.0, 1.0])
self.getStartingPoint(x0)
f = np.array([0.0])
adolc.trace_on(1)
ax = adolc.adouble(x0)
af = adolc.adouble(f)
adolc.independent(ax)
self.costFunction(af, ax)
adolc.dependent(af)
adolc.trace_off()
# trace constraint function
adolc.trace_on(2)
ax = adolc.adouble(x0)
g = np.zeros(self.getDimensionOfConstraints())
ag = adolc.adouble(g)
adolc.independent(ax)
self.constraintFunction(ag, ax)
adolc.dependent(ag)
adolc.trace_off()
# trace lagrangian function
adolc.trace_on(3)
ax = adolc.adouble(x0)
alagrange = adolc.adouble([1., 1.])
aobj_factor = adolc.adouble(1.)
adolc.independent(ax)
adolc.independent(alagrange)
adolc.independent(aobj_factor)
ay = self.eval_lagrangian(ax, alagrange, aobj_factor)
adolc.dependent(ay)
adolc.trace_off()
x = np.array([1.0, 5.0, 5.0, 1.0])
hoptions = np.array([0, 0], dtype=int)
hess_result = adolc.colpack.sparse_hess_no_repeat(3, x, hoptions)
self.hrow = np.asarray(hess_result[1], dtype=int)
self.hcol = np.asarray(hess_result[2], dtype=int)
self.hess_values = np.asarray(hess_result[3], dtype=float)
# need only upper left part of the Hessian
self.mask = np.where(self.hcol < 4)
def tape_A(self, xtrace):
"""
Tape the objective function.
"""
print('Taping action evaluation...')
tstart = time.time()
adolc.trace_on(self.adolcID)
# set the active independent variables
ax = adolc.adouble(xtrace)
adolc.independent(ax)
# set the dependent variable (or vector of dependent variables)
af = self.A(ax)
adolc.dependent(af)
adolc.trace_off()
self.taped = True
print('Done!')
print('Time = {0} s\n'.format(time.time()-tstart))
def dydgamma(self, x, gamma):
"""
Calculate derivative of the neural network output with respect to the
hyperparameters gamma.
"""
gamma_c = gamma.copy()
gamma = ad.adouble(gamma)
tag = 11
ad.trace_on(tag)
ad.independent(gamma)
self.set_from_array(gamma)
y = self.eval(x)
ad.dependent(y)
ad.trace_off()
gamma = gamma_c
self.set_from_array(gamma_c)
dJdgamma = calc_jacobian(gamma, tag=tag, sparse=False)
return dJdgamma.reshape(gamma_c.shape)
def create_tape(dev, vPort):
"""
Generate Adol-C tape
Normally there is no need to manually call this.
"""
assert dev.isNonlinear
try:
tag = dev._tag
except AttributeError:
tag = dev.adolcID
dev._tag = tag
ad.trace_on(tag)
# Create derivative vector
a_vPort = ad.adouble(vPort)
ad.independent(a_vPort)
# perform actual calculation (for now re-tape every time)
a_out = dev.eval_cqs(a_vPort)
ad.dependent(a_out)
ad.trace_off()
import numpy
import math
import adolc
# tape a function evaluation
ax = numpy.array([adolc.adouble(0.) for n in range(2)])
# ay = adolc.adouble(0)
adolc.trace_on(13)
adolc.independent(ax)
ay = numpy.sin(ax[0] + ax[1]*ax[0])
adolc.dependent(ay)
adolc.trace_off()
x = numpy.array([3., 7.])
y = numpy.zeros(1)
adolc.tape_to_latex(13, x, y)
y = adolc.function(13, x)
g = adolc.gradient(13, x)
J = adolc.jacobian(13, x)
print('function y=', y)
print('gradient g=', g)
print('Jacobian J=', J)
title('Trajectory of an ODE')
xlabel(r'states $x(t),x_{p_1}(t)$ and $x_{p_2}(t)$')
ylabel(r' time $t$')
savefig('variational_ode_trajectory.eps')
# generate pseudo measurement data
p[0]+=3.; p[1] += 2.
x0 = array([v[0], 1., 0.])
x = explicit_euler(x0,f1,ts,p,q)
h = measurement_model(x[:,0],p,q)
etas = h + numpy.random.normal(size=Nm)
p[0]-= 3.; p[1] -= 2.
# taping F
av = array([adolc.adouble(0) for i in range(Nv)])
y = zeros(Nm)
adolc.trace_on(1)
av[0].is_independent(p[0])
av[1].is_independent(p[1])
av[2].is_independent(q[0])
ay = F(av[:Np],av[Np:],ts,Sigma,etas)
for m in range(Nm):
y[m] = adolc.depends_on(ay[m])
adolc.trace_off()
# taping dFdp
av = array([adolc.adouble(0) for i in range(Nv)])
adolc.trace_on(1)
av[0].is_independent(p[0])
av[1].is_independent(p[1])
print 'running runtime tests for A.shape = (D,P,N,N) = %d, %d, %d, %d'%(D,P,N,N)
A_data = numpy.random.rand(N,N,P,D)
Qbar_data = numpy.random.rand(1,N,N,P,D)
Rbar_data = numpy.random.rand(1,N,N,P,D)
#----------------------------------------------
# STEP 1:
# QR decomposition by Givens Rotations
# using pyadolc for the differentiation
#----------------------------------------------
A = A_data[:,:,0,0]
# trace QR decomposition with adolc
AP = AdolcProgram()
AP.trace_on(1)
aA = adouble(A)
AP.independent(aA)
aQ, aR = qr(aA)
AP.dependent(aQ)
AP.dependent(aR)
AP.trace_off()
for r in range(repetitions):
A = A_data[:,:,0,0]
# compute push forward
VA = A_data[:,:,:,1:]
tic = time()
out = AP.forward([A],[VA])
toc = time()
runtime_pyadolc_push_forward = toc - tic
symdict[xs[n,d]] = x[n,d]
return array([[dfs[n,d].subs(symdict).evalf() for d in range(D)] for n in range(N)])
def sym_ddf(x):
symdict = dict()
for n in range(N):
for d in range(D):
symdict[xs[n,d]] = x[n,d]
return array([[[[ ddfs[m,e,n,d].subs(symdict).evalf() for d in range(D)] for n in range(N)] for e in range(D)] for m in range(N)],dtype=float)
###################################
# PART 1: computation with PYADOLC
###################################
adolc.trace_on(0)
x = adolc.adouble(numpy.random.rand(*(N,D)))
adolc.independent(x)
y = f(x)
adolc.dependent(y)
adolc.trace_off()
# point at which the derivatives should be evaluated
x = random((N,D))
print '\n\n'
print 'Sympy function = function check (should be almost zero)'
print f(x) - sym_f(x)
print '\n\n'
print 'Sympy vs Hand Derived Gradient check (should be almost zero)'
# OBJECTIVE FUNCTION # ------------------ def Phi(F): return trace( dot(F.T,F)) def ffcn(x): return 0.5*array( [[(x[0]-17.)*(x[0]-17.), (x[0]-17.)*(x[0]-17.)], [ x[1]-19. , x[1]-19.]]) # TAPING THE FUNCTIONS # -------------------- # taping function ffcn u = 3.; v = 7. ax = array([adolc.adouble(u), adolc.adouble(v)]) adolc.trace_on(1) ax[0].is_independent(u) ax[1].is_independent(v) ay = ffcn(ax) for n in range(2): for m in range(2): adolc.depends_on(ay[n,m]) adolc.trace_off() # taping matrix functions with algopy x = array([u,v]) F = ffcn(x) Fdot = zeros((2,2)) cg = CGraph() FF = Function(Mtc(F))
## taping
start_time = time()
cg = rm.tape(f, x)
end_time = time()
tape_eval_times.append(end_time - start_time)
## reverse evaluation
start_time = time()
g_reverse2 = rm.gradient_from_graph(cg)
end_time = time()
rev_eval_times.append(end_time - start_time)
## PyADOLC taping
start_time = time()
ax = numpy.array([adolc.adouble(0.0) for i in range(N)])
adolc.trace_on(0)
for n in range(N):
ax[n].is_independent(x[n])
ay = f(ax)
adolc.depends_on(ay)
adolc.trace_off()
end_time = time()
adolc_tape_times.append(end_time - start_time)
## PyADOLC gradient
start_time = time()
adolc_g = adolc.gradient(0, x)
end_time = time()
adolc_gradient_times.append(end_time - start_time)
import numpy; from numpy import sin,cos; import adolc
def f(x):
return sin(x[0] + cos(x[1])*x[0])
adolc.trace_on(1)
x = adolc.adouble([3,7]); adolc.independent(x)
y = f(x)
adolc.dependent(y); adolc.trace_off()
adolc.tape_to_latex(1,[3,7],[0.])
# trace with ALGOPY start_time = time.time() cg = CGraph() x = Function(UTPM(numpy.random.rand(1, 1, N))) y = f(x) cg.trace_off() cg.independentFunctionList = [x] cg.dependentFunctionList = [y] end_time = time.time() time_trace_algopy = end_time - start_time # trace with PYADOLC start_time = time.time() adolc.trace_on(1) x = adolc.adouble(numpy.random.rand(N)) adolc.independent(x) y = f(x) adolc.dependent(y) adolc.trace_off() end_time = time.time() time_trace_adolc = end_time - start_time # trace with PYADOLC.cgraph from adolc.cgraph import AdolcProgram start_time = time.time() ap = AdolcProgram() ap.trace_on(2) x = adolc.adouble(numpy.random.rand(N)) ap.independent(x)
atmp = []
for n in range(N):
atmp.append(numpy.sin( numpy.sum(ax[:n])))
ay = numpy.array( [ ax[0] * numpy.sin( numpy.sum(atmp)) ] )
f = pycppad.adfun(ax, ay)
x = numpy.random.rand(N)
w = numpy.array( [ 1.] ) # compute Hessian of x0 * sin(x1)
cppad_hessian_runtime = timeit.Timer('f.hessian(x, w)', 'from __main__ import f,x,w').timeit(number=reps)/reps
cppad_gradient_runtime = timeit.Timer('f.jacobian(x)', 'from __main__ import f,x').timeit(number=reps)/reps
# adolc timing
x = numpy.zeros(N, dtype=float)
adolc.trace_on(0)
ax = adolc.adouble(x)
adolc.independent(ax)
atmp = []
for n in range(N):
atmp.append(numpy.sin( numpy.sum(ax[:n])))
ay = numpy.array( [ ax[0] * numpy.sin( numpy.sum(atmp)) ] )
adolc.dependent(ay)
adolc.trace_off()
x = numpy.random.rand(N)
adolc_hessian_runtime = timeit.Timer('adolc.hessian(0, x)', 'import adolc;from __main__ import x').timeit(number=reps)/reps
adolc_gradient_runtime = timeit.Timer('adolc.gradient(0, x)', 'import adolc;from __main__ import x').timeit(number=reps)/reps
adolc_hessian_runtimes.append(adolc_hessian_runtime)
cppad_hessian_runtimes.append(cppad_hessian_runtime)
return (exp(q[0]*t)-1.)/q[0] p = array([10.,2.]) q = array([-1.]) v = concatenate((p,q)) Np = size(p) Nq = size(q) Nv = size(v) Nm = 100 ts = linspace(0,10,Nm) # TAPING THE INTEGRATOR adolc.trace_on(1) av = adolc.adouble(v) adolc.independent(av) ax = explicit_euler(av[0],f,ts,av[:Np],av[Np:]) adolc.dependent(ax) adolc.trace_off() # COMPUTING FUNCTION AND JACOBIAN FROM THE TAPE y = adolc.zos_forward(1,v,0) J = adolc.jacobian(1,v) x_plot = plot(ts, y,'b') x_analytical_plot = plot(ts,phi(ts,p,q),'b.') xp0_plot = plot(ts, J[:,0], 'g') xp0_analytical_plot = plot(ts, phip0(ts,p,q), 'g.')
values[2] += lagrange[1] * 2
values[5] += lagrange[1] * 2
values[9] += lagrange[1] * 2
return values
def apply_new(x):
return True
x0 = numpy.array([1.0, 5.0, 5.0, 1.0])
pi0 = numpy.array([1.0, 1.0])
# trace objective function
adolc.trace_on(1)
ax = adolc.adouble(x0)
adolc.independent(ax)
ay = eval_f(ax)
adolc.dependent(ay)
adolc.trace_off()
# trace constraint function
adolc.trace_on(2)
ax = adolc.adouble(x0)
adolc.independent(ax)
ay = eval_g(ax)
adolc.dependent(ay)
adolc.trace_off()
# trace lagrangian function
adolc.trace_on(3)
