def run_laplacian_debug(self, x):
"""Compute the trained derivative (debug version)."""
n = len(x)
m = len(self.eq.bc)
H = len(self.v)
w = self.w
u = self.u
v = self.v
z = np.zeros((n, H))
for i in range(n):
for k in range(H):
z[i, k] = u[k]
for j in range(m):
z[i, k] += w[j, k] * x[i, j]
s = np.zeros((n, H))
for i in range(n):
for k in range(H):
s[i, k] = sigma.s(z[i, k])
s1 = np.zeros((n, H))
for i in range(n):
for k in range(H):
s1[i, k] = sigma.s1(s[i, k])
s2 = np.zeros((n, H))
for i in range(n):
for k in range(H):
s2[i, k] = sigma.s2(s[i, k])
N = np.zeros(n)
for i in range(n):
for k in range(H):
N[i] += s[i, k] * v[k]
delN = np.zeros((n, m))
for i in range(n):
for j in range(m):
for k in range(H):
delN[i, j] += v[k] * s1[i, k] * w[j, k]
del2N = np.zeros((n, m))
for i in range(n):
for j in range(m):
for k in range(H):
del2N[i, j] += v[k] * s2[i, k] * w[j, k]**2
del2Yt = np.zeros((n, m))
for i in range(n):
del2Yt[i] = self.tf.del2Yt(x[i], N[i], delN[i], del2N[i])
return del2Yt
def _train_delta_debug(self, x, opts=DEFAULT_OPTS):
"""Train using the delta method (debug version). """
my_opts = dict(DEFAULT_OPTS)
my_opts.update(opts)
# Sanity-check arguments.
assert len(x) > 0
assert opts['maxepochs'] > 0
assert opts['eta'] > 0
assert opts['vmin'] < opts['vmax']
assert opts['wmin'] < opts['wmax']
assert opts['umin'] < opts['umax']
# Determine the number of training points, independent variables, and
# hidden nodes.
n = len(x) # Number of training points
m = len(self.eq.bc)
H = len(self.v) # Number of hidden nodes
# Change notation for convenience.
debug = my_opts['debug']
verbose = my_opts['verbose']
eta = my_opts['eta'] # Learning rate
maxepochs = my_opts['maxepochs'] # Number of training epochs
wmin = my_opts['wmin'] # Network parameter limits
wmax = my_opts['wmax']
umin = my_opts['umin']
umax = my_opts['umax']
vmin = my_opts['vmin']
vmax = my_opts['vmax']
# Create the hidden node weights, biases, and output node weights.
w = np.random.uniform(wmin, wmax, (m, H))
u = np.random.uniform(umin, umax, H)
v = np.random.uniform(vmin, vmax, H)
# Initial parameter deltas are 0.
dE_dw = np.zeros((m, H))
dE_du = np.zeros(H)
dE_dv = np.zeros(H)
# Train the network.
for epoch in range(maxepochs):
if debug:
print('Starting epoch %d.' % epoch)
# Compute the new values of the network parameters.
for j in range(m):
for k in range(H):
w[j, k] -= eta*dE_dw[j, k]
for k in range(H):
u[k] -= eta*dE_du[k]
for k in range(H):
v[k] -= eta*dE_dv[k]
# Compute the input, the sigmoid function, and its derivatives,
# for each hidden node and each training point.
z = np.zeros((n, H))
for i in range(n):
for k in range(H):
z[i, k] = u[k]
for j in range(m):
z[i, k] += w[j, k]*x[i, j]
s = np.zeros((n, H))
for i in range(n):
for k in range(H):
s[i, k] = sigma.s(z[i, k])
s1 = np.zeros((n, H))
for i in range(n):
for k in range(H):
s1[i, k] = sigma.s1(s[i, k])
s2 = np.zeros((n, H))
for i in range(n):
for k in range(H):
s2[i, k] = sigma.s2(s[i, k])
# Compute the network output and its derivatives, for each
# training point.
N = np.zeros(n)
for i in range(n):
for k in range(H):
N[i] += v[k]*s[i, k]
P = np.zeros(n)
for i in range(n):
P[i] = self._P(x[i])
delP = np.zeros((n, m))
for i in range(n):
for j in range(m):
delP[i, j] = self._delP[j](self, x[i])
delN = np.zeros((n, m))
for i in range(n):
for j in range(m):
for k in range(H):
delN[i, j] += v[k]*s1[i, k]*w[j, k]
dN_dw = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
dN_dw[i, j, k] = v[k]*s1[i, k]*x[i, j]
dN_du = np.zeros((n, H))
for i in range(n):
for k in range(H):
dN_du[i, k] = v[k]*s1[i, k]
dN_dv = np.zeros((n, H))
for i in range(n):
for k in range(H):
dN_dv[i, k] = s[i, k]
d2N_dwdx = np.zeros((n, m, m, H))
for i in range(n):
for j in range(m):
for jj in range(m):
for k in range(H):
d2N_dwdx[i, j, jj, k] = (
v[k]*(s1[i, k]*kdelta(j, jj)
+ s2[i, k]*w[jj, k]*x[i, j])
)
d2N_dudx = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
d2N_dudx[i, j, k] = v[k]*s2[i, k]*w[j, k]
d2N_dvdx = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
d2N_dvdx[i, j, k] = s1[i, k]*w[j, k]
# Compute the value of the trial solution and its derivatives,
# for each training point.
Yt = np.zeros(n)
for i in range(n):
Yt[i] = self._Yt(x[i], N[i])
delYt = np.zeros((n, m))
for i in range(n):
for j in range(m):
delYt[i, j] = self._delYt[j](self, x[i], N[i], delN[i])
dYt_dw = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
dYt_dw[i, j, k] = P[i]*dN_dw[i, j, k]
dYt_du = np.zeros((n, H))
for i in range(n):
for k in range(H):
dYt_du[i, k] = P[i]*dN_du[i, k]
dYt_dv = np.zeros((n, H))
for i in range(n):
for k in range(H):
dYt_dv[i, k] = P[i]*dN_dv[i, k]
d2Yt_dwdx = np.zeros((n, m, m, H))
for i in range(n):
for j in range(m):
for jj in range(m):
for k in range(H):
d2Yt_dwdx[i, j, jj, k] = (
P[i]*d2N_dwdx[i, j, jj, k]
+ delP[i, jj]*dN_dw[i, j, k]
)
d2Yt_dudx = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
d2Yt_dudx[i, j, k] = (
P[i]*d2N_dudx[i, j, k]
+ delP[i, j]*dN_du[i, k]
)
d2Yt_dvdx = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
d2Yt_dvdx[i, j, k] = (
P[i]*d2N_dvdx[i, j, k]
+ delP[i, j]*dN_dv[i, k]
)
# Compute the value of the original differential equation for
# each training point, and its derivatives.
G = np.zeros(n)
for i in range(n):
G[i] = self.eq.G(x[i], Yt[i], delYt[i])
dG_dYt = np.zeros(n)
for i in range(n):
dG_dYt[i] = self.eq.dG_dY(x[i], Yt[i], delYt[i])
dG_ddelYt = np.zeros((n, m))
for i in range(n):
for j in range(m):
dG_ddelYt[i, j] = (
self.eq.dG_ddelY[j](x[i], Yt[i], delYt[i])
)
dG_dw = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
dG_dw[i, j, k] = dG_dYt[i]*dYt_dw[i, j, k]
for jj in range(m):
dG_dw[i, j, k] += (
dG_ddelYt[i, jj]*d2Yt_dwdx[i, j, jj, k]
)
dG_du = np.zeros((n, H))
for i in range(n):
for k in range(H):
dG_du[i, k] = dG_dYt[i]*dYt_du[i, k]
for j in range(m):
dG_du[i, k] += dG_ddelYt[i, j]*d2Yt_dudx[i, j, k]
dG_dv = np.zeros((n, H))
for i in range(n):
for k in range(H):
dG_dv[i, k] = dG_dYt[i]*dYt_dv[i, k]
for j in range(m):
dG_dv[i, k] += dG_ddelYt[i, j]*d2Yt_dvdx[i, j, k]
# Compute the error function for this epoch.
E = 0
for i in range(n):
E += G[i]**2
# Compute the partial derivatives of the error with respect to
# the network parameters.
dE_dw = np.zeros((m, H))
for j in range(m):
for k in range(H):
for i in range(n):
dE_dw[j, k] += 2*G[i]*dG_dw[i, j, k]
dE_du = np.zeros(H)
for k in range(H):
for i in range(n):
dE_du[k] += 2*G[i]*dG_du[i, k]
dE_dv = np.zeros(H)
for k in range(H):
for i in range(n):
dE_dv[k] += 2*G[i]*dG_dv[i, k]
# Compute the RMS error for this epoch.
rmse = sqrt(E/n)
if verbose:
print(epoch, rmse)
# Save the optimized parameters.
self.w = w
self.u = u
self.v = v
def _compute_error_debug(self, p, x):
"""Compute the error function using the current parameter values
(debug version)."""
# Determine the number of training points, independent variables, and
# hidden nodes.
n = len(x)
m = len(self.eq.bc)
H = len(self.v)
# Unpack the network parameters.
w = np.zeros((m, H))
for j in range(m):
w[j] = p[j * H:(j + 1) * H]
u = p[m * H:(m + 1) * H]
v = p[(m + 1) * H:(m + 2) * H]
# Compute the forward pass through the network.
z = np.zeros((n, H))
for i in range(n):
for k in range(H):
z[i, k] = u[k]
for j in range(m):
z[i, k] += w[j, k] * x[i, j]
s = np.zeros((n, H))
for i in range(n):
for k in range(H):
s[i, k] = sigma.s(z[i, k])
s1 = np.zeros((n, H))
for i in range(n):
for k in range(H):
s1[i, k] = sigma.s1(s[i, k])
s2 = np.zeros((n, H))
for i in range(n):
for k in range(H):
s2[i, k] = sigma.s2(s[i, k])
N = np.zeros(n)
for i in range(n):
for k in range(H):
N[i] += v[k] * s[i, k]
delN = np.zeros((n, m))
for i in range(n):
for j in range(m):
for k in range(H):
delN[i, j] += v[k] * s1[i, k] * w[j, k]
del2N = np.zeros((n, m))
for i in range(n):
for j in range(m):
for k in range(H):
del2N[i, j] += v[k] * s2[i, k] * w[j, k]**2
Yt = np.zeros(n)
for i in range(n):
Yt[i] = self.tf.Yt(x[i], N[i])
delYt = np.zeros((n, m))
for i in range(n):
delYt[i] = self.tf.delYt(x[i], N[i], delN[i])
del2Yt = np.zeros((n, m))
for i in range(n):
del2Yt[i] = self.tf.del2Yt(x[i], N[i], delN[i], del2N[i])
G = np.zeros(n)
for i in range(n):
G[i] = self.eq.G(x[i], Yt[i], delYt[i], del2Yt[i])
E2 = 0
for i in range(n):
E2 += G[i]**2
return E2
def _train_delta_debug(self, x, opts=DEFAULT_OPTS):
"""Train using the delta method (debug version)."""
my_opts = dict(DEFAULT_OPTS)
my_opts.update(opts)
# Determine the number of training points, independent variables, and
# hidden nodes.
n = len(x)
m = len(self.eq.bc)
H = len(self.v)
# Change notation for convenience.
debug = my_opts["debug"]
verbose = my_opts["verbose"]
eta = my_opts["eta"] # Learning rate
maxepochs = my_opts["maxepochs"] # Number of training epochs
wmin = my_opts["wmin"] # Network parameter limits
wmax = my_opts["wmax"]
umin = my_opts["umin"]
umax = my_opts["umax"]
vmin = my_opts["vmin"]
vmax = my_opts["vmax"]
# Create the hidden node weights, biases, and output node weights.
w = np.random.uniform(wmin, wmax, (m, H))
u = np.random.uniform(umin, umax, H)
v = np.random.uniform(vmin, vmax, H)
# w = np.zeros((m, H))
# u = np.zeros(H)
# v = np.zeros(H)
# Initial parameter deltas are 0.
dE_dw = np.zeros((m, H))
dE_du = np.zeros(H)
dE_dv = np.zeros(H)
# Train the network.
for epoch in range(maxepochs):
if debug:
print("Starting epoch %d." % epoch)
# Compute the new values of the network parameters.
for j in range(m):
for k in range(H):
w[j, k] -= eta * dE_dw[j, k]
for k in range(H):
u[k] -= eta * dE_du[k]
for k in range(H):
v[k] -= eta * dE_dv[k]
# Compute the input, the sigmoid function, and its derivatives,
# for each hidden node and each training point.
z = np.zeros((n, H))
for i in range(n):
for k in range(H):
z[i, k] = u[k]
for j in range(m):
z[i, k] += w[j, k] * x[i, j]
s = np.zeros((n, H))
for i in range(n):
for k in range(H):
s[i, k] = sigma.s(z[i, k])
s1 = np.zeros((n, H))
for i in range(n):
for k in range(H):
s1[i, k] = sigma.s1(s[i, k])
s2 = np.zeros((n, H))
for i in range(n):
for k in range(H):
s2[i, k] = sigma.s2(s[i, k])
s3 = np.zeros((n, H))
for i in range(n):
for k in range(H):
s3[i, k] = sigma.s3(s[i, k])
# Compute the network output and its derivatives, for each
# training point.
N = np.zeros(n)
for i in range(n):
for k in range(H):
N[i] += v[k] * s[i, k]
delN = np.zeros((n, m))
for i in range(n):
for j in range(m):
for k in range(H):
delN[i, j] += v[k] * s1[i, k] * w[j, k]
del2N = np.zeros((n, m))
for i in range(n):
for j in range(m):
for k in range(H):
del2N[i, j] += v[k] * s2[i, k] * w[j, k]**2
dN_dw = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
dN_dw[i, j, k] = v[k] * s1[i, k] * x[i, j]
dN_du = np.zeros((n, H))
for i in range(n):
for k in range(H):
dN_du[i, k] = v[k] * s1[i, k]
dN_dv = np.zeros((n, H))
for i in range(n):
for k in range(H):
dN_dv[i, k] = s[i, k]
d2N_dwdx = np.zeros((n, m, m, H))
for i in range(n):
for j in range(m):
for jj in range(m):
for k in range(H):
d2N_dwdx[i, j, jj,
k] = (v[k] *
(s1[i, k] * kdelta(j, jj) +
s2[i, k] * w[jj, k] * x[i, j]))
d2N_dudx = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
d2N_dudx[i, j, k] = v[k] * s2[i, k] * w[j, k]
d2N_dvdx = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
d2N_dvdx[i, j, k] = s1[i, k] * w[j, k]
d3N_dwdx2 = np.zeros((n, m, m, H))
for i in range(n):
for j in range(m):
for jj in range(m):
for k in range(H):
d3N_dwdx2[i, j, jj, k] = (
v[k] *
(2 * s2[i, k] * w[jj, k] * kdelta(j, jj) +
s3[i, k] * w[j, k]**2 * x[i, j]))
d3N_dudx2 = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
d3N_dudx2[i, j, k] = v[k] * s3[i, k] * w[j, k]**2
d3N_dvdx2 = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
d3N_dvdx2[i, j, k] = s2[i, k] * w[j, k]**2
# Compute the value of the trial solution and its derivatives,
# for each training point.
P = np.zeros(n)
for i in range(n):
P[i] = self.tf.P(x[i])
delP = np.zeros((n, m))
for i in range(n):
delP[i] = self.tf.delP(x[i])
del2P = np.zeros((n, m))
for i in range(n):
del2P[i] = self.tf.del2P(x[i])
Yt = np.zeros(n)
for i in range(n):
Yt[i] = self.tf.Yt(x[i], N[i])
delYt = np.zeros((n, m))
for i in range(n):
delYt[i] = self.tf.delYt(x[i], N[i], delN[i])
del2Yt = np.zeros((n, m))
for i in range(n):
del2Yt[i] = self.tf.del2Yt(x[i], N[i], delN[i], del2N[i])
dYt_dw = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
dYt_dw[i, j, k] = P[i] * dN_dw[i, j, k]
dYt_du = np.zeros((n, H))
for i in range(n):
for k in range(H):
dYt_du[i, k] = P[i] * dN_du[i, k]
dYt_dv = np.zeros((n, H))
for i in range(n):
for k in range(H):
dYt_dv[i, k] = P[i] * dN_dv[i, k]
d2Yt_dwdx = np.zeros((n, m, m, H))
for i in range(n):
for j in range(m):
for jj in range(m):
for k in range(H):
d2Yt_dwdx[i, j, jj,
k] = (P[i] * d2N_dwdx[i, j, jj, k] +
delP[i, jj] * dN_dw[i, j, k])
d2Yt_dudx = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
d2Yt_dudx[i, j, k] = (P[i] * d2N_dudx[i, j, k] +
delP[i, j] * dN_du[i, k])
d2Yt_dvdx = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
d2Yt_dvdx[i, j, k] = (P[i] * d2N_dvdx[i, j, k] +
delP[i, j] * dN_dv[i, k])
d3Yt_dwdx2 = np.zeros((n, m, m, H))
for i in range(n):
for j in range(m):
for jj in range(m):
for k in range(H):
d3Yt_dwdx2[i, j, jj, k] = (
P[i] * d3N_dwdx2[i, j, jj, k] +
2 * delP[i, jj] * d2N_dwdx[i, j, jj, k] +
del2P[i, jj] * dN_dw[i, j, k])
d3Yt_dudx2 = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
d3Yt_dudx2[i, j,
k] = (P[i] * d3N_dudx2[i, j, k] +
2 * delP[i, j] * d2N_dudx[i, j, k] +
del2P[i, j] * dN_du[i, k])
d3Yt_dvdx2 = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
d3Yt_dvdx2[i, j,
k] = (P[i] * d3N_dvdx2[i, j, k] +
2 * delP[i, j] * d2N_dvdx[i, j, k] +
del2P[i, j] * dN_dv[i, k])
# Compute the value of the original differential equation
# for each training point, and its derivatives.
G = np.zeros(n)
for i in range(n):
G[i] = self.eq.G(x[i], Yt[i], delYt[i], del2Yt[i])
dG_dYt = np.zeros(n)
for i in range(n):
dG_dYt[i] = self.eq.dG_dY(x[i], Yt[i], delYt[i], del2Yt[i])
dG_ddelYt = np.zeros((n, m))
for i in range(n):
for j in range(m):
dG_ddelYt[i,
j] = (self.eq.dG_ddelY[j](x[i], Yt[i], delYt[i],
del2Yt[i]))
dG_ddel2Yt = np.zeros((n, m))
for i in range(n):
for j in range(m):
dG_ddel2Yt[i,
j] = (self.eq.dG_ddel2Y[j](x[i], Yt[i],
delYt[i], del2Yt[i]))
dG_dw = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
dG_dw[i, j, k] = dG_dYt[i] * dYt_dw[i, j, k]
for jj in range(m):
dG_dw[i, j, k] += (
dG_ddelYt[i, jj] * d2Yt_dwdx[i, j, jj, k] +
dG_ddel2Yt[i, jj] * d3Yt_dwdx2[i, j, jj, k])
dG_du = np.zeros((n, H))
for i in range(n):
for k in range(H):
dG_du[i, k] = dG_dYt[i] * dYt_du[i, k]
for j in range(m):
dG_du[i, k] += (dG_ddelYt[i, j] * d2Yt_dudx[i, j, k] +
dG_ddel2Yt[i, j] * d3Yt_dudx2[i, j, k])
dG_dv = np.zeros((n, H))
for i in range(n):
for k in range(H):
dG_dv[i, k] = dG_dYt[i] * dYt_dv[i, k]
for j in range(m):
dG_dv[i, k] += (dG_ddelYt[i, j] * d2Yt_dvdx[i, j, k] +
dG_ddel2Yt[i, j] * d3Yt_dvdx2[i, j, k])
# Compute the error function for this epoch.
E2 = 0
for i in range(n):
E2 += G[i]**2
if verbose:
rmse = sqrt(E2 / n)
print(epoch, rmse)
# Compute the partial derivatives of the error with respect to the
# network parameters.
dE_dw = np.zeros((m, H))
for j in range(m):
for k in range(H):
for i in range(n):
dE_dw[j, k] += 2 * G[i] * dG_dw[i, j, k]
dE_du = np.zeros(H)
for k in range(H):
for i in range(n):
dE_du[k] += 2 * G[i] * dG_du[i, k]
dE_dv = np.zeros(H)
for k in range(H):
for i in range(n):
dE_dv[k] += 2 * G[i] * dG_dv[i, k]
# Save the optimized parameters.
self.w = w
self.u = u
self.v = v
def _compute_error_jacobian_debug(self, p, x):
"""Compute the Jacobian of the error function (debug version)."""
# Determine the number of training points, independent variables, and
# hidden nodes.
n = len(x)
m = len(self.eq.bc)
H = len(self.v)
# Unpack the network parameters.
w = np.zeros((m, H))
for j in range(m):
w[j] = p[j * H:(j + 1) * H]
u = p[m * H:(m + 1) * H]
v = p[(m + 1) * H:(m + 2) * H]
# Compute the forward pass through the network.
z = np.zeros((n, H))
for i in range(n):
for k in range(H):
z[i, k] = u[k]
for j in range(m):
z[i, k] += w[j, k] * x[i, j]
s = np.zeros((n, H))
for i in range(n):
for k in range(H):
s[i, k] = sigma.s(z[i, k])
s1 = np.zeros((n, H))
for i in range(n):
for k in range(H):
s1[i, k] = sigma.s1(s[i, k])
s2 = np.zeros((n, H))
for i in range(n):
for k in range(H):
s2[i, k] = sigma.s2(s[i, k])
s3 = np.zeros((n, H))
for i in range(n):
for k in range(H):
s3[i, k] = sigma.s3(s[i, k])
N = np.zeros(n)
for i in range(n):
for k in range(H):
N[i] += v[k] * s[i, k]
delN = np.zeros((n, m))
for i in range(n):
for j in range(m):
for k in range(H):
delN[i, j] += v[k] * s1[i, k] * w[j, k]
del2N = np.zeros((n, m))
for i in range(n):
for j in range(m):
for k in range(H):
del2N[i, j] += v[k] * s2[i, k] * w[j, k]**2
dN_dw = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
dN_dw[i, j, k] = v[k] * s1[i, k] * x[i, j]
dN_du = np.zeros((n, H))
for i in range(n):
for k in range(H):
dN_du[i, k] = v[k] * s1[i, k]
dN_dv = np.zeros((n, H))
for i in range(n):
for k in range(H):
dN_dv[i, k] = s[i, k]
d2N_dwdx = np.zeros((n, m, m, H))
for i in range(n):
for j in range(m):
for jj in range(m):
for k in range(H):
d2N_dwdx[i, j, jj,
k] = (v[k] * (s1[i, k] * kdelta(j, jj) +
s2[i, k] * w[jj, k] * x[i, j]))
d2N_dudx = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
d2N_dudx[i, j, k] = v[k] * s2[i, k] * w[j, k]
d2N_dvdx = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
d2N_dvdx[i, j, k] = s1[i, k] * w[j, k]
d3N_dwdx2 = np.zeros((n, m, m, H))
for i in range(n):
for j in range(m):
for jj in range(m):
for k in range(H):
d3N_dwdx2[i, j, jj, k] = (
v[k] * (2 * s2[i, k] * w[jj, k] * kdelta(j, jj) +
s3[i, k] * w[j, k]**2 * x[i, j]))
d3N_dudx2 = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
d3N_dudx2[i, j, k] = v[k] * s3[i, k] * w[j, k]**2
d3N_dvdx2 = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
d3N_dvdx2[i, j, k] = s2[i, k] * w[j, k]**2
P = np.zeros(n)
for i in range(n):
P[i] = self.tf.P(x[i])
delP = np.zeros((n, m))
for i in range(n):
delP[i] = self.tf.delP(x[i])
del2P = np.zeros((n, m))
for i in range(n):
del2P[i] = self.tf.del2P(x[i])
Yt = np.zeros(n)
for i in range(n):
Yt[i] = self.tf.Yt(x[i], N[i])
delYt = np.zeros((n, m))
for i in range(n):
delYt[i] = self.tf.delYt(x[i], N[i], delN[i])
del2Yt = np.zeros((n, m))
for i in range(n):
del2Yt[i] = self.tf.del2Yt(x[i], N[i], delN[i], del2N[i])
dYt_dw = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
dYt_dw[i, j, k] = P[i] * dN_dw[i, j, k]
dYt_du = np.zeros((n, H))
for i in range(n):
for k in range(H):
dYt_du[i, k] = P[i] * dN_du[i, k]
dYt_dv = np.zeros((n, H))
for i in range(n):
for k in range(H):
dYt_dv[i, k] = P[i] * dN_dv[i, k]
d2Yt_dwdx = np.zeros((n, m, m, H))
for i in range(n):
for j in range(m):
for jj in range(m):
for k in range(H):
d2Yt_dwdx[i, j, jj,
k] = (P[i] * d2N_dwdx[i, j, jj, k] +
delP[i, jj] * dN_dw[i, j, k])
d2Yt_dudx = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
d2Yt_dudx[i, j, k] = (P[i] * d2N_dudx[i, j, k] +
delP[i, j] * dN_du[i, k])
d2Yt_dvdx = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
d2Yt_dvdx[i, j, k] = (P[i] * d2N_dvdx[i, j, k] +
delP[i, j] * dN_dv[i, k])
d3Yt_dwdx2 = np.zeros((n, m, m, H))
for i in range(n):
for j in range(m):
for jj in range(m):
for k in range(H):
d3Yt_dwdx2[i, j, jj, k] = (
P[i] * d3N_dwdx2[i, j, jj, k] +
2 * delP[i, jj] * d2N_dwdx[i, j, jj, k] +
del2P[i, jj] * dN_dw[i, j, k])
d3Yt_dudx2 = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
d3Yt_dudx2[i, j, k] = (P[i] * d3N_dudx2[i, j, k] +
2 * delP[i, j] * d2N_dudx[i, j, k] +
del2P[i, j] * dN_du[i, k])
d3Yt_dvdx2 = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
d3Yt_dvdx2[i, j, k] = (P[i] * d3N_dvdx2[i, j, k] +
2 * delP[i, j] * d2N_dvdx[i, j, k] +
del2P[i, j] * dN_dv[i, k])
G = np.zeros(n)
for i in range(n):
G[i] = self.eq.G(x[i], Yt[i], delYt[i], del2Yt[i])
dG_dYt = np.zeros(n)
for i in range(n):
dG_dYt[i] = self.eq.dG_dY(x[i], Yt[i], delYt[i], del2Yt[i])
dG_ddelYt = np.zeros((n, m))
for i in range(n):
for j in range(m):
dG_ddelYt[i, j] = (self.eq.dG_ddelY[j](x[i], Yt[i], delYt[i],
del2Yt[i]))
dG_ddel2Yt = np.zeros((n, m))
for i in range(n):
for j in range(m):
dG_ddel2Yt[i, j] = (self.eq.dG_ddel2Y[j](x[i], Yt[i], delYt[i],
del2Yt[i]))
dG_dw = np.zeros((n, m, H))
for i in range(n):
for j in range(m):
for k in range(H):
dG_dw[i, j, k] = dG_dYt[i] * dYt_dw[i, j, k]
for jj in range(m):
dG_dw[i, j, k] += (
dG_ddelYt[i, jj] * d2Yt_dwdx[i, j, jj, k] +
dG_ddel2Yt[i, jj] * d3Yt_dwdx2[i, j, jj, k])
dG_du = np.zeros((n, H))
for i in range(n):
for k in range(H):
dG_du[i, k] = dG_dYt[i] * dYt_du[i, k]
for j in range(m):
dG_du[i, k] += (dG_ddelYt[i, j] * d2Yt_dudx[i, j, k] +
dG_ddel2Yt[i, j] * d3Yt_dudx2[i, j, k])
dG_dv = np.zeros((n, H))
for i in range(n):
for k in range(H):
dG_dv[i, k] = dG_dYt[i] * dYt_dv[i, k]
for j in range(m):
dG_dv[i, k] += (dG_ddelYt[i, j] * d2Yt_dvdx[i, j, k] +
dG_ddel2Yt[i, j] * d3Yt_dvdx2[i, j, k])
dE_dw = np.zeros((m, H))
for j in range(m):
for k in range(H):
for i in range(n):
dE_dw[j, k] += 2 * G[i] * dG_dw[i, j, k]
dE_du = np.zeros(H)
for k in range(H):
for i in range(n):
dE_du[k] += 2 * G[i] * dG_du[i, k]
dE_dv = np.zeros(H)
for k in range(H):
for i in range(n):
dE_dv[k] += 2 * G[i] * dG_dv[i, k]
jac = np.hstack((dE_dw.flatten(), dE_du, dE_dv))
return jac