def compute_psi2(self):
merg_psi, idx = contra(self.tensors[self.current_site],
self.order_left[self.current_site],
self.tensors[self.current_site + 1],
self.order_left[self.current_site + 1])
merg_psi_r, idx_r = contra(self.image_tensors[self.current_site],
self.order_right[self.current_site],
self.image_tensors[self.current_site + 1],
self.order_right[self.current_site + 1])
merg, idx = contra(merg_psi, idx, merg_psi_r, idx_r)
if self.current_site == 0:
psi, idx_psi = contra(merg, idx, self.contraction_psi2[2][0],
self.contraction_psi2[2][1])
else:
if self.current_site == self.n - 2:
psi, idx_psi = contra(
merg, idx, self.contraction_psi2[self.current_site - 1][0],
self.contraction_psi2[self.current_site - 1][1])
# Z, idz = contra(Z, idz, merg_l, idm_l)
else:
psi_r, idx_r = contra(
merg, idx, self.contraction_psi2[self.current_site - 1][0],
self.contraction_psi2[self.current_site - 1][1])
psi, idx_psi = contra(
psi_r, idx_r,
self.contraction_psi2[self.current_site + 2][0],
self.contraction_psi2[self.current_site + 2][1])
# Z, idz = contra(Z, idz, merg_l, idm_l)
return psi
def compute_Z(self):
merg, idm = contra(self.tensors[self.current_site],
self.order[self.current_site],
self.tensors[self.current_site + 1],
self.order[self.current_site + 1])
merg_l, idm_l = contra(self.tensors[self.current_site],
self.order_left[self.current_site],
self.tensors[self.current_site + 1],
self.order_left[self.current_site + 1])
merg_m, idm_m = contra(merg, idm, merg_l, idm_l)
if self.current_site == 0:
Z, idz = contra(merg_m, idm_m,
self.contraction_z[self.current_site + 2][0],
self.contraction_z[self.current_site + 2][1])
# Z, idz = contra(Z, idz, merg_l,idm_l)
else:
if self.current_site == self.n - 2:
Z, idz = contra(merg_m, idm_m,
self.contraction_z[self.current_site - 1][0],
self.contraction_z[self.current_site - 1][1])
# Z, idz = contra(Z, idz, merg_l, idm_l)
else:
Z_l, idz_l = contra(
merg_m, idm_m,
self.contraction_z[self.current_site - 1][0],
self.contraction_z[self.current_site - 1][1])
Z, idz = contra(Z_l, idz_l,
self.contraction_z[self.current_site + 2][0],
self.contraction_z[self.current_site + 2][1])
# Z, idz = contra(Z, idz, merg_l, idm_l)
return Z
def contraction_update_all_left2(self):
# v1=torch.Tensor([0,1])
# v2=torch.Tensor([1,0])
self.going_righ = 0
merg, idm = contra(self.tensors[0], self.order_left[0],
self.tensors[0], self.order[0])
self.contraction_z[0] = ([merg, idm])
merg_psi, merg_psi_idx = contra(self.tensors[0], self.order_left[0],
self.image_tensors[0],
self.order_right[0])
self.contraction_psi2[0] = ([merg_psi, merg_psi_idx])
for i in range(1, self.n - 2):
merg, idm = contra(merg, idm, self.tensors[i], self.order_left[i])
merg, idm = contra(merg, idm, self.tensors[i], self.order[i])
self.contraction_z[i] = ([merg, idm])
merg_psi, merg_psi_idx = contra(merg_psi, merg_psi_idx,
self.tensors[i],
self.order_left[i])
merg_psi, merg_psi_idx = contra(merg_psi, merg_psi_idx,
self.image_tensors[i],
self.order_right[i])
self.contraction_psi2[i] = ([merg_psi, merg_psi_idx])
def contraction_update_all_left(self):
v1 = torch.Tensor([0, 1]).cuda()
v2 = torch.Tensor([1, 0]).cuda()
self.going_righ = 0
merg, idm = contra(self.tensors[0], self.order_left[0],
self.tensors[0], self.order[0])
self.contraction_z[0] = ([merg, idm])
merg_psi = []
for j in range(self.m):
if self.image[0, j] == 1: #n*m
mergv, idv = contra(self.tensors[0], self.order_left[0], v1,
[0])
else:
mergv, idv = contra(self.tensors[0], self.order_left[0], v2,
[0])
merg_psi.append(mergv)
self.contraction_psi[0] = ([merg_psi, idv])
for i in range(1, self.n - 2):
merg, idm = contra(merg, idm, self.tensors[i], self.order_left[i])
merg, idm = contra(merg, idm, self.tensors[i], self.order[i])
self.contraction_z[i] = ([merg, idm])
merg_psi = []
for j in range(0, self.m):
if self.image[i, j] == 1: # n*m
mergv, idvm = contra(self.contraction_psi[i - 1][0][j],
idv, self.tensors[i],
self.order_left[i])
mergv, idvm = contra(mergv, idvm, v1, [2 * i])
else:
mergv, idvm = contra(self.contraction_psi[i - 1][0][j],
idv, self.tensors[i],
self.order_left[i])
mergv, idvm = contra(mergv, idvm, v2, [2 * i])
merg_psi.append(mergv)
idv = idvm
self.contraction_psi[i] = ([merg_psi, idvm])
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "2"
d1 = 30
d2 = 30
r = 10
learning_rate = 0.1
loops = 30
his_loss = torch.zeros(1, loops)
A = torch.rand(d1, d2).cuda()
E = torch.rand(d1, d2).cuda()
M = torch.rand(d1, r).cuda()
N = torch.rand(r, d2).cuda()
A = A / A.norm()
E = E / E.norm()
M = M / M.norm()
N = N / N.norm()
t1 = time.time()
for loop in range(loops):
A0 = (contra(M, [1, 2], N, [2, 3]))[0]
loss = (contra((A - A0), [1, 2], E, [1, 2]))[0]
grad_M = 2 * loss * contra(E, [0, 2], N, [1, 2])[0]
grad_N = 2 * loss * contra(M, [1, 2], E, [1, 3])[0]
M = M + learning_rate * grad_M
N = N + learning_rate * grad_N
his_loss[0, loop] = loss**2
t2 = time.time()
print(his_loss, t2 - t1)
def train(self, loops):
for loop in range(loops):
self.going_righ = 0
for i in np.arange(self.n - 2, 0, -1):
self.current_site = i
self.merged_tensor, self.merged_idx = contra(
self.tensors[self.current_site],
self.order_left[self.current_site],
self.tensors[self.current_site + 1],
self.order_left[self.current_site + 1])
self.Z = self.compute_Z()
self.psi = self.compute_psi()
dmerge = self.gradient_descent
# nll=0
# for k in range(self.m):
# nll=nll+(torch.log(self.psi[k] *self.psi[k] /self.Z))
# nll=nll/self.m
# print nll,i
self.tensors[self.current_site], self.tensors[
self.current_site + 1] = svd_update(
self.tensors[self.current_site],
self.order[self.current_site],
self.tensors[self.current_site + 1],
self.order[self.current_site + 1], dmerge,
self.going_righ, 0.0002, self.max_bond)
self.contraction_updat_twosite()
self.going_righ = 1
for i in range(self.n - 2):
self.current_site = i
self.merged_tensor, self.merged_idx = contra(
self.tensors[self.current_site],
self.order_left[self.current_site],
self.tensors[self.current_site + 1],
self.order_left[self.current_site + 1])
self.Z = self.compute_Z()
self.psi = self.compute_psi()
# nll = 0
# for k in range(self.m):
# nll = nll + ( torch.log(self.psi[k] *self.psi[k] / self.Z))
# nll=nll/self.m
# print nll,i
dmerge = self.gradient_descent
self.tensors[self.current_site], self.tensors[
self.current_site + 1] = svd_update(
self.tensors[self.current_site],
self.order[self.current_site],
self.tensors[self.current_site + 1],
self.order[self.current_site + 1], dmerge,
self.going_righ, 0.0002, self.max_bond)
self.contraction_updat_twosite()
for j in range(len(self.links)):
self.Z = self.compute_Z()
self.psi = self.compute_psi()
k0 = self.links[j][0]
k1 = self.links[j][1]
dmerge = self.gradient_descent5(k0, k1)
self.tensors[k0], self.tensors[k1] = svd_update(
self.tensors[k0], self.order[k0], self.tensors[k1],
self.order[k1], dmerge, self.going_righ, 0.0002,
self.max_bond)
self.contraction_update_all_left()
nll = 0
for k in range(self.m):
nll = nll + (torch.log(self.psi[k] * self.psi[k] / self.Z))
nll = nll / self.m
print nll
if nll > self.nll_history[-1] + 0.002:
self.nll_history.append(nll)
else:
break
def gradient_descent5(self, j, k):
self.current_site = j
self.merged_tensor, self.merged_idx = contra(self.tensors[j],
self.order_left[j],
self.tensors[k],
self.order_left[k])
merged_idx2 = self.merged_idx[:]
for l in range(len(self.merged_idx)):
if self.merged_idx[l] % 2 == 1:
merged_idx2[l] = merged_idx2[l] - 8 * self.n
mer_l, mer_l_idx = contra(self.tensors[0], self.order_left[0],
self.tensors[0], self.order[0])
for i in range(1, j):
mer_l, mer_l_idx = contra(mer_l, mer_l_idx, self.tensors[i],
self.order_left[i])
mer_l, mer_l_idx = contra(mer_l, mer_l_idx, self.tensors[i],
self.order[i])
mer_r, mer_r_idx = contra(self.tensors[self.n - 1],
self.order_left[self.n - 1],
self.tensors[self.n - 1],
self.order[self.n - 1])
for i in np.arange(self.n - 2, k, -1):
mer_r, mer_r_idx = contra(mer_r, mer_r_idx, self.tensors[i],
self.order_left[i])
mer_r, mer_r_idx = contra(mer_r, mer_r_idx, self.tensors[i],
self.order[i])
mer_m, mer_m_idx = contra(self.tensors[j + 1], self.order_left[j + 1],
self.tensors[j + 1], self.order[j + 1])
for i in range(j + 2, k):
mer_m, mer_m_idx = contra(mer_m, mer_m_idx, self.tensors[i],
self.order_left[i])
mer_m, mer_m_idx = contra(mer_m, mer_m_idx, self.tensors[i],
self.order[i])
dZ, dZ_idx = contra(mer_l, mer_l_idx, self.merged_tensor, merged_idx2)
dZ, dZ_idx = contra(dZ, dZ_idx, mer_r, mer_r_idx)
dZ, dZ_idx = contra(dZ, dZ_idx, mer_m, mer_m_idx)
dpsi = []
dPsi = []
v1 = torch.Tensor([0, 1]).cuda()
v2 = torch.Tensor([1, 0]).cuda()
for jl in range(self.m):
if self.image[0, jl] == 1:
psi_l, psi_l_idx = contra(self.tensors[0], self.order_left[0],
v1, [0])
else:
psi_l, psi_l_idx = contra(self.tensors[0], self.order_left[0],
v2, [0])
dpsi.append(psi_l)
for i in range(1, j):
for jl in range(self.m):
dpsi[jl], psi_l_idx2 = contra(dpsi[jl], psi_l_idx,
self.tensors[i],
self.order_left[i])
if self.image[j, jl] == 1:
psi_l, psi_l_idx2 = contra(dpsi[jl], psi_l_idx2, v1,
[2 * i])
else:
psi_l, psi_l_idx2 = contra(dpsi[jl], psi_l_idx2, v2,
[2 * i])
dpsi[jl] = psi_l
psi_l_idx = psi_l_idx2[:]
for i in range(j + 1, k):
for jl in range(self.m):
dpsi[jl], psi_l_idx2 = contra(dpsi[jl], psi_l_idx,
self.tensors[i],
self.order_left[i])
if self.image[j, jl] == 1:
psi_l, psi_l_idx2 = contra(dpsi[jl], psi_l_idx2, v1,
[2 * i])
else:
psi_l, psi_l_idx2 = contra(dpsi[jl], psi_l_idx2, v2,
[2 * i])
dpsi[jl] = psi_l
psi_l_idx = psi_l_idx2[:]
for i in range(k + 1, self.n):
for jl in range(self.m):
dpsi[jl], psi_l_idx2 = contra(dpsi[jl], psi_l_idx,
self.tensors[i],
self.order_left[i])
if self.image[j, jl] == 1:
psi_l, psi_l_idx2 = contra(dpsi[jl], psi_l_idx2, v1,
[2 * i])
else:
psi_l, psi_l_idx2 = contra(dpsi[jl], psi_l_idx2, v2,
[2 * i])
dpsi[jl] = psi_l
psi_l_idx = psi_l_idx2[:]
dPsi.append(dpsi + [psi_l_idx])
dmerge = torch.zeros((self.merged_tensor).size()).cuda()
for im1 in [1, 2]:
for im2 in [1, 2]:
dpsi = torch.zeros((dPsi[0][0]).size()).cuda()
im = (self.image[self.current_site, :]
== im1) * (self.image[self.current_site + 1, :] == im2)
for jl in range(self.m):
if im[jl] == 1:
dpsi = 2 * dPsi[0][jl] / self.psi[jl] + dpsi
dmerge = tensor_slide(dmerge, self.merged_idx,
[2 - im1, 2 - im2], [j * 2, k * 2],
dpsi, (dPsi[0][-1]))
dmerge = tensor_add(dmerge / self.n, self.merged_idx, -2 * dZ / self.Z,
dZ_idx)
gnorm = torch.norm(dmerge) / 20
if (gnorm < 1.0): # % & & self.bond_dims(self.current_bond) <= 50;
dmerge = dmerge / gnorm
dmerge = self.merged_tensor + self.learning_rate * dmerge
return dmerge
def gradient_descent(self):
dpsi = []
dPsi = []
merged_idx2 = self.merged_idx[:]
for k in range(len(self.merged_idx)):
if self.merged_idx[k] % 2 == 1:
merged_idx2[k] = merged_idx2[k] - 8 * self.n
if self.current_site == 0:
dZ, dZ_idx = contra(self.merged_tensor, merged_idx2,
self.contraction_z[self.current_site + 2][0],
self.contraction_z[self.current_site + 2][1])
for j in range(self.m):
dpsi.append(self.contraction_psi[self.current_site + 2][0][j])
dpsi_idx = self.contraction_psi[self.current_site + 2][1]
dPsi.append(dpsi + [dpsi_idx])
else:
if self.current_site == self.n - 2:
dZ, dZ_idx = contra(
self.merged_tensor, merged_idx2,
self.contraction_z[self.current_site - 1][0],
self.contraction_z[self.current_site - 1][1])
for j in range(self.m):
dpsi.append(self.contraction_psi[self.current_site -
1][0][j])
dpsi_idx = self.contraction_psi[self.current_site - 1][1]
dPsi.append(dpsi + [dpsi_idx])
else:
merg_Z, idz = contra(
self.merged_tensor, merged_idx2,
self.contraction_z[self.current_site - 1][0],
self.contraction_z[self.current_site - 1][1])
dZ, dZ_idx = contra(
merg_Z, idz, self.contraction_z[self.current_site + 2][0],
self.contraction_z[self.current_site + 2][1])
for j in range(self.m):
merg_psi, dpsi_idx = contra(
self.contraction_psi[self.current_site - 1][0][j],
self.contraction_psi[self.current_site - 1][1],
self.contraction_psi[self.current_site + 2][0][j],
self.contraction_psi[self.current_site + 2][1])
dpsi.append(merg_psi)
dPsi.append(dpsi + [dpsi_idx])
dmerge = torch.zeros((self.merged_tensor).size()).cuda()
for im1 in [1, 2]:
for im2 in [1, 2]:
dpsi = torch.zeros((dPsi[0][0]).size()).cuda()
im = (self.image[self.current_site, :]
== im1) * (self.image[self.current_site + 1, :] == im2)
for j in range(self.m):
if im[j] == 1:
dpsi = 2 * dPsi[0][j] / self.psi[j] + dpsi
dmerge = tensor_slide(
dmerge, self.merged_idx, [2 - im1, 2 - im2],
[self.current_site * 2, self.current_site * 2 + 2],
dpsi, (dPsi[0][-1]))
dmerge = tensor_add(dmerge / self.n, self.merged_idx, -2 * dZ / self.Z,
dZ_idx)
gnorm = torch.norm(dmerge) / 20
if (gnorm < 1.0): #% & & self.bond_dims(self.current_bond) <= 50;
dmerge = dmerge / gnorm
dmerge = self.merged_tensor + self.learning_rate * dmerge
return dmerge
def compute_psi(self):
Psi = []
v1 = torch.Tensor([0, 1]).cuda()
v2 = torch.Tensor([1, 0]).cuda()
if self.current_site == 0:
for j in range(self.m):
if self.image[0, j] == 1: # n*m
mergv, idvm = contra(self.tensors[0], self.order_left[0],
v1, [0])
mergv, idvm = contra(mergv, idvm, self.tensors[1],
self.order_left[1])
if self.image[1, j] == 1:
mergv, idvm = contra(mergv, idvm, v1, [2])
else:
mergv, idvm = contra(mergv, idvm, v2, [2])
else:
mergv, idvm = contra(self.tensors[0], self.order_left[0],
v2, [0])
mergv, idvm = contra(mergv, idvm, self.tensors[1],
self.order_left[1])
if self.image[1, j] == 1:
mergv, idvm = contra(mergv, idvm, v1, [2])
else:
mergv, idvm = contra(mergv, idvm, v2, [2])
psi, id = contra(mergv, idvm, self.contraction_psi[2][0][j],
self.contraction_psi[2][1])
Psi.append(psi)
else:
if self.current_site == self.n - 2:
for j in range(self.m):
if self.image[self.n - 2, j] == 1: # n*m
mergv, idvm = contra(self.tensors[self.n - 2],
self.order_left[self.n - 2], v1,
[2 * self.n - 4])
mergv, idvm = contra(mergv, idvm,
self.tensors[self.n - 1],
self.order_left[self.n - 1])
if self.image[self.n - 1, j] == 1:
mergv, idvm = contra(mergv, idvm, v1,
[2 * self.n - 2])
else:
mergv, idvm = contra(mergv, idvm, v2,
[2 * self.n - 2])
else:
mergv, idvm = contra(self.tensors[self.n - 2],
self.order_left[self.n - 2], v2,
[2 * self.n - 4])
mergv, idvm = contra(mergv, idvm,
self.tensors[self.n - 1],
self.order_left[self.n - 1])
if self.image[self.n - 1, j] == 1:
mergv, idvm = contra(mergv, idvm, v1,
[2 * self.n - 2])
else:
mergv, idvm = contra(mergv, idvm, v2,
[2 * self.n - 2])
psi, id = contra(mergv, idvm,
self.contraction_psi[self.n - 3][0][j],
self.contraction_psi[self.n - 3][1])
Psi.append(psi)
else:
for j in range(self.m):
if self.image[self.current_site, j] == 1: # n*m
mergv, idvm = contra(
self.tensors[self.current_site],
self.order_left[self.current_site], v1,
[2 * self.current_site])
mergv, idvm = contra(
mergv, idvm, self.tensors[self.current_site + 1],
self.order_left[self.current_site + 1])
if self.image[self.current_site + 1, j] == 1:
mergv, idvm = contra(mergv, idvm, v1,
[2 * self.current_site + 2])
else:
mergv, idvm = contra(mergv, idvm, v2,
[2 * self.current_site + 2])
else:
mergv, idvm = contra(
self.tensors[self.current_site],
self.order_left[self.current_site], v2,
[2 * self.current_site])
mergv, idvm = contra(
mergv, idvm, self.tensors[self.current_site + 1],
self.order_left[self.current_site + 1])
if self.image[self.current_site + 1, j] == 1:
mergv, idvm = contra(mergv, idvm, v1,
[2 * self.current_site + 2])
else:
mergv, idvm = contra(mergv, idvm, v2,
[2 * self.current_site + 2])
mergv, idvm = contra(
mergv, idvm,
self.contraction_psi[self.current_site - 1][0][j],
self.contraction_psi[self.current_site - 1][1])
psi, id = contra(
mergv, idvm,
self.contraction_psi[self.current_site + 2][0][j],
self.contraction_psi[self.current_site + 2][1])
Psi.append(psi)
return Psi
def contraction_updat_twosite(self):
v1 = torch.Tensor([0, 1]).cuda()
v2 = torch.Tensor([1, 0]).cuda()
if self.going_righ == 0:
if self.current_site == self.n - 2:
merg, idm = contra(self.tensors[self.n - 1],
self.order[self.n - 1],
self.tensors[self.n - 1],
self.order_left[self.n - 1])
# self.contraction_z[self.current_site+1]=merg
merg_psi = []
for j in range(self.m):
if self.image[self.n - 1, j] == 1: # n*m
mergv, idvm = contra(self.tensors[self.n - 1],
self.order_left[self.n - 1], v1,
[self.n * 2 - 2])
else:
mergv, idvm = contra(self.tensors[self.n - 1],
self.order_left[self.n - 1], v2,
[self.n * 2 - 2])
merg_psi.append(mergv)
else:
merg, idm = contra(
self.tensors[self.current_site + 1],
self.order_left[self.current_site + 1],
self.contraction_z[self.current_site + 2][0],
self.contraction_z[self.current_site + 2][1])
merg, idm = contra(merg, idm,
self.tensors[self.current_site + 1],
self.order[self.current_site + 1])
merg_psi = []
for j in range(0, self.m):
if self.image[self.current_site + 1, j] == 1: # n*m
mergv, idvm = contra(
self.tensors[self.current_site + 1],
self.order_left[self.current_site + 1],
self.contraction_psi[self.current_site + 2][0][j],
self.contraction_psi[self.current_site + 2][1])
mergv, idvm = contra(mergv, idvm, v1,
[2 * self.current_site + 2])
else:
mergv, idvm = contra(
self.tensors[self.current_site + 1],
self.order_left[self.current_site + 1],
self.contraction_psi[self.current_site + 2][0][j],
self.contraction_psi[self.current_site + 2][1])
mergv, idvm = contra(mergv, idvm, v2,
[2 * self.current_site + 2])
merg_psi.append(mergv)
self.contraction_psi[self.current_site + 1] = ([merg_psi, idvm])
self.contraction_z[self.current_site + 1] = ([merg, idm])
if self.going_righ == 1:
if self.current_site == 0:
merg, idm = contra(self.tensors[0], self.order[0],
self.tensors[0], self.order_left[0])
# self.contraction_z[self.current_site+1]=merg
merg_psi = []
for j in range(self.m):
if self.image[0, j] == 1: # n*m
mergv, idvm = contra(self.tensors[0],
self.order_left[0], v1, [0])
else:
mergv, idvm = contra(self.tensors[0],
self.order_left[0], v2, [0])
merg_psi.append(mergv)
else:
merg, idm = contra(
self.tensors[self.current_site],
self.order_left[self.current_site],
self.contraction_z[self.current_site - 1][0],
self.contraction_z[self.current_site - 1][1])
merg, idm = contra(merg, idm, self.tensors[self.current_site],
self.order[self.current_site])
merg_psi = []
for j in range(0, self.m):
if self.image[self.current_site, j] == 1: # n*m
mergv, idvm = contra(
self.tensors[self.current_site],
self.order_left[self.current_site],
self.contraction_psi[self.current_site - 1][0][j],
self.contraction_psi[self.current_site - 1][1])
mergv, idvm = contra(mergv, idvm, v1,
[2 * self.current_site])
else:
mergv, idvm = contra(
self.tensors[self.current_site],
self.order_left[self.current_site],
self.contraction_psi[self.current_site - 1][0][j],
self.contraction_psi[self.current_site - 1][1])
mergv, idvm = contra(mergv, idvm, v2,
[2 * self.current_site])
merg_psi.append(mergv)
self.contraction_psi[self.current_site] = ([merg_psi, idvm])
self.contraction_z[self.current_site] = ([merg, idm])
def gradient_descent2(self):
dpsi = []
dPsi = []
merged_idx2 = self.merged_idx[:]
for k in range(len(self.merged_idx)):
if self.merged_idx[k] % 2 == 1:
merged_idx2[k] = merged_idx2[k] - 8 * self.n
merge_psi, merge_psi_idx = contra(
self.image_tensors[self.current_site],
self.order_right[self.current_site],
self.image_tensors[self.current_site + 1],
self.order_right[self.current_site + 1])
if self.current_site == 0:
dZ, dZ_idx = contra(self.merged_tensor, merged_idx2,
self.contraction_z[self.current_site + 2][0],
self.contraction_z[self.current_site + 2][1])
dPsi, dPsi_idx = contra(
merge_psi, merge_psi_idx,
self.contraction_psi2[self.current_site + 2][0],
self.contraction_psi2[self.current_site + 2][1])
else:
if self.current_site == self.n - 2:
dZ, dZ_idx = contra(
self.merged_tensor, merged_idx2,
self.contraction_z[self.current_site - 1][0],
self.contraction_z[self.current_site - 1][1])
dPsi, dPsi_idx = contra(
merge_psi, merge_psi_idx,
self.contraction_psi2[self.current_site - 1][0],
self.contraction_psi2[self.current_site - 1][1])
else:
merg_Z, idz = contra(
self.merged_tensor, merged_idx2,
self.contraction_z[self.current_site - 1][0],
self.contraction_z[self.current_site - 1][1])
dZ, dZ_idx = contra(
merg_Z, idz, self.contraction_z[self.current_site + 2][0],
self.contraction_z[self.current_site + 2][1])
merg_Psi, idpsi = contra(
merge_psi, merge_psi_idx,
self.contraction_psi2[self.current_site - 1][0],
self.contraction_psi2[self.current_site - 1][1])
dPsi, dPsi_idx = contra(
merg_Psi, idpsi,
self.contraction_psi2[self.current_site + 2][0],
self.contraction_psi2[self.current_site + 2][1])
# for im1 in [1,2]:
# for im2 in [1,2]:
# dpsi=torch.zeros((dPsi[0][0]).size()).cuda()
# im=(self.image[self.current_site,:]==im1)*(self.image[self.current_site+1,:]==im2)
# for j in range(self.m):
# if im[j]==1:
# dpsi=2*dPsi[0][j]/self.psi[j]+dpsi
# dmerge=tensor_slide(dmerge,self.merged_idx,[2-im1,2-im2],[self.current_site*2,self.current_site*2+2],dpsi,(dPsi[0][-1]))
psi_1 = 1.0 / self.psi
dPsi, dPsi_idx = contra(dPsi, dPsi_idx, psi_1, [-1])
dmerge = tensor_add(2 * dPsi, dPsi_idx, -2 * 4 * self.m * dZ / self.Z,
dZ_idx)
gnorm = torch.norm(dmerge) / 80
if (gnorm < 1.0): #% & & self.bond_dims(self.current_bond) <= 50;
dmerge = dmerge / gnorm
dmerge = tensor_add(self.merged_tensor, self.merged_idx,
self.learning_rate * dmerge, dPsi_idx)
return dmerge
def gradient_descent25(self, j, k):
self.current_site = j
self.merged_tensor, self.merged_idx = contra(self.tensors[j],
self.order_left[j],
self.tensors[k],
self.order_left[k])
merged_idx2 = self.merged_idx[:]
psi_merg, psi_merg_idx = contra(self.image_tensors[j],
self.order_right[j],
self.image_tensors[k],
self.order_right[k])
for l in range(len(self.merged_idx)):
if self.merged_idx[l] % 2 == 1:
merged_idx2[l] = merged_idx2[l] - 8 * self.n
mer_l, mer_l_idx = contra(self.tensors[0], self.order_left[0],
self.tensors[0], self.order[0])
psi_l, psi_l_idx = contra(self.tensors[0], self.order_left[0],
self.image_tensors[0], self.order_right[0])
for i in range(1, j):
mer_l, mer_l_idx = contra(mer_l, mer_l_idx, self.tensors[i],
self.order_left[i])
mer_l, mer_l_idx = contra(mer_l, mer_l_idx, self.tensors[i],
self.order[i])
psi_l, psi_l_idx = contra(psi_l, psi_l_idx, self.tensors[i],
self.order_left[i])
psi_l, psi_l_idx = contra(psi_l, psi_l_idx, self.image_tensors[i],
self.order_right[i])
mer_r, mer_r_idx = contra(self.tensors[self.n - 1],
self.order_left[self.n - 1],
self.tensors[self.n - 1],
self.order[self.n - 1])
psi_r, psi_r_idx = contra(self.tensors[self.n - 1],
self.order_left[self.n - 1],
self.image_tensors[self.n - 1],
self.order_right[self.n - 1])
for i in np.arange(self.n - 2, k, -1):
mer_r, mer_r_idx = contra(mer_r, mer_r_idx, self.tensors[i],
self.order_left[i])
mer_r, mer_r_idx = contra(mer_r, mer_r_idx, self.tensors[i],
self.order[i])
psi_r, psi_r_idx = contra(psi_r, psi_r_idx, self.tensors[i],
self.order_left[i])
psi_r, psi_r_idx = contra(psi_r, psi_r_idx, self.image_tensors[i],
self.order_right[i])
mer_m, mer_m_idx = contra(self.tensors[j + 1], self.order_left[j + 1],
self.tensors[j + 1], self.order[j + 1])
psi_m, psi_m_idx = contra(self.tensors[j + 1], self.order_left[j + 1],
self.image_tensors[j + 1],
self.order_right[j + 1])
for i in range(j + 2, k):
mer_m, mer_m_idx = contra(mer_m, mer_m_idx, self.tensors[i],
self.order_left[i])
mer_m, mer_m_idx = contra(mer_m, mer_m_idx, self.tensors[i],
self.order[i])
psi_m, psi_m_idx = contra(psi_m, psi_m_idx, self.tensors[i],
self.order_left[i])
psi_m, psi_m_idx = contra(psi_m, psi_m_idx, self.image_tensors[i],
self.order_right[i])
dZ, dZ_idx = contra(mer_l, mer_l_idx, self.merged_tensor, merged_idx2)
dZ, dZ_idx = contra(dZ, dZ_idx, mer_r, mer_r_idx)
dZ, dZ_idx = contra(dZ, dZ_idx, mer_m, mer_m_idx)
dpsi, dpsi_idx = contra(psi_l, psi_l_idx, psi_merg, psi_merg_idx)
dpsi, dpsi_idx = contra(dpsi, dpsi_idx, psi_r, psi_r_idx)
dpsi, dpsi_idx = contra(dpsi, dpsi_idx, psi_m, psi_m_idx)
psi_1 = 1.0 / self.psi
dpsi, dpsi_idx = contra(dpsi, dpsi_idx, psi_1, [-1])
dmerge = tensor_add(2 * dpsi / self.m, dpsi_idx, -2 * 4 * dZ / self.Z,
dZ_idx)
gnorm = torch.norm(dmerge) / 80
if (gnorm < 1.0): # % & & self.bond_dims(self.current_bond) <= 50;
dmerge = dmerge / gnorm
dmerge = tensor_add(self.merged_tensor, self.merged_idx,
self.learning_rate * dmerge, dpsi_idx)
return dmerge
def contraction_updat_twosite2(self):
if self.going_righ == 0:
if self.current_site == self.n - 2:
merg, idm = contra(self.tensors[self.n - 1],
self.order[self.n - 1],
self.tensors[self.n - 1],
self.order_left[self.n - 1])
# self.contraction_z[self.current_site+1]=merg
merg_psi, idm_psi = contra(self.image_tensors[self.n - 1],
self.order_right[self.n - 1],
self.tensors[self.n - 1],
self.order_left[self.n - 1])
else:
merg, idm = contra(
self.tensors[self.current_site + 1],
self.order_left[self.current_site + 1],
self.contraction_z[self.current_site + 2][0],
self.contraction_z[self.current_site + 2][1])
merg, idm = contra(merg, idm,
self.tensors[self.current_site + 1],
self.order[self.current_site + 1])
merg_psi, idm_psi = contra(
self.tensors[self.current_site + 1],
self.order_left[self.current_site + 1],
self.contraction_psi2[self.current_site + 2][0],
self.contraction_psi2[self.current_site + 2][1])
merg_psi, idm_psi = contra(
merg_psi, idm_psi,
self.image_tensors[self.current_site + 1],
self.order_right[self.current_site + 1])
self.contraction_psi2[self.current_site +
1] = ([merg_psi, idm_psi])
self.contraction_z[self.current_site + 1] = ([merg, idm])
if self.going_righ == 1:
if self.current_site == 0:
merg, idm = contra(self.tensors[0], self.order[0],
self.tensors[0], self.order_left[0])
# self.contraction_z[self.current_site+1]=merg
merg_psi, idm_psi = contra(self.image_tensors[0],
self.order_right[0],
self.tensors[0], self.order_left[0])
else:
merg, idm = contra(
self.tensors[self.current_site],
self.order_left[self.current_site],
self.contraction_z[self.current_site - 1][0],
self.contraction_z[self.current_site - 1][1])
merg, idm = contra(merg, idm, self.tensors[self.current_site],
self.order[self.current_site])
merg_psi, idm_psi = contra(
self.tensors[self.current_site],
self.order_left[self.current_site],
self.contraction_psi2[self.current_site - 1][0],
self.contraction_psi2[self.current_site - 1][1])
merg_psi, idm_psi = contra(
merg_psi, idm_psi, self.image_tensors[self.current_site],
self.order_right[self.current_site])
self.contraction_psi2[self.current_site] = ([
merg_psi, idm_psi
])
self.contraction_z[self.current_site] = ([merg, idm])
