def update_tensor_decision_mps_gradient_algo(self,
nt,
time_r=5,
threshold=0,
step=0.2):
self.update_tensor_decision_mps_svd(nt)
for t in range(0, time_r):
d_tensor = np.zeros(self.tensors[nt].shape)
for n in self.remaining_samples_train:
v1 = tm.absorb_vectors2tensors(
self.tensors[nt],
(self.v_ctr_train[n], self.vecsImages[n][:, nt]), (0, 1))
norm = np.linalg.norm(v1)
fid = self.fun_fidelity(v1 / norm)
fid_now = fid[self.classes.index(self.LabelNow[n])]
fid = [fid[nn] / fid_now for nn in range(0, self.num_classes)]
fid.pop(self.classes.index(self.LabelNow[n]))
if max(fid) > threshold:
tmp = np.kron(np.kron(self.v_ctr_train[n], self.vecsImages[n][:, nt]),
self.vLabel[self.classes.index(self.LabelNow[n])]) \
/ (fid_now * norm)
d_tensor += tmp.reshape(self.tensors[nt].shape)
d_tensor -= self.tensors[nt]
norm = np.linalg.norm(d_tensor.reshape(-1, ))
if norm > 1e-10:
d_tensor /= norm
self.tensors[nt] = self.tensors[nt] + step * d_tensor
self.tensors[nt] /= np.linalg.norm(self.tensors[nt].reshape(-1, ))
def update_tensor_decision_mps_svd_threshold_algo(self,
nt,
time_r=5,
threshold=0.9):
self.update_tensor_decision_mps_svd(nt)
env = 0
d0 = self.v_ctr_train[0].shape[0]
d1 = self.vecsImages[0][:, nt].shape[0]
for t in range(0, time_r):
for n in self.remaining_samples_train:
v1 = tm.absorb_vectors2tensors(
self.tensors[nt],
(self.v_ctr_train[n], self.vecsImages[n][:, nt]), (0, 1))
norm = np.linalg.norm(v1)
fid = self.fun_fidelity(v1 / norm)
fid_now = fid[self.classes.index(self.LabelNow[n])]
fid = [fid[nn] / fid_now for nn in range(0, self.num_classes)]
fid.pop(self.classes.index(self.LabelNow[n]))
if max(fid) > threshold:
env += np.kron(
np.kron(self.v_ctr_train[n], self.vecsImages[n][:,
nt]),
self.vLabel[self.classes.index(self.LabelNow[n])])
u, self.lm[nt], v = np.linalg.svd(
(env / np.linalg.norm(env.reshape(-1, ))).reshape(
d0 * d1, self.chi),
full_matrices=False)
self.tensors[nt] = u.dot(v).reshape([d0, d1, self.chi])
self.lm[nt] /= np.linalg.norm(self.lm[nt])
def fidelity_mps_image(self, mps_ref, ni):
# Calculate the fidelity between an MPS and one image
fid = 0
length = mps_ref.__len__()
v0 = np.ones((1, ))
image = self.vecsImages[ni]
for n in range(0, self.length):
v0 = tm.absorb_vectors2tensors(mps_ref[n], (v0, image[:, n]), (0, 1))
norm = np.linalg.norm(v0)
v0 /= norm
fid -= np.log(norm) / length
return fid
def update_v_ctr_test(self, nt):
for n in self.remaining_samples_test:
self.v_ctr_test[n] = tm.absorb_vectors2tensors(
self.tensors[nt], (self.v_ctr_test[n], self.vecsTest[:, nt, n]), (0, 1))
self.v_ctr_test[n] /= np.linalg.norm(self.v_ctr_test[n])
def update_v_ctr_train(self, nt):
for n in self.remaining_samples_train:
self.v_ctr_train[n] = tm.absorb_vectors2tensors(
self.tensors[nt], (self.v_ctr_train[n], self.vecsImages[n][:, nt]), (0, 1))
self.v_ctr_train[n] /= np.linalg.norm(self.v_ctr_train[n])
