def fit_new_doc(self, a):
# 固定w，根据新a求h
avg = np.sqrt(a.mean() / self.k)
h = avg * self.rng.randn(self.k, 1).astype(np.float32)
np.abs(h, h)
# 设置初始误差
tolh = max(0.001, self.tol)
error_at_init = self.error_new_doc(a, h)
print(error_at_init)
self.error_array = [error_at_init]
at = a.transpose()
for i in range(self.max_iter):
# 固定W求H
h, _, iter_count = nmf.nlssubprob(a, self.w, h, tolh, 1000)
if iter_count == 1:
tolh *= 0.1
error = self.error_new_doc(a, h)
print(error)
self.error_array.append(error)
if i + 1 >= self.min_iter and \
np.abs(error - self.error_array[i]) / error_at_init <= self.tol:
print(f'达到收敛条件，在第{i}次迭代后提前结束')
break
return h
def fit_original_nmf_cjlin_nnls(self):
"""
计算原始的nmf，损失函数：
"""
tolw = max(0.001, self.tol)
tolh = tolw
error_at_init = self.error_original_nmf()
self.error_array = [error_at_init]
self.callback_set_nmf_training_progress(0, error_at_init)
at = self.a.T
for i in range(self.max_iter):
if i % 2 == 0:
# 固定H求W
self.w, _, iter_count = nmf.nlssubprob(at, self.h.T, self.w.T,
tolw, 1000)
self.w = self.w.T
if iter_count == 1:
tolw *= 0.1
else:
# 固定W求H
self.h, _, iter_count = nmf.nlssubprob(self.a, self.w, self.h,
tolh, 1000)
if iter_count == 1:
tolh *= 0.1
error = self.error_original_nmf()
print(error)
self.error_array.append(error)
self.callback_set_nmf_training_progress(i + 1, error)
if i + 1 >= self.min_iter and \
np.abs(error - self.error_array[i]) / error_at_init <= self.tol:
print(f'达到收敛条件，在第{i}次迭代后提前结束')
break
def fit_semi_supervised_nmf_cjlin_nnls(self):
"""
计算半监督的nmf，损失函数：
"""
tolw = max(0.001, self.tol)
tolh_list = [tolw] * self.n
error_at_init = self.error_semi_supervised_nmf()
self.error_array = [error_at_init]
self.callback_set_nmf_training_progress(0, error_at_init)
vstack_ht_mw = np.vstack((self.h.T, self.mw))
vstack_at_hstack_mw_wrt = np.vstack(
(self.a.T, np.dot(self.mw, self.wr.T)))
for i in range(self.max_iter):
if i % 2 == 0:
# 固定H求W
vstack_ht_mw[0:self.n, 0:self.k] = self.h.T
'''
vstack_ht_mw = np.vstack((self.h.T, self.mw))
vstack_at_hstack_mw_wrt = np.vstack(
(self.a.T, np.dot(self.mw, self.wr.T))
)
'''
self.w, _, iter_count = nmf.nlssubprob(vstack_at_hstack_mw_wrt,
vstack_ht_mw, self.w.T,
tolw, 1000)
self.w = self.w.T
if iter_count == 1:
tolw *= 0.1
print(f'迭代次数{iter_count}')
else:
# 固定W求H
for ci in range(self.n):
vstack_w_mhi_ik = np.vstack(
(self.w, self.mh[ci, ci] * self.ik))
vstack_aci_mhi_dhi_hrci = np.vstack((self.a[:, ci].reshape(
(self.m, 1)), (self.mh[ci, ci] * self.dh[ci, ci] *
self.hr[:, ci]).reshape((self.k, 1))))
hci, _, iter_count = nmf.nlssubprob(
vstack_aci_mhi_dhi_hrci, vstack_w_mhi_ik,
self.h[:, ci].reshape(
(self.k, 1)), tolh_list[ci], 1000)
self.h[:, ci] = hci.reshape(self.k)
if iter_count == 1:
tolh_list[ci] *= 0.1
# 更新DH
for j in range(self.n):
if self.mh[j, j] == 0:
self.dh[j, j] = 0
else:
hri = self.hr[:, j]
# 这里向量乘向量不需要reshape
# 因为numpy中，np.dot(v1, v2)
# 就相当于v1.T * v2 (v1∈(n, 1), v2∈(n, 1))
self.dh[j, j] = (np.dot(hri, self.h[:, j]) /
np.dot(hri, hri))
error = self.error_semi_supervised_nmf()
print(error)
self.error_array.append(error)
self.callback_set_nmf_training_progress(i + 1, error)
if i + 1 >= self.min_iter and \
np.abs(error - self.error_array[i]) / error_at_init <= self.tol:
print(f'达到收敛条件，在第{i}次迭代后提前结束')
break