def calc_loss(self, sys_ys, ref_ys, dists):
loss = wrapper.make_var([[0.0]])
sys_Tscore = wrapper.make_var([[0.0]])
ref_Tscore = wrapper.make_var([[0.0]])
sys_Tscore, sys_vecs = self.calc_trans_score(
sys_ys) #chainer.Variable, 1hotvecがconcateされたもの
sys_matrix = wrapper.make_var([sys_vecs])
ref_Tscore, ref_vecs = self.calc_trans_score(
ref_ys) #chainer.Variable, 1hotvec
ref_matrix = wrapper.make_var([ref_vecs])
dists_matrix = functions.concat(tuple(dists))
#異なるラベル数のカウント
diff_cnt = wrapper.make_var([[0.0]])
for sys_y, ref_y in zip(sys_ys, ref_ys):
if sys_y != ref_y:
diff_cnt += wrapper.make_var([[1.0]])
#max 0
loss = functions.matmul(sys_matrix, dists_matrix, transb=True) + sys_Tscore\
- functions.matmul(ref_matrix, dists_matrix, transb=True) - ref_Tscore\
+ self.__eta * diff_cnt
"""
debug
print("sys_score trans : ", wrapper.get_data(functions.matmul(sys_matrix, dists_matrix, transb=True)), wrapper.get_data(sys_Tscore))
print("ref_score trans : ", wrapper.get_data(functions.matmul(ref_matrix, dists_matrix, transb=True)), wrapper.get_data(ref_Tscore))
print("diff_cnt penal : ",wrapper.get_data(diff_cnt), wrapper.get_data(self.__eta * diff_cnt))
"""
return loss
def calc_loss(self, sys_ys, ref_ys, dists):
loss = wrapper.make_var([[0.0]])
sys_Tscore = wrapper.make_var([[0.0]])
ref_Tscore = wrapper.make_var([[0.0]])
sys_Tscore, sys_vecs = self.calc_trans_score(sys_ys) #chainer.Variable, 1hotvecがconcateされたもの
sys_matrix = wrapper.make_var([sys_vecs])
ref_Tscore, ref_vecs = self.calc_trans_score(ref_ys) #chainer.Variable, 1hotvec
ref_matrix = wrapper.make_var([ref_vecs])
dists_matrix = functions.concat(tuple(dists))
#異なるラベル数のカウント
diff_cnt = wrapper.make_var([[0.0]])
for sys_y, ref_y in zip(sys_ys, ref_ys):
if sys_y != ref_y:
diff_cnt += wrapper.make_var([[1.0]])
#max 0
loss = functions.matmul(sys_matrix, dists_matrix, transb=True) + sys_Tscore\
- functions.matmul(ref_matrix, dists_matrix, transb=True) - ref_Tscore\
+ self.__eta * diff_cnt
"""
debug
print("sys_score trans : ", wrapper.get_data(functions.matmul(sys_matrix, dists_matrix, transb=True)), wrapper.get_data(sys_Tscore))
print("ref_score trans : ", wrapper.get_data(functions.matmul(ref_matrix, dists_matrix, transb=True)), wrapper.get_data(ref_Tscore))
print("diff_cnt penal : ",wrapper.get_data(diff_cnt), wrapper.get_data(self.__eta * diff_cnt))
"""
return loss
def __forward(self, is_training, text):
m = self.__model
tanh = functions.tanh
lstm = functions.lstm
letters, labels = self.__make_input(is_training, text)
n_letters = len(letters)
accum_loss = wrapper.zeros(()) if is_training else None
hidden_zeros = wrapper.zeros((1, self.__n_hidden))
# embedding
list_e = []
for i in range(n_letters):
s_x = wrapper.make_var([letters[i]], dtype=np.int32)
list_e.append(tanh(m.w_xe(s_x)))
# forward encoding
s_a = hidden_zeros
c = hidden_zeros
list_a = []
for i in range(n_letters):
c, s_a = lstm(c, m.w_ea(list_e[i]) + m.w_aa(s_a))
list_a.append(s_a)
# backward encoding
s_b = hidden_zeros
c = hidden_zeros
list_b = []
for i in reversed(range(n_letters)):
c, s_b = lstm(c, m.w_eb(list_e[i]) + m.w_bb(s_b))
list_b.append(s_b)
# segmentation
scores = []
for i in range(n_letters - 1):
s_y = tanh(m.w_ay1(list_a[i]) + m.w_by1(list_b[i]) + m.w_ay2(list_a[i + 1]) + m.w_by2(list_b[i + 1]))
scores.append(float(wrapper.get_data(s_y)))
if is_training:
s_t = wrapper.make_var([[labels[i]]])
accum_loss += functions.mean_squared_error(s_y, s_t)
return scores, accum_loss
def __forward(self, is_training, text):
m = self.__model
tanh = functions.tanh
letters, labels = self.__make_input(is_training, text)
scores = []
accum_loss = wrapper.zeros(()) if is_training else None
for n in range(len(letters) - 2 * self.__n_context + 1):
s_hu = wrapper.zeros((1, self.__n_hidden))
for k in range(2 * self.__n_context):
wid = k * len(self.__vocab) + letters[n + k]
s_x = wrapper.make_var([wid], dtype=np.int32)
s_hu += m.w_xh(s_x)
s_hv = tanh(s_hu)
s_y = tanh(m.w_hy(s_hv))
scores.append(float(wrapper.get_data(s_y)))
if is_training:
s_t = wrapper.make_var([[labels[n]]])
accum_loss += functions.mean_squared_error(s_y, s_t)
return scores, accum_loss
def __forward(self, is_training, text):
m = self.__model
tanh = functions.tanh
letters, labels = self.__make_input(is_training, text)
scores = []
accum_loss = wrapper.zeros(()) if is_training else None
for n in range(len(letters) - 2 * self.__n_context + 1):
s_hu = wrapper.zeros((1, self.__n_hidden))
for k in range(2 * self.__n_context):
wid = k * len(self.__vocab) + letters[n + k]
s_x = wrapper.make_var([wid], dtype=np.int32)
s_hu += m.w_xh(s_x)
s_hv = tanh(s_hu)
s_y = tanh(m.w_hy(s_hv))
scores.append(float(wrapper.get_data(s_y)))
if is_training:
s_t = wrapper.make_var([[labels[n]]])
accum_loss += functions.mean_squared_error(s_y, s_t)
return scores, accum_loss
def calc_trans_score(self, labels):
trans_score = wrapper.make_var([[0.0]])
labels_vec = list()
pre_label = None
for label in labels:
if pre_label != None:
"""
debug
print("pre_label : ",pre_label)
print("label : ",label)
"""
trans_score += wrapper.get_data(self.get_trans_prob(pre_label))[0][label] #trans_score(pre_label → 0,1)をsoftmaxをかけてスコアリング
#1hot CONCATE vector
for i in range(self.__n_labels):
if i == label:
labels_vec.append(1)
else:
labels_vec.append(0)
pre_label = label
return trans_score, labels_vec
def calc_trans_score(self, labels):
trans_score = wrapper.make_var([[0.0]])
labels_vec = list()
pre_label = None
for label in labels:
if pre_label != None:
"""
debug
print("pre_label : ",pre_label)
print("label : ",label)
"""
trans_score += wrapper.get_data(
self.get_trans_prob(pre_label))[0][
label] #trans_score(pre_label → 0,1)をsoftmaxをかけてスコアリング
#1hot CONCATE vector
for i in range(self.__n_labels):
if i == label:
labels_vec.append(1)
else:
labels_vec.append(0)
pre_label = label
return trans_score, labels_vec
def __forward(self, is_training, text):
m = self.__model
tanh = functions.tanh
softmax = functions.softmax
letters, labels = self.__make_input(is_training, text) #文字列(スペースなし), 単語境界ラベル(-1,+1)
dists = []
accum_loss = wrapper.make_var([[0.0]]) if is_training else None
"""
debug
print("length, letters : ",len(letters), letters)
print("len(self.__vocab) : ",len(self.__vocab))
"""
for n in range(len(letters) - 2 * self.__n_context + 1):
"""
debug
print("n : ",n)
"""
s_hu = wrapper.zeros((1, self.__n_hidden))
for k in range(2 * self.__n_context):
wid = k * len(self.__vocab) + letters[n + k]
s_x = wrapper.make_var([wid], dtype=np.int32)
s_hu += m.w_xh(s_x)
"""
debug
print("k : ",k)
print("wid : ",wid)
"""
s_hv = tanh(s_hu)
s_y = softmax(m.w_hy(s_hv))
"""
debug
print("s_y : ",s_y.data)
"""
dists.append(s_y)
"""
debug
for i,dist in enumerate(dists):
print("dist_", i,": ", wrapper.get_data(dist))
self.check_trans()
"""
sys_ys, end_score = self.viterbi(dists)
"""
debug
print("end_score : ", end_score)
"""
if is_training:
"""
debug
print("labels : ", labels)
print("sys_ys : ", sys_ys)
"""
loss = self.calc_loss(sys_ys, labels, dists)
"""
debug
print("loss : ", wrapper.get_data(loss))
"""
if float(wrapper.get_data(loss)) > 0:
accum_loss += loss
return sys_ys, accum_loss
def __forward(self, is_training, text):
m = self.__model
tanh = functions.tanh
softmax = functions.softmax
letters, labels = self.__make_input(is_training,
text) #文字列(スペースなし), 単語境界ラベル(-1,+1)
dists = []
accum_loss = wrapper.make_var([[0.0]]) if is_training else None
"""
debug
print("length, letters : ",len(letters), letters)
print("len(self.__vocab) : ",len(self.__vocab))
"""
for n in range(len(letters) - 2 * self.__n_context + 1):
"""
debug
print("n : ",n)
"""
s_hu = wrapper.zeros((1, self.__n_hidden))
for k in range(2 * self.__n_context):
wid = k * len(self.__vocab) + letters[n + k]
s_x = wrapper.make_var([wid], dtype=np.int32)
s_hu += m.w_xh(s_x)
"""
debug
print("k : ",k)
print("wid : ",wid)
"""
s_hv = tanh(s_hu)
s_y = softmax(m.w_hy(s_hv))
"""
debug
print("s_y : ",s_y.data)
"""
dists.append(s_y)
"""
debug
for i,dist in enumerate(dists):
print("dist_", i,": ", wrapper.get_data(dist))
self.check_trans()
"""
sys_ys, end_score = self.viterbi(dists)
"""
debug
print("end_score : ", end_score)
"""
if is_training:
"""
debug
print("labels : ", labels)
print("sys_ys : ", sys_ys)
"""
loss = self.calc_loss(sys_ys, labels, dists)
"""
debug
print("loss : ", wrapper.get_data(loss))
"""
if float(wrapper.get_data(loss)) > 0:
accum_loss += loss
return sys_ys, accum_loss
