def calculate_score(feature, count):
feature_score = self.dense_layer(
dn.concatenate(
[span_feature,
dn.inputTensor(np.array(feature))]))
if self.options.use_count:
count_score = self.count_scale_2.expr() * dn.log(
dn.abs(self.count_scale.expr() * count + 1))
else:
count_score = 0
return feature_score + count_score
def backward(self, word_vectors, label):
dy.renew_cg()
x = dy.inputTensor(word_vectors)
y = dy.inputTensor(label)
logit = self.build_graph(x)
# q表示对正样本的加权
# 公式见https://www.tensorflow.org/api_docs/python/tf/nn/weighted_cross_entropy_with_logits
q = 15
l = 1 + (q - 1) * y
loss = (1 - y) * logit + l * (dy.log(1 + dy.exp(-dy.abs(logit))) +
dy.rectify(-logit))
res = loss.value()
loss.backward()
return res
def __getExpr(self, sentence, i, j):
if sentence[i].headfov is None:
sentence[i].headfov = concatenate([sentence[i].lstms[0], sentence[i].lstms[1]])
if sentence[j].modfov is None:
sentence[j].modfov = concatenate([sentence[j].lstms[0], sentence[j].lstms[1]])
_inputVector = concatenate(
[sentence[i].headfov, sentence[j].modfov, dynet.abs(sentence[i].headfov - sentence[j].modfov),
dynet.cmult(sentence[i].headfov, sentence[j].modfov)])
if self.hidden_units > 0:
output = self.outLayer.expr() * self.activation(
self.hid2Bias.expr() + self.hidLayer.expr() * self.activation(
_inputVector + self.hidBias.expr()))
else:
output = self.outLayer.expr() * self.activation(_inputVector + self.hidBias.expr())
return output
def __call__(self, premise, hypothesis, dropout=0.0):
precode = self._ssbilstm(premise)
hypcode = self._ssbilstm(hypothesis)
conc_pre_hyp = dy.concatenate([precode, hypcode])
dist_pre_hyp = dy.abs(precode - hypcode)
mult_pre_hyp = dy.cmult(precode, hypcode)
x = dy.concatenate([conc_pre_hyp, dist_pre_hyp, mult_pre_hyp])
expr = list()
for exp in self._params:
expr.append(dy.parameter(exp))
for i in range(0, len(expr) - 2, 2):
if '-relu' in sys.argv:
x = dy.rectify((expr[i] * x) + expr[i + 1])
else:
x = dy.tanh((expr[i] * x) + expr[i + 1])
if dropout != 0.0:
x = dy.dropout(x, dropout)
output = dy.softmax((expr[-2] * x) + expr[-1])
return output
def l2_normalize(x):
epsilon = np.finfo(float).eps * dy.ones(pred.dim()[0])
norm = dy.sqrt(dy.sum_elems(dy.square(x)))
sign = dy.cdiv(x, dy.bmax(dy.abs(x), epsilon))
return dy.cdiv(dy.cmult(sign, dy.bmax(dy.abs(x), epsilon)), dy.bmax(norm, epsilon[0]))
#wez y i usun sezonowosc i level
for i in range(n, len(df) - h):
inputs = y[i - n:i] #n okresy
curr_season = s[i - n:i]
inputs = dy.cdiv(inputs, l[i])
inputs = dy.cdiv(inputs, curr_season)
inputs = dy.log(inputs)
reseasonalize = s[i + 1] #poprzedni okres +1 krok
preds.append(dy.exp(fcstr(inputs)) * l[i] * reseasonalize)
outputs.append(y[i + 1]) #+1 krok
predictions = dy.concatenate(preds)
outputs = dy.concatenate(outputs)
#log_err = dy.mean_elems(dy.abs(dy.log(outputs)-dy.log(predictions)))
err = dy.mean_elems(dy.abs(outputs - predictions))
loss = err + level_loss
trainer = dy.SimpleSGDTrainer(m, learning_rate=0.25)
loss_value = loss.value()
print(seasonInit.npvalue())
for i in range(2000):
loss.backward()
trainer.update()
loss_value = loss.value(recalculate=True)
trainer.learning_rate *= 0.992 #0.99 lr decay
if i % 50 == 0 or i < 10:
print("the mae after step is:", err.value(recalculate=True))
def build_graph(pre_words, hy_words, holder):
fl1_init = holder.fwdRNN_layer1.initial_state()
bl1_init = holder.bwdRNN_layer1.initial_state()
fl2_init = holder.fwdRNN_layer2.initial_state()
bl2_init = holder.bwdRNN_layer2.initial_state()
fl3_init = holder.fwdRNN_layer3.initial_state()
bl3_init = holder.bwdRNN_layer3.initial_state()
pre_wembs = [get_word_rep(w, holder) for w in pre_words]
hy_wembs = [get_word_rep(w, holder) for w in hy_words]
pre_fws = fl1_init.transduce(pre_wembs)
pre_bws = bl1_init.transduce(reversed(pre_wembs))
pre_bi = [
dy.concatenate([word, f, b])
for word, f, b in zip(pre_wembs, pre_fws, reversed(pre_bws))
]
pre_fws2 = fl2_init.transduce(pre_bi)
pre_bws2 = bl2_init.transduce(reversed(pre_bi))
pre_b_tag = [
dy.concatenate([word, f1, b1, f2, b2])
for word, f1, b1, f2, b2 in zip(pre_wembs, pre_fws, reversed(pre_bws),
pre_fws2, reversed(pre_bws2))
]
pre_fws3 = fl3_init.transduce(pre_b_tag)
pre_bws3 = bl3_init.transduce(reversed(pre_b_tag))
pre_b_tagtag = [
dy.concatenate([f3, b3])
for f3, b3 in zip(pre_fws3, reversed(pre_bws3))
]
pre_v_elemets_size = len(pre_b_tagtag[0].npvalue())
pre_row_num = len(pre_b_tagtag)
pre_vecs_concat = dy.concatenate([v for v in pre_b_tagtag])
pre_mat = dy.reshape(pre_vecs_concat, (pre_v_elemets_size, pre_row_num))
pre_final = dy.concatenate([dy.kmax_pooling(v, 1, 0) for v in pre_mat])
hy_fws = fl1_init.transduce(hy_wembs)
hy_bws = bl1_init.transduce(reversed(hy_wembs))
hy_bi = [
dy.concatenate([word, f, b])
for word, f, b in zip(hy_wembs, hy_fws, reversed(hy_bws))
]
hy_fws2 = fl2_init.transduce(hy_bi)
hy_bws2 = bl2_init.transduce(reversed(hy_bi))
hy_b_tag = [
dy.concatenate([word, f1, b1, f2, b2]) for word, f1, b1, f2, b2 in zip(
hy_wembs, hy_fws, reversed(hy_bws), hy_fws2, reversed(hy_bws2))
]
hy_fws3 = fl3_init.transduce(hy_b_tag)
hy_bws3 = bl3_init.transduce(reversed(hy_b_tag))
hy_b_tagtag = [
dy.concatenate([f3, b3]) for f3, b3 in zip(hy_fws3, reversed(hy_bws3))
]
hy_v_elemets_size = len(hy_b_tagtag[0].npvalue())
hy_row_num = len(hy_b_tagtag)
hy_vecs_concat = dy.concatenate([v for v in hy_b_tagtag])
hy_mat = dy.reshape(hy_vecs_concat, (hy_v_elemets_size, hy_row_num))
hy_final = dy.concatenate([dy.kmax_pooling(v, 1, 0) for v in hy_mat])
final = dy.concatenate([
pre_final, hy_final,
dy.abs(pre_final - hy_final),
dy.cmult(pre_final, hy_final)
])
W1 = dy.parameter(holder.W1)
b1 = dy.parameter(holder.b1)
W2 = dy.parameter(holder.W2)
b2 = dy.parameter(holder.b2)
mid = dy.rectify(W1 * final + b1)
return W2 * mid + b2