#opt = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08)
#opt = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='categorical_crossentropy',
optimizer=opt,
metrics=['accuracy'])
#为每组实验记录F值得迭代
fpath = u'../data/fscore/Fscore-bils%s-ls%s-%s.txt' % (str(bils), str(ls),
label_mode)
mpath = u'../data/model/bils%s-ls%s-%s.h5' % (str(bils), str(ls), label_mode)
fpout = cs.open(fpath, 'w', 'utf-8')
for i in range(100):
model.fit(xtrain, ytrain, batch_size=128, epochs=1) #epochs练过程中数据将被“轮”多少次
#使用开发集的f来决定是否保存模型，实体F最大的模型和关系F最大的模型分别保存
y_predict = model.predict(xvaild, batch_size=128)
pre_label = ypre2label(y_predict)
pre_e, e_token = label2answer(pre_label, vaildtokens)
print(u'第%s次训练开发集的结果为' % (str(i)))
pe, re, fe = computeFe(gold_e_vaild, pre_e)
fpout.write(u'第%s次开发集F值分别为：%f\n' % (str(i + 1), fe))
if fe > maxf_e:
maxf_e = fe
save_model(mpath, model)
#每轮迭代都额外考察在测试集上的数据，但测试集不参与模型的选择
y_predict = model.predict(xtest, batch_size=128)
pre_label = ypre2label(y_predict)
pre_e, e_token = label2answer(pre_label, testtokens)
print(u'第%s次训练测试集的结果为' % (str(i)))
pe, re, fe = computeFe(gold_e, pre_e)
fpout.write(u'第%s次测试集F值分别为：%f\n\n' % (str(i + 1), fe))
fpout.close()
# cnn = Convolution1D(nb_filter=100, filter_length=3, activation='tanh')(cnn)
# flattened = Flatten()(cnn)
# dense = Dense(100, activation='tanh')(flattened)
#dense = Dense(50, activation='tanh')(dense)
predict = Dense(len(reptest.l2i_dic), activation='softmax')(bilstm)
model = Model(input=[word, distance_e1, distance_e2], output=predict)
opt = RMSprop(lr=0.001, rho=0.9, epsilon=1e-06)
model.compile(loss='categorical_crossentropy', optimizer=opt)
model.summary()
best_f = 0
for epoch in range(epoch_size):
print('running the epoch:', (epoch + 1))
model.fit([xtrain, pos1train, pos2train], ytrain, batch_size=256, epochs=1)
answer_array_d = model.predict([xtest, pos1test, pos2test], batch_size=256)
pre_label = ypre2label(answer_array_d)
pre_r = [[] for i in range(len(goldr))]
print(len(pre_r))
for i in range(len(pre_label)):
if pre_label[i] != 0:
label = reptest.i2l_dic[pre_label[i]] #类别
sen_index = indexmap_test[i] #句子索引
ans = epostest[i] + [label]
pre_r[sen_index].append(ans)
p, r, f, f_label = computeFr(goldr, pre_r)
if f > best_f:
print('New Best F-score')
best_f = f
save_model('../data/model/bils%s.h5' % (str(bils)), model)
# else:
# print ('Lower than before')