def __init__(self, n_voc, trainset, testset,dataname, classes, prefix):
if prefix != None:
prefix += '/'
self.trainset = trainset
self.testset = testset
docs = T.imatrix()
label = T.ivector()
length = T.fvector()
sentencenum = T.fvector()
wordmask = T.fmatrix()
sentencemask = T.fmatrix()
maxsentencenum = T.iscalar()
isTrain = T.iscalar()
rng = numpy.random
layers = []
layers.append(EmbLayer(rng, docs, n_voc, 200, 'emblayer', dataname, prefix))
layers.append(LSTMLayer(rng, layers[-1].output, wordmask, 200, 200, 'wordlstmlayer', prefix))
layers.append(MeanPoolLayer(layers[-1].output, length))
layers.append(SentenceSortLayer(layers[-1].output,maxsentencenum))
layers.append(LSTMLayer(rng, layers[-1].output, sentencemask, 200, 200, 'sentencelstmlayer', prefix))
layers.append(MeanPoolLayer(layers[-1].output, sentencenum))
layers.append(HiddenLayer(rng, layers[-1].output, 200, 200, 'fulllayer', prefix))
layers.append(HiddenLayer(rng, layers[-1].output, 200, int(classes), 'softmaxlayer', prefix, activation=T.nnet.softmax))
self.layers = layers
cost = -T.mean(T.log(layers[-1].output)[T.arange(label.shape[0]), label], acc_dtype='float32')
correct = T.sum(T.eq(T.argmax(layers[-1].output, axis=1), label), acc_dtype='int32')
err = T.argmax(layers[-1].output, axis=1) - label
mse = T.sum(err * err)
params = []
for layer in layers:
params += layer.params
L2_rate = numpy.float32(1e-5)
for param in params[1:]:
cost += T.sum(L2_rate * (param * param), acc_dtype='float32')
gparams = [T.grad(cost, param) for param in params]
updates = AdaUpdates(params, gparams, 0.95, 1e-6)
self.train_model = theano.function(
inputs=[docs, label,length,sentencenum,wordmask,sentencemask,maxsentencenum],
outputs=cost,
updates=updates,
)
self.test_model = theano.function(
inputs=[docs, label,length,sentencenum,wordmask,sentencemask,maxsentencenum],
outputs=[correct, mse],
)
def __init__(self, n_voc, trainset, testset, em_path, classes, prefix):
if prefix is not None:
prefix += '/'
self.trainset = trainset
self.testset = testset
docs = T.imatrix()
label = T.ivector()
length = T.fvector()
sentencenum = T.fvector()
wordmask = T.fmatrix()
sentencemask = T.fmatrix()
maxsentencenum = T.iscalar()
isTrain = T.iscalar()
rng = numpy.random
layers = []
layers.append(
EmbLayer(em_path, rng, docs, n_voc, 200, "emblayer", prefix))
layers.append(
LSTMLayer(rng, layers[-1].output, wordmask, 200, 200,
'wordlstmlayer', prefix))
layers.append(MeanPoolLayer(layers[-1].output, length))
layers.append(SentenceSortLayer(layers[-1].output, maxsentencenum))
layers.append(
LSTMLayer(rng, layers[-1].output, sentencemask, 200, 200,
'sentencelstmlayer', prefix))
layers.append(MeanPoolLayer(layers[-1].output, sentencenum))
layers.append(
HiddenLayer(rng, layers[-1].output, 200, 200, 'fulllayer', prefix))
# 保存这层的输出 是numpy.ndarray类型
feature_vector = layers[-1].output
layers.append(
HiddenLayer(rng,
layers[-1].output,
200,
int(classes),
'softmaxlayer',
prefix,
activation=T.nnet.softmax))
# softmax这层输出是 n行2维向量（2分类），表示每个句子属于每个分类的概率，之和为1
predict_probability = layers[-1].output
self.layers = layers
cost = -T.mean(T.log(layers[-1].output)[T.arange(label.shape[0]),
label],
acc_dtype='float32')
correct = T.sum(T.eq(T.argmax(layers[-1].output, axis=1), label),
acc_dtype='int32')
err = T.argmax(layers[-1].output, axis=1) - label
mse = T.sum(err * err)
# 返回预测的各个文本的类别（0和1）
predict_label = T.argmax(layers[-1].output, axis=1)
params = []
for layer in layers:
params += layer.params
L2_rate = numpy.float32(1e-5)
for param in params[1:]:
cost += T.sum(L2_rate * (param * param), acc_dtype='float32')
gparams = [T.grad(cost, param) for param in params]
updates = AdaUpdates(params, gparams, 0.95, 1e-6)
self.train_model = theano.function(inputs=[
docs, label, length, sentencenum, wordmask, sentencemask,
maxsentencenum
],
outputs=[cost, feature_vector],
updates=updates,
mode='FAST_RUN')
self.test_model = theano.function(inputs=[
docs, label, length, sentencenum, wordmask, sentencemask,
maxsentencenum
],
outputs=[
correct, mse, predict_label,
feature_vector,
predict_probability
],
mode='FAST_RUN')
def __init__(self, n_voc, n_usr, n_prd, trainset, testset, dataname, classes, prefix):
if prefix != None:
prefix += '/'
self.trainset = trainset
self.testset = testset
docs = T.imatrix()
label = T.ivector()
usr = T.ivector()
prd = T.ivector()
wordmask = T.fmatrix()
sentencemask = T.fmatrix()
maxsentencenum = T.iscalar()
isTrain = T.iscalar()
rng = numpy.random
layers = []
docsemb = EmbLayer(rng, docs, n_voc, 200, 'emblayer', dataname, prefix)
Uemb = UsrEmbLayer(rng, n_usr, 200, 'usremblayer', prefix)
Pemb = PrdEmbLayer(rng, n_prd, 200, 'prdemblayer', prefix)
layers.append(docsemb)
layers.append(Uemb)
layers.append(Pemb)
layers.append(LSTMLayer(rng, docsemb.output, wordmask, 200, 200, 'wordlstmlayer', prefix))
uemb_sentence = GetuEmbLayer(usr, Uemb.output, maxsentencenum, 'uemb_sentence', prefix)
pemb_sentence = GetpEmbLayer(prd, Pemb.output, maxsentencenum, 'pemb_sentence', prefix)
layers.append(AttentionLayer(rng, layers[-1].output, uemb_sentence.output, pemb_sentence.output, wordmask, 200,200,200,200, 'wordattentionLayer', prefix))
layers.append(SentenceSortLayer(layers[-1].output, maxsentencenum, prefix))
layers.append(LSTMLayer(rng, layers[-1].output, sentencemask, 200, 200, 'sentencelstmlayer', prefix))
uemb_doc = GetuEmbLayer(usr, Uemb.output, maxsentencenum, 'uemb_doc', prefix)
pemb_doc = GetpEmbLayer(prd, Pemb.output, maxsentencenum, 'pemb_doc', prefix)
layers.append(AttentionLayer(rng, layers[-1].output, uemb_doc.output, pemb_doc.output, sentencemask, 200,200,200,200, 'sentenceattentionLayer', prefix))
layers.append(HiddenLayer(rng, layers[-1].output, 200, 200, 'fulllayer', prefix))
layers.append(HiddenLayer(rng, layers[-1].output, 200, int(classes), 'softmaxlayer', prefix, activation=T.nnet.softmax))
self.layers = layers
cost = -T.mean(T.log(layers[-1].output)[T.arange(label.shape[0]), label], acc_dtype='float32')
correct = T.sum(T.eq(T.argmax(layers[-1].output, axis=1), label), acc_dtype='int32')
err = T.argmax(layers[-1].output, axis=1) - label
mse = T.sum(err * err)
params = []
for layer in layers:
params += layer.params
L2_rate = numpy.float32(1e-5)
for param in params[3:]:
cost += T.sum(L2_rate * (param * param), acc_dtype='float32')
gparams = [T.grad(cost, param) for param in params]
updates = AdaUpdates(params, gparams, 0.95, 1e-6)
self.train_model = theano.function(
inputs=[docs, label, usr, prd, wordmask,sentencemask,maxsentencenum],
outputs=cost,
updates=updates,
)
self.test_model = theano.function(
inputs=[docs, label, usr, prd, wordmask,sentencemask,maxsentencenum],
outputs=[correct, mse],
)
def __init__(self, n_voc, trainset, testset, dataname, classes, prefix):
if prefix != None:
prefix += '/'
self.trainset = trainset
self.testset = testset
self.classes = int(classes)
docs = T.imatrix()
label = T.imatrix()
length = T.fvector()
wordmask = T.fmatrix()
sentencemask = T.fmatrix()
maxsentencenum = T.iscalar()
sentencenum = T.fvector()
isTrain = T.iscalar()
rng = numpy.random
# layers = []
# layers.append(EmbLayer(rng, docs, n_voc, 50, 'emblayer', dataname, prefix))
# layers.append(LSTMLayer(rng, layers[-1].output, wordmask, 50, 50, 'wordlstmlayer', prefix))
# layers.append(SimpleAttentionLayer(rng, layers[-1].output, wordmask,50, 50, 'wordattentionlayer', prefix))
# layers.append(SentenceSortLayer(layers[-1].output,maxsentencenum,prefix))
# layers.append(LSTMLayer(rng, layers[-1].output, sentencemask, 50, 50, 'sentencelstmlayer', prefix))
# layers.append(SimpleAttentionLayer(rng, layers[-1].output, sentencemask,50, 50, 'sentenceattentionlayer', prefix))
# layers.append(HiddenLayer(rng, layers[-1].output, 50, 50, 'fulllayer', prefix))
# layers.append(HiddenLayer(rng, layers[-1].output, 50, int(classes), 'softmaxlayer', prefix, activation=T.nnet.sigmoid))
# self.layers = layers
layers = []
layers.append(
EmbLayer(rng, docs, n_voc, 50, 'emblayer', dataname, prefix))
layers.append(
LSTMLayer(rng, layers[-1].output, wordmask, 50, 50,
'wordlstmlayer', prefix))
layers.append(MeanPoolLayer(layers[-1].output, length))
layers.append(SentenceSortLayer(layers[-1].output, maxsentencenum))
layers.append(
LSTMLayer(rng, layers[-1].output, sentencemask, 50, 50,
'sentencelstmlayer', prefix))
layers.append(MeanPoolLayer(layers[-1].output, sentencenum))
layers.append(
HiddenLayer(rng, layers[-1].output, 50, 50, 'fulllayer', prefix))
layers.append(
HiddenLayer(rng,
layers[-1].output,
50,
int(classes),
'softmaxlayer',
prefix,
activation=T.nnet.sigmoid))
self.layers = layers
predict = layers[-1].output
cost = T.nnet.binary_crossentropy(layers[-1].output, label).sum(1)
cost = cost.mean()
# modifu corrrect.
# predicted_value = ((layers[-1].output) >= EVALUTION_THRESHOLD_FOR_MULTI_LABEL).astype(int)
# predicted_value = predicted_value.astype(bool)
# true_value = label.astype(bool)
# equal = true_value == predicted_value
# match = np.sum(equal, axis=1) == np.size(equal, axis=1)
# # value 1 match_ratio
# exact_match_ratio = np.sum(match) / np.size(match)
# true_and_predict = np.sum(true_value & predicted_value, axis=1)
# true_or_predict = np.sum(true_value | predicted_value, axis=1)
# # value 2 accuracy
# accuracy = np.mean(true_and_predict / true_or_predict)
# # value 3 pression
# precison = np.mean(true_and_predict / (np.sum(predicted_value, axis=1) + 1e-9))
# # recall 4 recall
# recall = np.mean(true_and_predict / np.sum(true_value, axis=1))
# # f1_Measure
# F1_Measure = np.mean((true_and_predict * 2) / (np.sum(true_value, axis=1) + np.sum(predicted_value, axis=1)))
# # HammingLoss
# HammingLoss = np.mean(true_value ^ total_predicted_value)
# TP
# TP = np.sum(true_value & predicted_value,axis=0,dtype=np.int32)
# FP = np.sum((~true_value) & predicted_value,axis=0,dtype=np.int32)
# FN = np.sum(true_value & (~predicted_value),axis=0,dtype=np.int32)
# _P = np.sum(TP) / (np.sum(TP) + np.sum(FP) + 1e-9 )
# _R = np.sum(TP) / (np.sum(TP) + np.sum(FN) + 1e-9 )
# Micro_F1 = (2 * _P *_R) / (_P + _R)
# _P_t = TP / (TP + FP + 1e-9)
# _R_t = TP / (TP + FN + 1e-9)
# Macro_F1 = np.mean((2 * _P_t * _R_t) / (_P_t + _R_t + 1e-9))
#cost = -T.mean(T.log(layers[-1].output)[T.arange(label.shape[0]), label], acc_dtype='float32')
#modify this
#correct = T.sum(T.eq(T.argmax(layers[-1].output, axis=1), label), acc_dtype='int32')
#err = T.argmax(layers[-1].output, axis=1) - label
#mse = T.sum(err * err)
params = []
for layer in layers:
params += layer.params
L2_rate = numpy.float32(1e-5)
for param in params[1:]:
cost += T.sum(L2_rate * (param * param), acc_dtype='float32')
gparams = [T.grad(cost, param) for param in params]
updates = AdaUpdates(params, gparams, 0.95, 1e-6)
self.train_model = theano.function(
inputs=[
docs, label, length, sentencenum, wordmask, sentencemask,
maxsentencenum
],
outputs=cost,
updates=updates,
)
self.test_model = theano.function(
inputs=[
docs, length, sentencenum, wordmask, sentencemask,
maxsentencenum
],
outputs=predict,
)