def __init__(self):
dt = DataTools()
self.ys, self.tokenMatrix, self.positionMatrix1, self.positionMatrix2, self.sdpMatrix = dt.get_data("pkl/test.pkl.gz")
class KejModel(object):
WORD_DIM = 250
CLASS_DIM = 30
NUM_CLASSES = 11
MAX_SENTENCE_LEN = 100
multiple = 1.0
dt = None
relationsMapping = {
'other': 0,
'locaA': 1,
'locAa': 2,
'med-ill': 3,
'ill-med': 4,
"clsaA": 5,
"clsAa": 6,
"w-c": 7,
"c-w": 8,
"cs-ef": 9,
"ef-cs": 10
}
idx2relation = {
0: 'other',
1: 'locaA',
2: 'locAa',
3: 'med-ill',
4: 'ill-med',
5: "clsaA",
6: "clsAa",
7: "w-c",
8: "c-w",
9: "cs-ef",
10: "ef-cs"
}
def __init__(self):
self.dt = DataTools()
def load_kej_model(self, modelpath):
model = load_model(modelpath)
print(model.summary())
return model
def build_new_model(self,
n_out,
max_sentence_len,
max_position,
embedding,
printflag=False):
print("build new model")
num_filter = 140
filter_length = 3
position_dims = 80
print(max_sentence_len)
words_input = Input(shape=(max_sentence_len, ),
dtype='int32',
name='words_input')
# words = Embedding(70000, self.WORD_DIM)(words_input)
words = embedding(words_input)
# if printflag:
# print(words)
# sdp_input = Input(shape=(max_sentence_len,), dtype='float32', name='sdp_input')
distance1_input = Input(shape=(max_sentence_len, ),
dtype='int32',
name='distance1_input')
distance1 = Embedding(max_position, position_dims)(distance1_input)
print(distance1)
distance2_input = Input(shape=(max_sentence_len, ),
dtype='int32',
name='distance2_input')
distance2 = Embedding(max_position, position_dims)(distance2_input)
print(distance2)
output = concatenate([words, distance1, distance2])
output = Convolution1D(filters=num_filter,
kernel_size=filter_length,
padding='same',
activation='relu',
strides=1)(output)
print(output)
output = GlobalMaxPooling1D()(output)
print(output)
output = Dropout(0.25)(output)
output = Dense(50, activation='relu')(output)
output = Dropout(0.25)(output)
output = Dense(n_out, activation='softmax')(output)
# model = Model(inputs = [sdp_input, words_input,distance1_input,distance2_input],outputs=[output])
model = Model(inputs=[words_input, distance1_input, distance2_input],
outputs=[output])
model.compile(loss='sparse_categorical_crossentropy',
optimizer='Adam',
metrics=['accuracy'])
return model
def train_model(self,
tosave,
savepath,
num_epoch=20,
batch_size=64,
load=True):
ys, tokenMatrix, positionMatrix1, positionMatrix2, sdpMatrix = self.dt.get_data(
'pkl/train.pkl.gz')
n_out = max(ys) + 1
max_sentence_len = tokenMatrix.shape[1]
print(max_sentence_len)
max_position = max(np.max(positionMatrix1),
np.max(positionMatrix2)) + 1
if load:
model = load_model('model/kej_model.h5')
else:
genmodel = GenModels.Word2Vec.load("w2vmodel/word2vec2.model")
embedding = genmodel.wv.get_keras_embedding(False)
model = self.build_new_model(n_out, max_sentence_len, max_position,
embedding)
print("Start training")
max_prec, max_rec, max_acc, max_f1 = 0, 0, 0, 0
# for epoch in range(num_epoch):
model.fit([tokenMatrix, positionMatrix1, positionMatrix2],
ys,
batch_size=batch_size,
verbose=True,
epochs=num_epoch)
if tosave:
model.save(savepath)
print("保存成功")
def predict_classes(self, prediction):
return prediction.argmax(axis=-1)
def model_predict(self, testpkl_path, model_path='model/kej_model.h5'):
ys, tokenMatrix, positionMatrix1, positionMatrix2, sdpMatrix = self.dt.get_data(
testpkl_path)
model = load_model(model_path)
pred_test = self.predict_classes(
model.predict([tokenMatrix, positionMatrix1, positionMatrix2],
verbose=False))
acc = np.sum(pred_test == ys) / float(len(ys))
print("测试样例数量: " + str(len(ys)))
print("准确率: " + str(acc))
print("错误样例：")
wrong_outputs = []
for i in range(len(ys)):
if pred_test[i] != ys[i]:
wrong_outputs.append(
str(i) + " " + self.idx2relation[ys[i]] + " " +
self.idx2relation[pred_test[i]])
for item in wrong_outputs:
print(item)
def model_predict_one(self, model, relationidx, positionMatrix1,
positionMatrix2, tokenMatrix):
pred_test = self.predict_classes(
model.predict([tokenMatrix, positionMatrix1, positionMatrix2],
verbose=False))
print(self.idx2relation[pred_test[0]])
def __init__(self):
dt = DataTools()
self.ys, self.tokenMatrix, self.positionMatrix1, self.positionMatrix2, self.sdpMatrix = dt.get_data(
"./pkl/final_test.pkl.gz")
print(self.ys.shape[0])
class KejModel(object):
WORD_DIM = 250
CLASS_DIM = 30
NUM_CLASSES = 11
MAX_SENTENCE_LEN = 100
multiple = 1.0
dt = None
relationsMapping = {
'other': 0,
'locaA': 1,
'locAa': 2,
'med-ill': 3,
'ill-med': 4,
"clsaA": 5,
"clsAa": 6,
"w-c": 7,
"c-w": 8,
"cs-ef": 9,
"ef-cs": 10
}
idx2relation = {
0: 'other',
1: 'locaA',
2: 'locAa',
3: 'med-ill',
4: 'ill-med',
5: "clsaA",
6: "clsAa",
7: "w-c",
8: "c-w",
9: "cs-ef",
10: "ef-cs"
}
def __init__(self):
self.dt = DataTools()
def load_kej_model(self, modelpath):
return load_model(modelpath)
def build_new_model(self,
n_out,
max_sentence_len,
max_position,
embedding,
printflag=False):
print("build new model")
num_filter = 140
filter_length = 3
position_dims = 80
print(max_sentence_len)
words_input = Input(shape=(max_sentence_len, ),
dtype='int32',
name='words_input')
# words = Embedding(70000, self.WORD_DIM)(words_input)
words = embedding(words_input)
# if printflag:
# print(words)
sdp_input = Input(shape=(max_sentence_len, ),
dtype='float32',
name='sdp_input')
print(sdp_input)
# K.variable(K.random_uniform(self.))
init_att = math.sqrt(6.0 / (self.WORD_DIM + self.CLASS_DIM))
U = K.variable(K.random_uniform([self.WORD_DIM, self.CLASS_DIM],
minval=-init_att,
maxval=init_att,
dtype=tf.float32),
name="U")
classes_matrix = K.variable(K.random_uniform(
[self.CLASS_DIM, self.NUM_CLASSES], dtype=tf.float32),
name="classmatrix")
G = tf.matmul(tf.reshape(words, [-1, self.WORD_DIM]), U, name="G")
print(G)
G = tf.reshape(tf.matmul(G, classes_matrix),
[-1, self.MAX_SENTENCE_LEN, self.NUM_CLASSES],
name="G")
print(G)
init_m = math.sqrt(6.0 / (self.MAX_SENTENCE_LEN + self.NUM_CLASSES))
M = tf.Variable(tf.random_uniform(
[self.MAX_SENTENCE_LEN, self.NUM_CLASSES],
minval=-init_m,
maxval=init_m,
dtype=tf.float32),
name="M")
alpha = tf.expand_dims(tf.matmul(sdp_input, M), axis=1, name="alpha")
print(alpha)
alpha = tf.matmul(alpha, tf.transpose(G, [0, 2, 1]), name="alpha")
print(alpha)
alpha = tf.nn.l2_normalize(tf.squeeze(alpha, axis=1),
axis=-1,
name="alpha")
print(alpha)
alpha_v = tf.add(sdp_input,
tf.scalar_mul(self.multiple, alpha),
name="alpha_v")
alpha = tf.matrix_diag(alpha_v, name="alpha")
print(alpha)
weighted_data = tf.matmul(alpha, words, name="weighted_data")
print(weighted_data)
distance1_input = Input(shape=(max_sentence_len, ),
dtype='int32',
name='distance1_input')
distance1 = Embedding(max_position, position_dims)(distance1_input)
print(distance1)
distance2_input = Input(shape=(max_sentence_len, ),
dtype='int32',
name='distance2_input')
distance2 = Embedding(max_position, position_dims)(distance2_input)
print(distance2)
# my_concat = Lambda(lambda x: K.concatenate([x[0], x[1],x[2],x[3]], axis=-1))
# output = my_concat([words,distance1,distance2,alpha])
outputtemp = tf.concat([words, distance1, distance2], axis=2)
# outputtemp = concatenate([words,distance1,distance2])
print(outputtemp)
output = Convolution1D(filters=num_filter,
kernel_size=filter_length,
padding='same',
activation='relu',
strides=1)(outputtemp)
print(output)
output = GlobalMaxPooling1D()(output)
print(output)
output = Dropout(0.25)(output)
output = Dense(50, activation='relu')(output)
output = Dropout(0.25)(output)
output = Dense(n_out, activation='softmax')(output)
model = Model(
inputs=[sdp_input, words_input, distance1_input, distance2_input],
outputs=[output])
model.compile(loss='sparse_categorical_crossentropy',
optimizer='Adam',
metrics=['accuracy'])
return model
def train_model(self,
tosave,
savepath,
num_epoch=20,
batch_size=64,
load=True):
ys, tokenMatrix, positionMatrix1, positionMatrix2, sdpMatrix = self.dt.get_data(
'pkl/train2.pkl.gz')
n_out = max(ys) + 1
max_sentence_len = tokenMatrix.shape[1]
print(max_sentence_len)
max_position = max(np.max(positionMatrix1),
np.max(positionMatrix2)) + 1
if load:
model = load_model('model/kej_model.h5')
else:
genmodel = GenModels.Word2Vec.load("w2vmodel/word2vec2.model")
embedding = genmodel.wv.get_keras_embedding(False)
model = self.build_new_model(n_out, max_sentence_len, max_position,
embedding)
print("Start training")
max_prec, max_rec, max_acc, max_f1 = 0, 0, 0, 0
# for epoch in range(num_epoch):
model.fit([sdpMatrix, tokenMatrix, positionMatrix1, positionMatrix2],
ys,
batch_size=batch_size,
verbose=True,
epochs=num_epoch)
if tosave:
model.save(savepath)
print("保存成功")
def predict_classes(self, prediction):
return prediction.argmax(axis=-1)
def model_predict(self, testpkl_path, model_path='model/kej_model.h5'):
ys, tokenMatrix, positionMatrix1, positionMatrix2 = self.dt.get_data(
testpkl_path)
model = load_model(model_path)
pred_test = self.predict_classes(
model.predict([tokenMatrix, positionMatrix1, positionMatrix2],
verbose=False))
acc = np.sum(pred_test == ys) / float(len(ys))
print("测试样例数量: " + str(len(ys)))
print("准确率: " + str(acc))
print("错误样例：")
wrong_outputs = []
for i in range(len(ys)):
if pred_test[i] != ys[i]:
wrong_outputs.append(
str(i) + " " + self.idx2relation[ys[i]] + " " +
self.idx2relation[pred_test[i]])
for item in wrong_outputs:
print(item)
def model_predict_one(self, model, relationidx, positionMatrix1,
positionMatrix2, tokenMatrix):
pred_test = self.predict_classes(
model.predict([tokenMatrix, positionMatrix1, positionMatrix2],
verbose=False))
print(self.idx2relation[pred_test[0]])