def attention_3d_block(input_shapeR, input_shapeQ):
relation_maxlen = gl.get_relation_maxlen()
LSTM_DIM=gl.get_LSTM_DIM()
# inputs.shape = (batch_size, time_steps, seq_len)
print("input_shapeR: ")
print(K.int_shape(input_shapeR))
print("input_shapeQ: ")
print(K.int_shape(input_shapeQ))
mid = Dense(2*LSTM_DIM,name="att_dense",kernel_regularizer=regularizers.l2(0.01))(input_shapeR)
print("mid: ")
print(K.int_shape(mid))
print("rq: ")
rq = Permute((2, 1))(input_shapeQ)
print(K.int_shape(rq))
#a = K.batch_dot(mid,rq,axes=[2,1])
a = Dot(axes=[2,2])([mid,input_shapeQ])
# a = K.batch_dot(a,mid,axes=2)
a = Activation('softmax')(a)
##rtt =Permute((2, 1))(input_shapeR)
# x.shape = (batch_size, seq_len, time_steps)
print("a: ")
print(K.int_shape(a))
outputs = Dot(axes=[1, 1])([input_shapeR, a])
outputs = Permute((2, 1))(outputs)
print("outputs: ")
print(K.int_shape(outputs))
return outputs
gl.set_NUM_FILTERS(150)
gl.set_filter_sizes([1, 3, 5])
NUM_FILTERS = gl.get_NUM_FILTERS()
filter_sizes = gl.get_filter_sizes()
#预处理
ques_train, rela_train, label_train, ques_test, rela_test, label_test, wd_idx = preprocess(
train_rela_files, train_ques_file, train_label_file, test_rela_files,
test_ques_file, test_label_file)
embedding_index = loadEmbeddingsIndex(preprocessWordVector_path,
preprocessWordVector_files)
embedding_matrix = generateWord2VectorMatrix(embedding_index, wd_idx)
#model
relation_maxlen = gl.get_relation_maxlen()
ques_maxlen = gl.get_ques_maxlen()
tweet_relation = Input(shape=(relation_maxlen, ))
tweet_ques = Input(shape=(ques_maxlen, ))
relation_embedding_layer = Embedding(len(wd_idx) + 1,
EMBEDDING_DIM,
input_length=relation_maxlen,
weights=[embedding_matrix],
trainable=True)(tweet_relation)
relation_conv1 = Conv1D(128, 3, activation='tanh')(relation_embedding_layer)
relation_drop_1 = Dropout(0.2)(relation_conv1)
relation_max_1 = MaxPooling1D(relation_maxlen - 3 + 1)(relation_drop_1)
relation_conv2 = Conv1D(128, 1, activation='tanh')(relation_max_1)
relation_drop_2 = Dropout(0.2)(relation_conv2)