今天在研读角色识别代码时，一时兴起想看看共指消解系统的summary，结果在summar打印的各层信息之间多了一堆警告，本来共指消解系统就是3个神经网络模型彼此连接，每隔一小段就会在网络层之间打出一堆WARNING，看着就难受，结果查了一下怎么关掉警告，百度了一下午，没找到能解决的方法，全是让设置os.environ的，而角色识别代码使用了keras.backend模块，所以使用os.environ根本就不起作用，还有一堆胡扯八道的，告诉版本有问题，后来发现好几个告警的都是tensorflow_backend.py文件，如下所示
WARNING:tensorflow:From /Users/yufei/git_code/information_extraction/character_identification/python/experiments/latest/model/coref.py:32:
The name tf.GPUOptions is deprecated. Please use tf.compat.v1.GPUOptions instead.
2019-09-29 20:54:10.659388: I tensorflow/compiler/xla/service/service.cc:168]
XLA service 0x7f8081c01700 initialized for platform Host (this does not guarantee that XLA will be used). Devices:
2019-09-29 20:54:10.659460: I tensorflow/compiler/xla/service/service.cc:176] StreamExecutor device (0): Host, Default Version
WARNING:tensorflow:From
/Users/yufei/anaconda3/envs/py36/lib/python3.6/site-packages/Keras-2.2.5-py3.6.egg/keras/backend/tensorflow_backend.py:4274:
The name tf.nn.max_pool is deprecated. Please use tf.nn.max_pool2d instead.
WARNING:tensorflow:From
/Users/yufei/anaconda3/envs/py36/lib/python3.6/site-packages/Keras-2.2.5-py3.6.egg/keras/backend/tensorflow_backend.py:3740:
calling dropout (from tensorflow.python.ops.nn_ops) with keep_prob is deprecated and will be removed in a future version.于是用Google搜了一下"tensorflow_backend.py warning"，在keras的github上找到了答案
The TensorFlow warnings when I'm trying to create a Keras model · Issue #8742 · keras-team/kerasgithub.com
下边给出角色识别中共指消解系统的模型构建代码，代码中将summary和模型图都输出了。注意添加下边的这行代码，表示关掉警告。
import 完整代码如下：
import tensorflow as tf
tf.logging.set_verbosity(tf.logging.ERROR)
pftdim = 6
mftdim = 72
eftdims = [(3, 50), (8, 50), (5, 50), (5, 50)]
nb_fltrs = 280
nb_efts = 4
# 设置GPU选项
gpu=None
gpu_num = 1
gpu_opts = tf.GPUOptions(allow_growth=True, visible_device_list=','.join(map(str, gpu) if gpu else []))
# 实例化会话对象
sess = tf.InteractiveSession(config=tf.ConfigProto(allow_soft_placement=True, gpu_options=gpu_opts))
# with tf.compat.v1.Session(config=tf.ConfigProto(allow_soft_placement=True, gpu_options=gpu_opts)) as sess:
# 将Keras作为tensorflow的精简接口
K.set_session(sess)
# 指定tensorflow运行的GPU或CPU设备
with tf.device('/device:GPU:%d' % gpu_num):
# 定义共指消解输入层
# 1. mention-pair的输入
mm_pft = Input(shape=(pftdim,))
# 2. mention1的输入， mention2的输入
m1_mft, m2_mft = Input(shape=(mftdim,)), Input(shape=(mftdim,))
# 3. 针对某个mention抽取出的所有特征的输入
m1_efts = [Input(shape=(r, d)) for r, d in eftdims]
# 4. 针对某个mention抽取出的所有特征的输入
m2_efts = [Input(shape=(r, d)) for r, d in eftdims]
# mention表达的命名空间
with tf.name_scope('mrepr'):
expand_dims, reshape_df, dropout = [], Reshape((nb_fltrs,)), Dropout(0.8)
# 所有特征的1-gram卷积层序列，2-gram卷积层序列，3-gram卷积层序列，1-gram池化层序列，2-gram池化层序列，3-gram池化层序列
eft_c1rs, eft_c2rs, eft_c3rs, eft_p1rs, eft_p2rs, eft_p3rs = [], [], [], [], [], []
# 遍历每1个抽取出的特征的维度信息
for r, d in eftdims:
expand_dims.append(Reshape((r, d, -1)))
# CONV1 卷积核窗口为1个单词，即1-gram滤波器，从卷积层输入形状(None, 2, 286, 1)表示channels为1，因此相当于Conv1D
eft_c1rs.append(Conv2D(nb_fltrs, (1, d), activation='tanh'))
# CONV1 卷积核窗口为2个单词，即2-gram滤波器
eft_c2rs.append(Conv2D(nb_fltrs, (2, d), activation='tanh'))
# CONV1 卷积核窗口为3个单词，即3-gram滤波器
eft_c3rs.append(Conv2D(nb_fltrs, (3, d), activation='tanh'))
# CONV1 卷积核窗口为1个单词的池化层
eft_p1rs.append(MaxPooling2D(pool_size=(r - 0, 1)))
# CONV1 卷积核窗口为2个单词的池化层
eft_p2rs.append(MaxPooling2D(pool_size=(r - 1, 1)))
# CONV1 卷积核窗口为3个单词的池化层
eft_p3rs.append(MaxPooling2D(pool_size=(r - 2, 1)))
# CONV2
nb_rows, mrepr_conv = 3 * nb_efts, Conv2D(nb_fltrs, (1, nb_fltrs), activation='tanh')
reshape_cefts, mrepr_pool = Reshape((nb_rows, nb_fltrs, -1)), MaxPooling2D(pool_size=(nb_rows, 1))
def mrepr(m_efts, m_mft, name=None):
eft_v1rs, eft_v2rs, eft_v3rs = [], [], []
for expand_dim, eft, c1r, c2r, c3r, p1r, p2r, p3r
in zip(expand_dims, m_efts, eft_c1rs, eft_c2rs, eft_c3rs, eft_p1rs, eft_p2rs, eft_p3rs):
eft_e = expand_dim(eft)
eft_v1rs.append(dropout(reshape_df(p1r(c1r(eft_e)))))
eft_v2rs.append(dropout(reshape_df(p2r(c2r(eft_e)))))
eft_v3rs.append(dropout(reshape_df(p3r(c3r(eft_e)))))
eft_mat = reshape_cefts(concatenate(eft_v1rs + eft_v2rs + eft_v3rs))
eft_vec = reshape_df(mrepr_pool(mrepr_conv(eft_mat)))
return concatenate([eft_vec, m_mft], name=name)
m1_mrepr = mrepr(m1_efts, m1_mft, name='m1_mrpr')
m2_mrepr = mrepr(m2_efts, m2_mft, name='m2_mrpr')
# mention-pair表达的命名空间
with tf.name_scope('mpair'):
mpair_conv = Conv2D(nb_fltrs, (1, nb_fltrs + mftdim), activation='tanh')
reshape_reprs, mpair_pool = Reshape((2, nb_fltrs + mftdim, -1)), MaxPooling2D(pool_size=(2, 1))
def mpair(mrepr1, mrepr2, pwft, name=None):
mat = reshape_reprs(concatenate([mrepr1, mrepr2]))
pft_vec = reshape_df(mpair_pool(mpair_conv(mat)))
return concatenate([pft_vec, pwft], name=name)
mm_mpair = mpair(m1_mrepr, m2_mrepr, mm_pft, name='mm_mpair')
# 预测mention-pair中两个mention关系的命名空间
with tf.name_scope('preds'):
mm_hidden = dropout(Dense(nb_fltrs, activation='relu')(mm_mpair))
mm_probs = Dense(3, activation='softmax', name='mm_probs', kernel_regularizer=l2(0.005))(mm_hidden)
# 构造mention嵌入表达的模型
mrepr_inputs = m2_efts + [m2_mft]
mrepr_model = Model(inputs=mrepr_inputs, outputs=[m2_mrepr], name='mm_mrepr_model')
# 构造mention-parir嵌入表达的模型
mpair_inputs = m1_efts + m2_efts + [m1_mft, m2_mft, mm_pft]
mpair_model = Model(inputs=mpair_inputs, outputs=[mm_mpair], name='mm_mpair_model')
# 构建mention-pair排序得分模型
ranking_inputs = m1_efts + m2_efts + [m1_mft, m2_mft, mm_pft]
ranking_model = Model(inputs=ranking_inputs, outputs=[mm_probs], name='mm_ranking_model')
ranking_model.compile(optimizer=RMSprop(),
loss=['sparse_categorical_crossentropy'],
metrics=['sparse_categorical_accuracy'])
ranking_model.summary()
plot_model(ranking_model, to_file='./model.png')输出的summary
Model: "mm_ranking_model"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_4 (InputLayer) (None, 3, 50) 0
__________________________________________________________________________________________________
input_5 (InputLayer) (None, 8, 50) 0
__________________________________________________________________________________________________
input_6 (InputLayer) (None, 5, 50) 0
__________________________________________________________________________________________________
input_7 (InputLayer) (None, 5, 50) 0
__________________________________________________________________________________________________
reshape_2 (Reshape) (None, 3, 50, 1) 0 input_4[0][0]
input_8[0][0]
__________________________________________________________________________________________________
reshape_3 (Reshape) (None, 8, 50, 1) 0 input_5[0][0]
input_9[0][0]
__________________________________________________________________________________________________
reshape_4 (Reshape) (None, 5, 50, 1) 0 input_6[0][0]
input_10[0][0]
__________________________________________________________________________________________________
reshape_5 (Reshape) (None, 5, 50, 1) 0 input_7[0][0]
input_11[0][0]
__________________________________________________________________________________________________
conv2d_1 (Conv2D) (None, 3, 1, 280) 14280 reshape_2[0][0]
reshape_2[1][0]
__________________________________________________________________________________________________
conv2d_4 (Conv2D) (None, 8, 1, 280) 14280 reshape_3[0][0]
reshape_3[1][0]
__________________________________________________________________________________________________
conv2d_7 (Conv2D) (None, 5, 1, 280) 14280 reshape_4[0][0]
reshape_4[1][0]
__________________________________________________________________________________________________
conv2d_10 (Conv2D) (None, 5, 1, 280) 14280 reshape_5[0][0]
reshape_5[1][0]
__________________________________________________________________________________________________
conv2d_2 (Conv2D) (None, 2, 1, 280) 28280 reshape_2[0][0]
reshape_2[1][0]
__________________________________________________________________________________________________
conv2d_5 (Conv2D) (None, 7, 1, 280) 28280 reshape_3[0][0]
reshape_3[1][0]
__________________________________________________________________________________________________
conv2d_8 (Conv2D) (None, 4, 1, 280) 28280 reshape_4[0][0]
reshape_4[1][0]
__________________________________________________________________________________________________
conv2d_11 (Conv2D) (None, 4, 1, 280) 28280 reshape_5[0][0]
reshape_5[1][0]
__________________________________________________________________________________________________
conv2d_3 (Conv2D) (None, 1, 1, 280) 42280 reshape_2[0][0]
reshape_2[1][0]
__________________________________________________________________________________________________
conv2d_6 (Conv2D) (None, 6, 1, 280) 42280 reshape_3[0][0]
reshape_3[1][0]
__________________________________________________________________________________________________
conv2d_9 (Conv2D) (None, 3, 1, 280) 42280 reshape_4[0][0]
reshape_4[1][0]
__________________________________________________________________________________________________
conv2d_12 (Conv2D) (None, 3, 1, 280) 42280 reshape_5[0][0]
reshape_5[1][0]
__________________________________________________________________________________________________
max_pooling2d_1 (MaxPooling2D) (None, 1, 1, 280) 0 conv2d_1[0][0]
conv2d_1[1][0]
__________________________________________________________________________________________________
max_pooling2d_4 (MaxPooling2D) (None, 1, 1, 280) 0 conv2d_4[0][0]
conv2d_4[1][0]
__________________________________________________________________________________________________
max_pooling2d_7 (MaxPooling2D) (None, 1, 1, 280) 0 conv2d_7[0][0]
conv2d_7[1][0]
__________________________________________________________________________________________________
max_pooling2d_10 (MaxPooling2D) (None, 1, 1, 280) 0 conv2d_10[0][0]
conv2d_10[1][0]
__________________________________________________________________________________________________
max_pooling2d_2 (MaxPooling2D) (None, 1, 1, 280) 0 conv2d_2[0][0]
conv2d_2[1][0]
__________________________________________________________________________________________________
max_pooling2d_5 (MaxPooling2D) (None, 1, 1, 280) 0 conv2d_5[0][0]
conv2d_5[1][0]
__________________________________________________________________________________________________
max_pooling2d_8 (MaxPooling2D) (None, 1, 1, 280) 0 conv2d_8[0][0]
conv2d_8[1][0]
__________________________________________________________________________________________________
max_pooling2d_11 (MaxPooling2D) (None, 1, 1, 280) 0 conv2d_11[0][0]
conv2d_11[1][0]
__________________________________________________________________________________________________
max_pooling2d_3 (MaxPooling2D) (None, 1, 1, 280) 0 conv2d_3[0][0]
conv2d_3[1][0]
__________________________________________________________________________________________________
max_pooling2d_6 (MaxPooling2D) (None, 1, 1, 280) 0 conv2d_6[0][0]
conv2d_6[1][0]
__________________________________________________________________________________________________
max_pooling2d_9 (MaxPooling2D) (None, 1, 1, 280) 0 conv2d_9[0][0]
conv2d_9[1][0]
__________________________________________________________________________________________________
max_pooling2d_12 (MaxPooling2D) (None, 1, 1, 280) 0 conv2d_12[0][0]
conv2d_12[1][0]
__________________________________________________________________________________________________
reshape_1 (Reshape) (None, 280) 0 max_pooling2d_1[0][0]
max_pooling2d_2[0][0]
max_pooling2d_3[0][0]
max_pooling2d_4[0][0]
max_pooling2d_5[0][0]
max_pooling2d_6[0][0]
max_pooling2d_7[0][0]
max_pooling2d_8[0][0]
max_pooling2d_9[0][0]
max_pooling2d_10[0][0]
max_pooling2d_11[0][0]
max_pooling2d_12[0][0]
max_pooling2d_13[0][0]
max_pooling2d_1[1][0]
max_pooling2d_2[1][0]
max_pooling2d_3[1][0]
max_pooling2d_4[1][0]
max_pooling2d_5[1][0]
max_pooling2d_6[1][0]
max_pooling2d_7[1][0]
max_pooling2d_8[1][0]
max_pooling2d_9[1][0]
max_pooling2d_10[1][0]
max_pooling2d_11[1][0]
max_pooling2d_12[1][0]
max_pooling2d_13[1][0]
max_pooling2d_14[0][0]
__________________________________________________________________________________________________
dropout_1 (Dropout) (None, 280) 0 reshape_1[0][0]
reshape_1[1][0]
reshape_1[2][0]
reshape_1[3][0]
reshape_1[4][0]
reshape_1[5][0]
reshape_1[6][0]
reshape_1[7][0]
reshape_1[8][0]
reshape_1[9][0]
reshape_1[10][0]
reshape_1[11][0]
reshape_1[13][0]
reshape_1[14][0]
reshape_1[15][0]
reshape_1[16][0]
reshape_1[17][0]
reshape_1[18][0]
reshape_1[19][0]
reshape_1[20][0]
reshape_1[21][0]
reshape_1[22][0]
reshape_1[23][0]
reshape_1[24][0]
dense_1[0][0]
__________________________________________________________________________________________________
concatenate_1 (Concatenate) (None, 3360) 0 dropout_1[0][0]
dropout_1[3][0]
dropout_1[6][0]
dropout_1[9][0]
dropout_1[1][0]
dropout_1[4][0]
dropout_1[7][0]
dropout_1[10][0]
dropout_1[2][0]
dropout_1[5][0]
dropout_1[8][0]
dropout_1[11][0]
__________________________________________________________________________________________________
input_8 (InputLayer) (None, 3, 50) 0
__________________________________________________________________________________________________
input_9 (InputLayer) (None, 8, 50) 0
__________________________________________________________________________________________________
input_10 (InputLayer) (None, 5, 50) 0
__________________________________________________________________________________________________
input_11 (InputLayer) (None, 5, 50) 0
__________________________________________________________________________________________________
reshape_6 (Reshape) (None, 12, 280, 1) 0 concatenate_1[0][0]
concatenate_2[0][0]
__________________________________________________________________________________________________
conv2d_13 (Conv2D) (None, 12, 1, 280) 78680 reshape_6[0][0]
reshape_6[1][0]
__________________________________________________________________________________________________
max_pooling2d_13 (MaxPooling2D) (None, 1, 1, 280) 0 conv2d_13[0][0]
conv2d_13[1][0]
__________________________________________________________________________________________________
concatenate_2 (Concatenate) (None, 3360) 0 dropout_1[12][0]
dropout_1[15][0]
dropout_1[18][0]
dropout_1[21][0]
dropout_1[13][0]
dropout_1[16][0]
dropout_1[19][0]
dropout_1[22][0]
dropout_1[14][0]
dropout_1[17][0]
dropout_1[20][0]
dropout_1[23][0]
__________________________________________________________________________________________________
input_2 (InputLayer) (None, 72) 0
__________________________________________________________________________________________________
input_3 (InputLayer) (None, 72) 0
__________________________________________________________________________________________________
m1_mrpr (Concatenate) (None, 352) 0 reshape_1[12][0]
input_2[0][0]
__________________________________________________________________________________________________
m2_mrpr (Concatenate) (None, 352) 0 reshape_1[25][0]
input_3[0][0]
__________________________________________________________________________________________________
concatenate_3 (Concatenate) (None, 704) 0 m1_mrpr[0][0]
m2_mrpr[0][0]
__________________________________________________________________________________________________
reshape_7 (Reshape) (None, 2, 352, 1) 0 concatenate_3[0][0]
__________________________________________________________________________________________________
conv2d_14 (Conv2D) (None, 2, 1, 280) 98840 reshape_7[0][0]
__________________________________________________________________________________________________
max_pooling2d_14 (MaxPooling2D) (None, 1, 1, 280) 0 conv2d_14[0][0]
__________________________________________________________________________________________________
input_1 (InputLayer) (None, 6) 0
__________________________________________________________________________________________________
mm_mpair (Concatenate) (None, 286) 0 reshape_1[26][0]
input_1[0][0]
__________________________________________________________________________________________________
dense_1 (Dense) (None, 280) 80360 mm_mpair[0][0]
__________________________________________________________________________________________________
mm_probs (Dense) (None, 3) 843 dropout_1[24][0]
==================================================================================================
Total params: 598,083
Trainable params: 598,083
Non-trainable params: 0
__________________________________________________________________________________________________模型图
5371
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
