Tensorflow常见模型及工程化方法
姓名 李林涛 学号 16020199032
转自：http://mp.weixin.qq.com/s?__biz=MzA5NDM4MjMxOQ==&mid=2447579217&idx=1&sn=93eede10b5c027917b91f9cc819e74e7&chksm=8458c9d1b32f40c75c6c3a0ec0434842e1c9efa92c7d0625d7760042acf7cf7ae13c05b1db58&mpshare=1&scene=23&srcid=1209EwoWaCSocVcEitXFbtTc#rd
【嵌牛导读】：作者介绍了Tensorflow中的常见模型和工程化方法
【嵌牛鼻子】：Tensorflow
【嵌牛提问】：Tensorflow的常见框架和工程化方法是什么？
【嵌牛正文】：
Tensorflow在深度学习模型研究中起到了很大的促进作用，灵活的框架免去了研究人员、开发者大量的自动求导代码工作。本文总结一下常用的模型代码和工程化需要的代码。有需求的同学收藏一下，以便日后查阅。
Tensorflow常见模型
A. LSTM模型结构
import tensorflow as tf
import tensorflow.contrib as contrib
from tensorflow.python.ops import array_ops
class lstm(object):
def __init__(self, in_data, hidden_dim, batch_seqlen=None, flag='concat'):
self.in_data = in_data
self.hidden_dim = hidden_dim
self.batch_seqlen = batch_seqlen
self.flag = flag
lstm_cell = contrib.rnn.LSTMCell(self.hidden_dim)
out, _ = tf.nn.dynamic_rnn(cell=lstm_cell, inputs=self.in_data, sequence_length=self.batch_seqlen,dtype=tf.float32)
if flag=='all_ht':
self.out = out
if flag = 'first_ht':
self.out = out[:,0,:]
if flag = 'last_ht':
self.out = out[:,-1,:]
if flag = 'concat':
self.out = tf.concat([out[:,0,:], out[:,-1,:]],1)
B. Bi-LSTM模型结构
import tensorflow as tf
import tensorflow.contrib as contrib
from tensorflow.python.ops import array_ops
from tensorflow.python.framework import dtypes
class bilstm(object):
def __init__(self, in_data, hidden_dim, batch_seqlen=None, flag='concat'):
self.in_data = in_data
self.hidden_dim = hidden_dim
self.batch_seqlen = batch_seqlen
self.flag = flag
lstm_cell_fw = contrib.rnn.LSTMCell(self.hidden_dim)
lstm_cell_bw = contrib.rnn.LSTMCell(self.hidden_dim)
out, state =
tf.nn.bidirectional_dynamic_rnn(cell_fw=lstm_cell_fw,cell_bw=lstm_cell_bw,inputs=self.in_data,
sequence_lenth=self.batch_seqlen,dtype=tf.float32)
bi_out = tf.concat(out, 2)
if flag=='all_ht':
self.out = bi_out
if flag=='first_ht':
self.out = bi_out[:,0,:]
if flag=='last_ht':
self.out = tf.concat([state[0].h,state[1].h], 1)
if flag=='concat':
self.out = tf.concat([bi_out[:,0,:],tf.concat([state[0].h,state[1].h], 1)],1)
C multi-channel CNN
import tensorflow as tf
import tensorflow.contrib as contrib
from tensorflow.python.ops import array_ops
class lstm(object):
def __init__(self, in_data, hidden_dim, batch_seqlen=None, flag='concat'):
self.in_data = in_data
self.hidden_dim = hidden_dim
self.batch_seqlen = batch_seqlen
self.flag = flag
lstm_cell = contrib.rnn.LSTMCell(self.hidden_dim)
out, _ = tf.nn.dynamic_rnn(cell=lstm_cell, inputs=self.in_data, sequence_length=self.batch_seqlen,dtype=tf.float32)
if flag=='all_ht':
self.out = out
if flag = 'first_ht':
self.out = out[:,0,:]
if flag = 'last_ht':
self.out = out[:,-1,:]
if flag = 'concat':
self.out = tf.concat([out[:,0,:], out[:,-1,:]],1)
D depth-wise cnn
import tensorflow as tf
def depth_wise_conv(in_data, scope, kernel_size, dim):
with tf.variable_scope(scope):
shapes = in_data.shape.as_list()
depthwise_filter = tf.get_varibale("depthwise_conv.weight",
(kernel_size[0], kernel_size[1], shapes[-1]
dtype=tf.float32, )
pointwise_filter = tf.get_variable("pointwise_conv.weight",
(1,1, shapes[-1], dim),
dtype=tf.float32, )
outputs = tf.nn.separable_conv2d(in_data,
depthwise_filter,
pointwise_filter,
strides=(1,1,1,1),
padding="SAME"
)
return outputs
D multi-layer depth-wise cnn
def multi_convs(input_x, dim, conv_number=2, k=5):
# input_x: 输入数据，为batch * seq * dim
# dim：对应的输入的维度
# conv_number: 对应的卷积的层数，一般2，
# k对应的是卷积核的窗口大小
res = input_x
for index in range(conv_number):
out = norm(res) # layer norm
out = tf.expand_dims(out, 2) # bach * seq * 1 * dim
out = depth_wise_conv(out, kernel_size=(k, 1), dim=dim, scope="convs.%d" % index)
out = tf.squeeze(out, 2) # batch * seq * dim
out = tf.nn.relu(out)
out = out + res
res = out
out = norm(out) # 输出为 batch * seq * 1 * dim
out = tf.squeeze(out, squeeze_dims=2) # 输出为 batch * seq * dim
return out
模型参数查看
已知模型文件的ckpt文件，通过pywrap_tensorflow获取模型的各参数名。
import tensoflow as tf
from tensorflow.python import pywrap_tensorflow
model_dir = "./ckpt/"
ckpt = tf.train.get_checkpoint_state(model_dir)
ckpt_path = ckpt.model_checkpoint_path
reader = pywrap_tensorflow.NewCheckpointReader(ckpt_path)
param_dict = reader.get_variable_to_shape_map()
for key, val in param_dict.items():
try:
print key, val
except:
pass
工程化方法
A. tennsorflow模型文件打包成PB文件
import tensorflow as tf
from tensorflow.python.tools import freeze_graph
with tf.Graph().as_default():
with tf.device("/cpu:0"):
config = tf.ConfigProto(allow_soft_placement=True)
with tf.Session(config=config).as_default() as sess:
model = Your_Model_Name()
model.build_graph()
sess.run(tf.initialize_all_variables())
saver = tf.train.Saver()
ckpt_path = "/your/model/path"
saver.restore(sess, ckpt_path)
graphdef = tf.get_default_graph().as_graph_def()
tf.train.write_graph(sess.graph_def,"/your/save/path/","save_name.pb",as_text=False)
frozen_graph = tf.graph_util.convert_variables_to_constants(sess,graphdef,['output/node/name'])
frozen_graph_trim = tf.graph_util.remove_training_nodes(frozen_graph)
freeze_graph.freeze_graph('/your/save/path/save_name.pb','',True,
ckpt_path,'output/node/name','save/restore_all','save/Const:0','frozen_name.pb',True,"")
B.PB文件读取使用
output_graph_def = tf.GraphDef()
with open("your_name.pb","rb") as f:
output_graph_def.ParseFromString(f.read())
_ = tf.import_graph_def(output_graph_def, name="")
node_in = sess.graph.get_tensor_by_name("input_node_name")
model_out = sess.graph.get_tensor_by_name("out_node_name")
feed_dict = {node_in:in_data}
pred = sess.run(model_out, feed_dict)
