Tensorflow常见模型及工程化方法

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)

注：本文代码均为笔者手敲留存，如代码有误可以咨询探讨。