TensorFlow多任务多标签分类

发表于 | 分类于 | | 阅读次数:

本文是应用TensorFlow进行多任务&多标签分类的一个示例(TensorFlow Multi-Task Multi-Label Example)。

以Mnist数据集为燃料，以简单的MLP为引擎，代码走起。

本文代码见：https://github.com/xylcbd/blogs_code/blob/master/tensorflow-multitask-multilabel/main.py

任务概述

任务共有2个，分别如下；

  • 第1个任务是识别数字（即10类分类问题）
  • 第2个任务是识别属性（图片中的数字是否是奇数、图片中的数字是否大于5）

可以看到第2个任务是有多个属性，即多标签任务。

  • [0, 0]代表[不是奇数, 不大于5]
  • [1, 0]代表[是奇数, 不大于5]
  • [0, 1]代表[不是奇数, 大于5]
  • [1, 1]代表[是奇数, 大于5]

属性可以再增加，不是固定为2个属性。

模型概述

简单的解释一下。

这里模型采用的是一个3层MLP模型，前2层共用，最后一层为2个任务各自所有。

  • 对于预测数字的任务，采用的是tf.losses.softmax_cross_entropy
  • 对于预测属性的任务，采用的是tf.losses.sigmoid_cross_entropy

示例代码

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
#coding: utf-8
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import time
import os
import sys
import numpy as np

######################################
#tools
class Tools(object):
    @staticmethod
    def currect_time():
        return time.strftime("%H:%M:%S", time.localtime()) + '.%03d' % (time.time() % 1 * 1000)
    @staticmethod
    def log_print(content):
        print("[" + Tools.currect_time() + "] " + content)

######################################
#setting
class Setting(object):
    @staticmethod
    def checkpoint_dir():
        return "model"

######################################
#network
class Network(object):
    def __init__(self):
        # Network Parameters
        self.num_input = 28 # MNIST data input (img shape: 28*28)
        self.num_class = 10 # MNIST total class (0-9 digits)
        self.num_attrs = 2 #MNIST number attrs, is_odd_number & is_larger_than_5
        self.num_hidden = 128

        #inputs <-> goldens
        self.inputs = tf.placeholder("float", [None, self.num_input * self.num_input])
        self.golden_class = tf.placeholder("float", [None, self.num_class])
        self.golden_attrs = tf.placeholder("float", [None, self.num_attrs])

        out = self.fc('fc1', self.inputs, self.num_input*self.num_input, self.num_hidden, True)
        out = self.fc('fc2', out, self.num_hidden, self.num_hidden, True)

        #last tail layer
        self.logits_class = self.fc('fc_class', out, self.num_hidden, self.num_class, False)
        self.logits_attrs = self.fc('fc_attrs', out, self.num_hidden, self.num_attrs, False)

    def get_weight_varible(self, name, shape):
        return tf.get_variable(name, shape=shape,
                       initializer=tf.contrib.layers.xavier_initializer())

    def get_bias_varible(self, name, shape):
        return tf.get_variable(name, shape=shape,
                           initializer=tf.contrib.layers.xavier_initializer())
    
    def fc(self, layer_name, x, inp_dim, out_dim, with_act):
        with tf.variable_scope(layer_name):
            y = tf.reshape(x, shape=[-1, inp_dim])
            w = self.get_weight_varible('w', [inp_dim, out_dim])
            b = self.get_bias_varible('b', [out_dim])
            y = tf.add(tf.matmul(y, w), b)
            if with_act:
                y = tf.nn.relu(y)
        return y

#####################################
#main route
def save_model(saver, sess, model_path):    
    Tools.log_print('save model to {0}.'.format(model_path))
    saver.save(sess, model_path)

def load_model(saver, sess, model_path):
    Tools.log_print('try to load model from {0}.'.format(model_path))    
    saver.restore(sess, model_path)
    Tools.log_print('load model success')
    return True

def custom_lables(batch_class):
    batch_attrs = []
    for label in batch_class:
        idx = np.argmax(label)
        attrs = [float(idx % 2), float(idx > 5)]
        batch_attrs.append(attrs)
    return batch_class, batch_attrs 

def train():
    Tools.log_print('loading dataset...')
    mnist = input_data.read_data_sets("/tmp/data/", one_hot=True)
    Tools.log_print('loade dataset success.\n')

    Tools.log_print('building network...')
    model_dir = 'model'
    model_path = os.path.join(model_dir, 'mnist_MLP')
    network = Network()
    #training parameters
    num_input = 28
    learning_rate = 0.01
    display_step = 100
    train_epochs = 10
    train_batchsize = 64
    test_batchsize = 64

    #loss of class
    loss_class_op = tf.reduce_mean(tf.losses.softmax_cross_entropy(
        logits= network.logits_class, onehot_labels=network.golden_class))
    #loss of attrs
    loss_attrs_op = tf.reduce_mean(tf.losses.sigmoid_cross_entropy(
        logits= network.logits_attrs, multi_class_labels=network.golden_attrs))
    #total loss
    loss_op = loss_class_op + loss_attrs_op

    #optimizer
    optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
    #train operator
    train_op = optimizer.minimize(loss_op)

    #accuracy of class
    correct_class_op = tf.equal(tf.argmax(network.logits_class, 1), tf.argmax(network.golden_class, 1))
    accuracy_class_op = tf.reduce_mean(tf.cast(correct_class_op, tf.float32))
    #accuracy of attrs
    correct_attrs_op = tf.equal(tf.cast(tf.greater_equal(tf.sigmoid(network.logits_attrs), 0.5), tf.int32), tf.cast(network.golden_attrs, tf.int32))
    accuracy_attrs_op = tf.reduce_mean(tf.reduce_min(tf.cast(correct_attrs_op, tf.float32), 1))

    #initialize the variables (i.e. assign their default value)
    init_op = tf.global_variables_initializer()
    Tools.log_print('build network success.\n')

    #start training
    with tf.Session() as sess:
        #run the initializer
        sess.run(init_op)
        
        #model dir
        saver = tf.train.Saver(tf.global_variables())
        if not os.path.exists(model_dir):
            os.makedirs(model_dir)

        #run epochs
        Tools.log_print('start training...')
        for epoch in range(1, train_epochs+1):
            train_steps = len(mnist.train.labels) / train_batchsize   
            total_loss = 0.0
            total_acc_class = 0.0
            total_acc_attrs = 0.0
            for step in range(1, train_steps+1):
                batch_x, batch_y = mnist.train.next_batch(train_batchsize)
                batch_class, batch_attrs = custom_lables(batch_y)
                batch_x = batch_x.reshape((train_batchsize, num_input * num_input))
                _, batch_loss, batch_acc_class, batch_acc_attrs = sess.run([train_op, loss_op, accuracy_class_op, accuracy_attrs_op], feed_dict={network.inputs: batch_x, network.golden_class: batch_class, network.golden_attrs: batch_attrs})
                total_loss += batch_loss
                total_acc_class += batch_acc_class
                total_acc_attrs += batch_acc_attrs
                if step % display_step == 0:
                    avg_loss = total_loss / display_step
                    avg_acc_class = total_acc_class / display_step
                    avg_acc_attrs = total_acc_attrs / display_step
                    Tools.log_print("Epoch[%d/%d] Step[%d/%d] Train Minibatch Loss= %.4f, Class Accuracy= %.4f, Attrs Accuracy= %.4f" % (epoch, train_epochs, step, train_steps, avg_loss, avg_acc_class, avg_acc_attrs))
                    total_loss = 0.0
                    total_acc_class = 0.0
                    total_acc_attrs = 0.0
        Tools.log_print("finished training!")
        
        #save model
        save_model(saver, sess, model_path)

        #load model and test
        Tools.log_print('start testing...')
        if load_model(saver, sess, model_path):                     
            test_steps = len(mnist.test.labels) / test_batchsize   
            acc_class_list = []
            acc_attrs_list = []
            for step in range(1, test_steps+1):
                batch_x, batch_y = mnist.test.next_batch(test_batchsize)                
                batch_x = batch_x.reshape((test_batchsize, num_input * num_input))
                batch_class, batch_attrs = custom_lables(batch_y)
                batch_acc_class, batch_acc_attrs = sess.run([accuracy_class_op, accuracy_attrs_op], feed_dict={network.inputs: batch_x, network.golden_class: batch_class, network.golden_attrs: batch_attrs})
                acc_class_list.append(batch_acc_class)
                acc_attrs_list.append(batch_acc_attrs)
            acc_class = np.mean(acc_class_list)
            acc_attrs = np.mean(acc_attrs_list)
            Tools.log_print("Testing Class Accuracy: %.4f, Attrs Accuracy: %.4f" % (acc_class, acc_attrs))
        Tools.log_print('finished testing...')

if __name__ == '__main__':
    train()

模型效果

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
[17:46:51.972] Epoch[10/10] Step[100/859] Train Minibatch Loss= 0.1084, Class Accuracy= 0.9791, Attrs Accuracy= 0.9848
[17:46:52.442] Epoch[10/10] Step[200/859] Train Minibatch Loss= 0.1188, Class Accuracy= 0.9802, Attrs Accuracy= 0.9852
[17:46:52.905] Epoch[10/10] Step[300/859] Train Minibatch Loss= 0.1122, Class Accuracy= 0.9797, Attrs Accuracy= 0.9859
[17:46:53.357] Epoch[10/10] Step[400/859] Train Minibatch Loss= 0.1019, Class Accuracy= 0.9814, Attrs Accuracy= 0.9875
[17:46:53.833] Epoch[10/10] Step[500/859] Train Minibatch Loss= 0.0833, Class Accuracy= 0.9841, Attrs Accuracy= 0.9886
[17:46:54.297] Epoch[10/10] Step[600/859] Train Minibatch Loss= 0.1323, Class Accuracy= 0.9745, Attrs Accuracy= 0.9828
[17:46:54.759] Epoch[10/10] Step[700/859] Train Minibatch Loss= 0.1333, Class Accuracy= 0.9752, Attrs Accuracy= 0.9808
[17:46:55.224] Epoch[10/10] Step[800/859] Train Minibatch Loss= 0.1227, Class Accuracy= 0.9767, Attrs Accuracy= 0.9838
[17:46:55.493] finished training!
[17:46:55.493] save model to model/mnist_MLP.
[17:46:55.584] start testing...
[17:46:55.584] try to load model from model/mnist_MLP.
[17:46:55.636] load model success
[17:46:56.014] Testing Class Accuracy: 0.9637, Attrs Accuracy: 0.9722
[17:46:56.014] finished testing...
洪柏敏 wechat
欢迎留言评论，邮件给我(dahanhbm AT gmail DOT com)，或者关注我的微信公众号