（原）tensorflow使用eager在mnist上训练的简单例子

转载请注明出处：

http://www.javashuo.com/article/p-ocbrtjgk-t.html

代码网址：

https://github.com/darkknightzh/trainEagerMnist

参考网址：

https://github.com/tensorflow/models/blob/master/official/mnist/mnist_eager.py

https://github.com/madalinabuzau/tensorflow-eager-tutorials/blob/master/07_convolutional_neural_networks_for_emotion_recognition.ipynb

一 整体流程

tensorflow使用eager时，须要下面几句话（若是不使用第三句话，则依旧能够使用静态图）：

import tensorflow as tf
import tensorflow.contrib.eager as tfe
tfe.enable_eager_execution()

tensorflow使用eager模式后，感受和pytorch同样方便。使用eager后，不须要tf.placeholder，用起来更加方便。

目前貌似tf.keras.layers和tf.layers支持eager，slim不支持。

整体流程以下：

initial optimizer
for I in range(epochs):
    for imgs, targets in training_data:
        with tf.GradientTape() as tape:
            logits = model(imgs, training=True)
            loss_value = calc_loss(logits, targets)
        grads = tape.gradient(loss_value, model.variables)
        optimizer.apply_gradients(zip(grads, model.variables), global_step=step_counter)
        update training_accurate, total_loss
    test model
    save model

二 建立模型

能够使用下面三种方式建立模型

1. 相似pytorch的方式

先在__init__中定义用到的层，而后重载call函数，构建网络。模型前向计算时，会调用call函数。以下面代码所示：

 1 class simpleModel(tf.keras.Model):
 2     def __init__(self, num_classes):
 3         super(simpleModel, self).__init__()
 4 
 5         input_shape = [28, 28, 1]
 6         data_format = 'channels_last'
 7         self.reshape = tf.keras.layers.Reshape(target_shape=input_shape, input_shape=(input_shape[0] * input_shape[1],))
 8 
 9         self.conv1 = tf.keras.layers.Conv2D(16, 5, padding="same", activation='relu')
10         self.batch1 = tf.keras.layers.BatchNormalization()
11         self.pool1 = tf.keras.layers.MaxPooling2D((2, 2), (2, 2), padding='same', data_format=data_format)
12 
13         self.conv2 = tf.keras.layers.Conv2D(32, 5, padding="same", activation='relu')
14         self.batch2 = tf.keras.layers.BatchNormalization()
15         self.pool2 = tf.keras.layers.MaxPooling2D((2, 2), (2, 2), padding='same', data_format=data_format)
16 
17         self.conv3 = tf.keras.layers.Conv2D(64, 5, padding="same", activation='relu')
18         self.batch3 = tf.keras.layers.BatchNormalization()
19         self.pool3 = tf.keras.layers.MaxPooling2D((2, 2), (2, 2), padding='same', data_format=data_format)
20 
21         self.conv4 = tf.keras.layers.Conv2D(64, 5, padding="same", activation='relu')
22         self.batch4 = tf.keras.layers.BatchNormalization()
23         self.pool4 = tf.keras.layers.MaxPooling2D((2, 2), (2, 2), padding='same', data_format=data_format)
24 
25         self.flat = tf.keras.layers.Flatten()
26         self.fc5 = tf.keras.layers.Dense(1024, activation='relu')
27         self.batch5 = tf.keras.layers.BatchNormalization()
28 
29         self.fc6 = tf.keras.layers.Dense(num_classes)
30         self.batch6 = tf.keras.layers.BatchNormalization()
31 
32     def call(self, inputs, training=None):
33         x = self.reshape(inputs)
34 
35         x = self.conv1(x)
36         x = self.batch1(x, training=training)
37         x = self.pool1(x)
38 
39         x = self.conv2(x)
40         x = self.batch2(x, training=training)
41         x = self.pool2(x)
42 
43         x = self.conv3(x)
44         x = self.batch3(x, training=training)
45         x = self.pool3(x)
46 
47         x = self.conv4(x)
48         x = self.batch4(x, training=training)
49         x = self.pool4(x)
50 
51         x = self.flat(x)
52         x = self.fc5(x)
53         x = self.batch5(x, training=training)
54 
55         x = self.fc6(x)
56         x = self.batch6(x, training=training)
57         # x = tf.layers.dropout(x, rate=0.3, training=training)
58         return x
59 
60     def get_acc(self, target):
61         correct_prediction = tf.equal(tf.argmax(self.logits, 1), tf.argmax(target, 1))
62         acc = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
63         return acc
64 
65     def get_loss(self):
66         return self.loss
67 
68     def loss_fn(self, images, target, training):
69         self.logits = self(images, training)  # call call(self, inputs, training=None) function
70         self.loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=self.logits, labels=target))
71         return self.loss
72 
73     def grads_fn(self, images, target, training):  # do not return loss and acc if unnecessary
74         with tfe.GradientTape() as tape:
75             loss = self.loss_fn(images, target, training)
76         return tape.gradient(loss, self.variables)

2. 直接使用tf.keras.Sequential

以下面代码所示：

 1 def create_model1():
 2     data_format = 'channels_last'
 3     input_shape = [28, 28, 1]
 4     l = tf.keras.layers
 5     max_pool = l.MaxPooling2D((2, 2), (2, 2), padding='same', data_format=data_format)
 6     # The model consists of a sequential chain of layers, so tf.keras.Sequential (a subclass of tf.keras.Model) makes for a compact description.
 7     return tf.keras.Sequential(
 8         [
 9             l.Reshape(target_shape=input_shape, input_shape=(28 * 28,)),
10             l.Conv2D(16, 5, padding='same', data_format=data_format, activation=tf.nn.relu),
11             l.BatchNormalization(),
12             max_pool,
13 
14             l.Conv2D(32, 5, padding='same', data_format=data_format, activation=tf.nn.relu),
15             l.BatchNormalization(),
16             max_pool,
17 
18             l.Conv2D(64, 5, padding='same', data_format=data_format, activation=tf.nn.relu),
19             l.BatchNormalization(),
20             max_pool,
21 
22             l.Conv2D(64, 5, padding='same', data_format=data_format, activation=tf.nn.relu),
23             l.BatchNormalization(),
24             max_pool,
25 
26             l.Flatten(),
27             l.Dense(1024, activation=tf.nn.relu),
28             l.BatchNormalization(),
29 
30             # # l.Dropout(0.4),
31             l.Dense(10),
32             l.BatchNormalization()
33         ])

3. 使用tf.keras.Sequential()及add函数

以下面代码所示：

 1 def create_model2():
 2     data_format = 'channels_last'
 3     input_shape = [28, 28, 1]
 4 
 5     model = tf.keras.Sequential()
 6 
 7     model.add(tf.keras.layers.Reshape(target_shape=input_shape, input_shape=(input_shape[0] * input_shape[1],)))
 8 
 9     model.add(tf.keras.layers.Conv2D(16, 5, padding="same", activation='relu'))
10     model.add(tf.keras.layers.BatchNormalization())
11     model.add(tf.keras.layers.MaxPooling2D((2, 2), (2, 2), padding='same', data_format=data_format))
12 
13     model.add(tf.keras.layers.Conv2D(32, 5, padding="same", activation='relu'))
14     model.add(tf.keras.layers.BatchNormalization())
15     model.add(tf.keras.layers.MaxPooling2D((2, 2), (2, 2), padding='same', data_format=data_format))
16 
17     model.add(tf.keras.layers.Conv2D(64, 5, padding="same", activation='relu'))
18     model.add(tf.keras.layers.BatchNormalization())
19     model.add(tf.keras.layers.MaxPooling2D((2, 2), (2, 2), padding='same', data_format=data_format))
20 
21     model.add(tf.keras.layers.Conv2D(64, 5, padding="same", activation='relu'))
22     model.add(tf.keras.layers.BatchNormalization())
23     model.add(tf.keras.layers.MaxPooling2D((2, 2), (2, 2), padding='same', data_format=data_format))
24 
25     model.add(tf.keras.layers.Flatten())
26     model.add(tf.keras.layers.Dense(1024, activation='relu'))
27     model.add(tf.keras.layers.BatchNormalization())
28 
29     model.add(tf.keras.layers.Dense(10))
30     model.add(tf.keras.layers.BatchNormalization())
31 
32 return model

三 使用动态图更新梯度

在更新梯度时，须要加上下面的几句话

1 with tf.GradientTape() as tape:
2     logits = model(imgs, training=True)
3     loss_value = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=labs))
4 grads = tape.gradient(loss_value, model.variables)
5 optimizer.apply_gradients(zip(grads, model.variables), global_step=step_counter)

第二行获得特征，第三行获得损失，第四行获得梯度，第五行将梯度应用到模型，更新模型参数。

四 保存及载入模型

1. 使用tfe.Saver

代码以下

1 def saveModelV1(model_dir, model, global_step, modelname='model1'):
2     tfe.Saver(model.variables).save(os.path.join(model_dir, modelname), global_step=global_step)
3 def restoreModelV1(model_dir, model):
4     dummy_input = tf.constant(tf.zeros((1, 28, 28, 1)))  # Run the model once to initialize variables
5     dummy_pred = model(dummy_input, training=False)
6 
7     saver = tfe.Saver(model.variables)  # Restore the variables of the model
8     saver.restore(tf.train.latest_checkpoint(model_dir))

2. 使用tf.train.Checkpoint

代码以下

1 step_counter = tf.train.get_or_create_global_step()
2 checkpoint = tf.train.Checkpoint(model=model, optimizer=optimizer, step_counter=step_counter)
3 
4 def saveModelV2(model_dir, checkpoint, modelname='model2'):
5     checkpoint_prefix = os.path.join(model_dir, modelname)
6     checkpoint.save(checkpoint_prefix)
7 
8 def restoreModelV2(model_dir, checkpoint):
9     checkpoint.restore(tf.train.latest_checkpoint(model_dir))

五 具体代码

代码未严格按照整体流程的步骤，仅供参考，见https://github.com/darkknightzh/trainEagerMnist

其中eagerFlag为使用eager的方式，0为不使用eager（使用静态图），1为使用V1的方式，2为使用V2的方式。当使用静态图时，不要加tfe.enable_eager_execution()，不然会报错。具体可参考代码。