（原）tensorflow使用eager在mnist上训练的简单例子

转载请注明出处：

https://www.cnblogs.com/darkknightzh/p/9989586.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 总体流程

tensorflow使用eager时，需要下面几句话（如果不使用第三句话，则依旧可以使用静态图）：

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

tensorflow使用eager模式后，感觉和pytorch一样方便。使用eager后，不需要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。

总体流程如下：

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 创建模型

可以使用下面三种方式创建模型

1. 类似pytorch的方式

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

class simpleModel(tf.keras.Model):
    def __init__(self, num_classes):
        super(simpleModel, self).__init__()

        input_shape = [28, 28, 1]
        data_format = 'channels_last'
        self.reshape = tf.keras.layers.Reshape(target_shape=input_shape, input_shape=(input_shape[0] * input_shape[1],))

        self.conv1 = tf.keras.layers.Conv2D(16, 5, padding="same", activation='relu')
        self.batch1 = tf.keras.layers.BatchNormalization()
        self.pool1 = tf.keras.layers.MaxPooling2D((2, 2), (2, 2), padding='same', data_format=data_format)

        self.conv2 = tf.keras.layers.Conv2D(32, 5, padding="same", activation='relu')
        self.batch2 = tf.keras.layers.BatchNormalization()
        self.pool2 = tf.keras.layers.MaxPooling2D((2, 2), (2, 2), padding='same', data_format=data_format)

        self.conv3 = tf.keras.layers.Conv2D(64, 5, padding="same", activation='relu')
        self.batch3 = tf.keras.layers.BatchNormalization()
        self.pool3 = tf.keras.layers.MaxPooling2D((2, 2), (2, 2), padding='same', data_format=data_format)

        self.conv4 = tf.keras.layers.Conv2D(64, 5, padding="same", activation='relu')
        self.batch4 = tf.keras.layers.BatchNormalization()
        self.pool4 = tf.keras.layers.MaxPooling2D((2, 2), (2, 2), padding='same', data_format=data_format)

        self.flat = tf.keras.layers.Flatten()
        self.fc5 = tf.keras.layers.Dense(1024, activation='relu')
        self.batch5 = tf.keras.layers.BatchNormalization()

        self.fc6 = tf.keras.layers.Dense(num_classes)
        self.batch6 = tf.keras.layers.BatchNormalization()

    def call(self, inputs, training=None):
        x = self.reshape(inputs)

        x = self.conv1(x)
        x = self.batch1(x, training=training)
        x = self.pool1(x)

        x = self.conv2(x)
        x = self.batch2(x, training=training)
        x = self.pool2(x)

        x = self.conv3(x)
        x = self.batch3(x, training=training)
        x = self.pool3(x)

        x = self.conv4(x)
        x = self.batch4(x, training=training)
        x = self.pool4(x)

        x = self.flat(x)
        x = self.fc5(x)
        x = self.batch5(x, training=training)

        x = self.fc6(x)
        x = self.batch6(x, training=training)
        # x = tf.layers.dropout(x, rate=0.3, training=training)
        return x

    def get_acc(self, target):
        correct_prediction = tf.equal(tf.argmax(self.logits, 1), tf.argmax(target, 1))
        acc = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
        return acc

    def get_loss(self):
        return self.loss

    def loss_fn(self, images, target, training):
        self.logits = self(images, training)  # call call(self, inputs, training=None) function
        self.loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=self.logits, labels=target))
        return self.loss

    def grads_fn(self, images, target, training):  # do not return loss and acc if unnecessary
        with tfe.GradientTape() as tape:
            loss = self.loss_fn(images, target, training)
        return tape.gradient(loss, self.variables)

2. 直接使用tf.keras.Sequential

如下面代码所示：

def create_model1():
    data_format = 'channels_last'
    input_shape = [28, 28, 1]
    l = tf.keras.layers
    max_pool = l.MaxPooling2D((2, 2), (2, 2), padding='same', data_format=data_format)
    # The model consists of a sequential chain of layers, so tf.keras.Sequential (a subclass of tf.keras.Model) makes for a compact description.
    return tf.keras.Sequential(
        [
            l.Reshape(target_shape=input_shape, input_shape=(28 * 28,)),
            l.Conv2D(16, 5, padding='same', data_format=data_format, activation=tf.nn.relu),
            l.BatchNormalization(),
            max_pool,

            l.Conv2D(32, 5, padding='same', data_format=data_format, activation=tf.nn.relu),
            l.BatchNormalization(),
            max_pool,

            l.Conv2D(64, 5, padding='same', data_format=data_format, activation=tf.nn.relu),
            l.BatchNormalization(),
            max_pool,

            l.Conv2D(64, 5, padding='same', data_format=data_format, activation=tf.nn.relu),
            l.BatchNormalization(),
            max_pool,

            l.Flatten(),
            l.Dense(1024, activation=tf.nn.relu),
            l.BatchNormalization(),

            # # l.Dropout(0.4),
            l.Dense(10),
            l.BatchNormalization()
        ])

3. 使用tf.keras.Sequential()及add函数

如下面代码所示：

def create_model2():
    data_format = 'channels_last'
    input_shape = [28, 28, 1]

    model = tf.keras.Sequential()

    model.add(tf.keras.layers.Reshape(target_shape=input_shape, input_shape=(input_shape[0] * input_shape[1],)))

    model.add(tf.keras.layers.Conv2D(16, 5, padding="same", activation='relu'))
    model.add(tf.keras.layers.BatchNormalization())
    model.add(tf.keras.layers.MaxPooling2D((2, 2), (2, 2), padding='same', data_format=data_format))

    model.add(tf.keras.layers.Conv2D(32, 5, padding="same", activation='relu'))
    model.add(tf.keras.layers.BatchNormalization())
    model.add(tf.keras.layers.MaxPooling2D((2, 2), (2, 2), padding='same', data_format=data_format))

    model.add(tf.keras.layers.Conv2D(64, 5, padding="same", activation='relu'))
    model.add(tf.keras.layers.BatchNormalization())
    model.add(tf.keras.layers.MaxPooling2D((2, 2), (2, 2), padding='same', data_format=data_format))

    model.add(tf.keras.layers.Conv2D(64, 5, padding="same", activation='relu'))
    model.add(tf.keras.layers.BatchNormalization())
    model.add(tf.keras.layers.MaxPooling2D((2, 2), (2, 2), padding='same', data_format=data_format))

    model.add(tf.keras.layers.Flatten())
    model.add(tf.keras.layers.Dense(1024, activation='relu'))
    model.add(tf.keras.layers.BatchNormalization())

    model.add(tf.keras.layers.Dense(10))
    model.add(tf.keras.layers.BatchNormalization())

return model

三 使用动态图更新梯度

在更新梯度时，需要加上下面的几句话

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

第二行得到特征，第三行得到损失，第四行得到梯度，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

代码如下

def saveModelV1(model_dir, model, global_step, modelname='model1'):
    tfe.Saver(model.variables).save(os.path.join(model_dir, modelname), global_step=global_step)
def restoreModelV1(model_dir, model):
    dummy_input = tf.constant(tf.zeros((1, 28, 28, 1)))  # Run the model once to initialize variables
    dummy_pred = model(dummy_input, training=False)

    saver = tfe.Saver(model.variables)  # Restore the variables of the model
    saver.restore(tf.train.latest_checkpoint(model_dir))

2. 使用tf.train.Checkpoint

代码如下

step_counter = tf.train.get_or_create_global_step()
checkpoint = tf.train.Checkpoint(model=model, optimizer=optimizer, step_counter=step_counter)

def saveModelV2(model_dir, checkpoint, modelname='model2'):
    checkpoint_prefix = os.path.join(model_dir, modelname)
    checkpoint.save(checkpoint_prefix)

def restoreModelV2(model_dir, checkpoint):
    checkpoint.restore(tf.train.latest_checkpoint(model_dir))

五 具体代码

代码未严格按照总体流程的步骤，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()，否则会报错。具体可参考代码。