TensorFlow2神经网络训练

文章目录

TensorFlow2神经网络训练

梯度下降
反向传播
训练可视化
补充说明

梯度下降

梯度 $\nabla f=\left(\frac{\partial f}{\partial x_{1}} ; \frac{\partial f}{\partial x_{2}} ; \ldots ; \frac{\partial f}{\partial x_{n}}\right)$ 指函数关于变量x的导数，梯度的方向表示函数值增大的方向，梯度的模表示函数值增大的速率。那么只要不断将参数的值向着梯度的反方向更新一定大小，就能得到函数的最小值（全局最小值或者局部最小值）。
$\theta_{t+1}=\theta_{t}-\alpha_{t} \nabla f\left(\theta_{t}\right)$
上述参数更新的过程就叫做梯度下降法，但是一般利用梯度更新参数时会将梯度乘以一个小于1的学习速率（learning rate），这是因为往往梯度的模还是比较大的，直接用其更新参数会使得函数值不断波动，很难收敛到一个平衡点（这也是学习率不宜过大的原因）。
但是对于不同的函数，GD（梯度下降法）未必都能找到最优解，很多时候它只能收敛到一个局部最优解就不再变动了（尽管这个局部最优解已经很接近全局最优解了），这是函数性质决定的，实验证明，梯度下降法对于凸函数有着较好的表现。
TensorFlow和PyTorch这类深度学习框架是支持自动梯度求解的，在TensorFlow2中只要将需要进行梯度求解的代码段包裹在GradientTape中，TensorFlow就会自动求解相关运算的梯度。但是通过tape.gradient(loss, [w1, w2, ...])只能调用一次，梯度作为占用显存较大的资源在被获取一次后就会被释放掉，要想多次调用需要设置tf.GradientTape(persistent=True)（此时注意及时释放资源）。TensorFlow2也支持多阶求导，只要将求导进行多层包裹即可。示例如下。

反向传播

反向传播算法（BP）是训练深度神经网络的核心算法，它的实现是基于链式法则的。将输出层的loss通过权值反向传播(前向传播的逆运算)回第i层（这是个反复迭代返回的过程），计算i层的梯度更新参数。具体原理见之前的BP神经网络博客。

在TensorFlow2中，对于经典的BP神经网络层进行了封装，称为全连接层，自动完成BP神经网络隐层的操作。下面为使用Dense层构建BP神经网络训练Fashion_MNIST数据集进行识别的代码。

"""
Author: Zhou Chen
Date: 2019/10/15
Desc: About
"""
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import datasets, layers, optimizers, Sequential, metrics

(x, y), (x_test, y_test) = datasets.fashion_mnist.load_data()
print(x.shape, y.shape)


def preprocess(x, y):
    x = tf.cast(x, dtype=tf.float32) / 255.
    y = tf.cast(y, dtype=tf.int32)
    return x, y


batch_size = 64
db = tf.data.Dataset.from_tensor_slices((x, y))
db = db.map(preprocess).shuffle(10000).batch(batch_size)
db_test = tf.data.Dataset.from_tensor_slices((x_test, y_test))
db_test = db_test.map(preprocess).shuffle(10000).batch(batch_size)

model = Sequential([
    layers.Dense(256, activation=tf.nn.relu),  # [b, 784] => [b, 256]
    layers.Dense(128, activation=tf.nn.relu),  # [b, 256] => [b, 128]
    layers.Dense(64, activation=tf.nn.relu),  # [b, 128] => [b, 64]
    layers.Dense(32, activation=tf.nn.relu),  # [b, 64] => [b, 32]
    layers.Dense(10),  # [b, 32] => [b, 10]
])
model.build(input_shape=([None, 28*28]))
optimizer = optimizers.Adam(lr=1e-3)


def main():
    # forward
    for epoch in range(30):
        for step, (x, y) in enumerate(db):
            x = tf.reshape(x, [-1, 28*28])
            with tf.GradientTape() as tape:
                logits = model(x)
                y_onthot = tf.one_hot(y, depth=10)

                loss_mse = tf.reduce_mean(tf.losses.MSE(y_onthot, logits))
                loss_ce = tf.reduce_mean(tf.losses.categorical_crossentropy(y_onthot, logits, from_logits=True))
            grads = tape.gradient(loss_ce, model.trainable_variables)
            # backward
            optimizer.apply_gradients(zip(grads, model.trainable_variables))

            if step % 100 == 0:
                print(epoch, step, "loss:", float(loss_mse), float(loss_ce))
        # test
        total_correct, total_num = 0, 0
        for x, y in db_test:
            x = tf.reshape(x, [-1, 28*28])
            logits = model(x)
            prob = tf.nn.softmax(logits, axis=1)
            pred = tf.cast(tf.argmax(prob, axis=1), dtype=tf.int32)
            correct = tf.reduce_sum(tf.cast(tf.equal(pred, y), dtype=tf.int32))
            total_correct += int(correct)
            total_num += int(x.shape[0])
        acc = total_correct / total_num
        print("acc", acc)


if __name__ == '__main__':
    main()

训练可视化

TensorFlow有一套伴生的可视化工具包TensorBoard（使用pip安装，最新版本的TensorFlow会自动安装TensorBoard），它是基于Web端的方便监控训练过程和训练数据的工具，监控数据来源于本地磁盘指定的一个目录。一般使用TensorBoard需要三步，创建log目录，创建summary实例，指定数据给summary实例。
tensorboard --logdir logs监听设定的log目录，此时由于并没有写入文件，所以显示如下。

后面两步一般在训练过程中嵌入，示例如下。（注意，TensorBoard并没有设计组合多个sample图片而是一个个显示，组合需要自己写接口，下面的代码就写了这个接口。）

"""
Author: Zhou Chen
Date: 2019/10/15
Desc: About
"""
import tensorflow as tf
from tensorflow.keras import datasets, layers, optimizers, Sequential, metrics
import datetime
from matplotlib import pyplot as plt
import io


def preprocess(x, y):
    x = tf.cast(x, dtype=tf.float32) / 255.
    y = tf.cast(y, dtype=tf.int32)

    return x, y


def plot_to_image(figure):
    # Save the plot to a PNG in memory.
    buf = io.BytesIO()
    plt.savefig(buf, format='png')
    # Closing the figure prevents it from being displayed directly inside the notebook.
    plt.close(figure)
    buf.seek(0)
    # Convert PNG buffer to TF image
    image = tf.image.decode_png(buf.getvalue(), channels=4)
    # Add the batch dimension
    image = tf.expand_dims(image, 0)
    return image


def image_grid(images):
    """
    Return a 5x5 grid of the MNIST images as a matplotlib figure.
    """
    # Create a figure to contain the plot.
    figure = plt.figure(figsize=(10, 10))
    for i in range(25):
        # Start next subplot.
        plt.subplot(5, 5, i + 1, title='name')
        plt.xticks([])
        plt.yticks([])
        plt.grid(False)
        plt.imshow(images[i], cmap=plt.cm.binary)

    return figure


batchsz = 128
(x, y), (x_val, y_val) = datasets.mnist.load_data()
print('datasets:', x.shape, y.shape, x.min(), x.max())

db = tf.data.Dataset.from_tensor_slices((x, y))
db = db.map(preprocess).shuffle(60000).batch(batchsz).repeat(10)

ds_val = tf.data.Dataset.from_tensor_slices((x_val, y_val))
ds_val = ds_val.map(preprocess).batch(batchsz, drop_remainder=True)

network = Sequential([layers.Dense(256, activation='relu'),
                    layers.Dense(128, activation='relu'),
                    layers.Dense(64, activation='relu'),
                    layers.Dense(32, activation='relu'),
                    layers.Dense(10)])
network.build(input_shape=(None, 28 * 28))
network.summary()

optimizer = optimizers.Adam(lr=0.01)

current_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
log_dir = 'logs/' + current_time
summary_writer = tf.summary.create_file_writer(log_dir)

# get x from (x,y)
sample_img = next(iter(db))[0]
# get first image instance
sample_img = sample_img[0]
sample_img = tf.reshape(sample_img, [1, 28, 28, 1])
with summary_writer.as_default():
    tf.summary.image("Training sample:", sample_img, step=0)

for step, (x, y) in enumerate(db):

    with tf.GradientTape() as tape:
        # [b, 28, 28] => [b, 784]
        x = tf.reshape(x, (-1, 28 * 28))
        # [b, 784] => [b, 10]
        out = network(x)
        # [b] => [b, 10]
        y_onehot = tf.one_hot(y, depth=10)
        # [b]
        loss = tf.reduce_mean(tf.losses.categorical_crossentropy(y_onehot, out, from_logits=True))

    grads = tape.gradient(loss, network.trainable_variables)
    optimizer.apply_gradients(zip(grads, network.trainable_variables))

    if step % 100 == 0:
        print(step, 'loss:', float(loss))
        with summary_writer.as_default():
            tf.summary.scalar('train-loss', float(loss), step=step)

            # evaluate
    if step % 500 == 0:
        total, total_correct = 0., 0

        for _, (x, y) in enumerate(ds_val):
            # [b, 28, 28] => [b, 784]
            x = tf.reshape(x, (-1, 28 * 28))
            # [b, 784] => [b, 10]
            out = network(x)
            # [b, 10] => [b]
            pred = tf.argmax(out, axis=1)
            pred = tf.cast(pred, dtype=tf.int32)
            # bool type
            correct = tf.equal(pred, y)
            # bool tensor => int tensor => numpy
            total_correct += tf.reduce_sum(tf.cast(correct, dtype=tf.int32)).numpy()
            total += x.shape[0]

        print(step, 'Evaluate Acc:', total_correct / total)

        # print(x.shape)
        val_images = x[:25]
        val_images = tf.reshape(val_images, [-1, 28, 28, 1])
        with summary_writer.as_default():
            tf.summary.scalar('test-acc', float(total_correct / total), step=step)
            tf.summary.image("val-onebyone-images:", val_images, max_outputs=25, step=step)

            val_images = tf.reshape(val_images, [-1, 28, 28])
            figure = image_grid(val_images)
            tf.summary.image('val-images:', plot_to_image(figure), step=step)