使用深度神经网络训练手写数字识别

下载数据

import tensorflow as tf
from tensorflow import keras

(X_train_full, y_train_full), (X_test, y_test) = keras.datasets.mnist.load_data()

数据处理

X_valid, X_train = X_train_full[:5000] / 255., X_train_full[5000:] / 255.
y_valid, y_train = y_train_full[:5000], y_train_full[5000:]

简单查看数据

import matplotlib.pyplot as plt
n_rows = 4
n_cols = 10
plt.figure(figsize=(n_cols * 1.2, n_rows * 1.2))
for row in range(n_rows):
    for col in range(n_cols):
        index = n_cols * row + col
        plt.subplot(n_rows, n_cols, index + 1)#要生成4行10列，这是第index + 1个图
        plt.imshow(X_train[index], cmap="binary", interpolation="nearest")
        plt.axis('off')
        plt.title(y_train[index], fontsize=12)
plt.subplots_adjust(wspace=0.2, hspace=0.5)#调整子图布局
plt.show()

寻找合适的学习率

#设置回调函数，每个数据训练轮次，将学习率*factor
K = keras.backend

class ExponentialLearningRate(keras.callbacks.Callback):
    def __init__(self, factor):
        self.factor = factor
        self.rates = []
        self.losses = []
    def on_batch_end(self, batch, logs):
        self.rates.append(K.get_value(self.model.optimizer.lr))
        self.losses.append(logs["loss"])
        K.set_value(self.model.optimizer.lr, self.model.optimizer.lr * self.factor)

import numpy as np
keras.backend.clear_session()
np.random.seed(42)
tf.random.set_seed(42)

#设置神经网络参数
model = keras.models.Sequential([
    keras.layers.Flatten(input_shape=[28, 28]),
    keras.layers.Dense(300, activation="relu"),
    keras.layers.Dense(100, activation="relu"),
    keras.layers.Dense(10, activation="softmax")
])

model.compile(loss="sparse_categorical_crossentropy",
              optimizer=keras.optimizers.SGD(learning_rate=1e-3),
              metrics=["accuracy"])
expon_lr = ExponentialLearningRate(factor=1.005)

训练模型找出最优学习率

history = model.fit(X_train, y_train, epochs=1,
                    validation_data=(X_valid, y_valid),
                    callbacks=[expon_lr])

plt.plot(expon_lr.rates, expon_lr.losses)
plt.gca().set_xscale('log')#设置为对数坐标
plt.hlines(min(expon_lr.losses), min(expon_lr.rates), max(expon_lr.rates))#绘制水平线，即绘制最小值的水平线
plt.axis([min(expon_lr.rates), max(expon_lr.rates), 0, expon_lr.losses[0]])#设置XY轴范围
plt.grid()#显示网格线
plt.xlabel("Learning rate")
plt.ylabel("Loss")

可以看到，学习率在0.6之后就变动较大，因此最优学习率取一半即0.3

训练神经网络

keras.backend.clear_session()
np.random.seed(42)
tf.random.set_seed(42)

model = keras.models.Sequential([
    keras.layers.Flatten(input_shape=[28, 28]),
    keras.layers.Dense(300, activation="relu"),
    keras.layers.Dense(100, activation="relu"),
    keras.layers.Dense(10, activation="softmax")
])

model.compile(loss="sparse_categorical_crossentropy",
              optimizer=keras.optimizers.SGD(lr=3e-1),
              metrics=["accuracy"])

#设置保 存 路 径
import os
run_index = 1 # increment this at every run
run_logdir = os.path.join(os.curdir, "my_mnist_logs", "run_{:03d}".format(run_index))
run_logdir

开始训练

early_stopping_cb = keras.callbacks.EarlyStopping(patience=20)#patience: 没有进步的训练轮数，在这之后训练就会被停止
checkpoint_cb = keras.callbacks.ModelCheckpoint("my_mnist_model.h5", save_best_only=True)#在每个训练期之后保存模型。
tensorboard_cb = keras.callbacks.TensorBoard(run_logdir)#这个回调函数为 Tensorboard 编写一个日志

history = model.fit(X_train, y_train, epochs=100,
                    validation_data=(X_valid, y_valid),
                    callbacks=[checkpoint_cb, early_stopping_cb, tensorboard_cb])

在测试集上评估

model = keras.models.load_model("my_mnist_model.h5") # rollback to best model
model.evaluate(X_test, y_test)