Keras 实现多层感知机

这里以 mnist 数据集为例，代码如下：

import  tensorflow as tf
from    tensorflow.keras import datasets, layers, optimizers, Sequential, metrics

# 设置GPU使用方式
# 获取GPU列表
gpus = tf.config.experimental.list_physical_devices('GPU')
if gpus:
    try:
        for gpu in gpus:
            # 设置GPU为增长式占用
            tf.config.experimental.set_memory_growth(gpu, True)
    except RuntimeError as e:
        # 打印异常
        print(e)

# 导入数据
(x_train, y_train),(x_val, y_val) = datasets.mnist.load_data()
print('datasets:', x_train.shape, y_train.shape, x_train.min(), x_train.max())

# 将样本属性转换为张量
x_train = tf.convert_to_tensor(x_train, dtype=tf.float32) / 255.
x_val = tf.convert_to_tensor(x_val, dtype=tf.float32) / 255.

# 每批次的样本个数
batch_size = 512

# 模型迭代次数
epochs = 100

# 序列模型 Sequential 适用于每层只有一个输入张量和一个输出张量的简单层堆栈
model = Sequential([layers.Dense(256, activation='relu'),
                     layers.Dense(128, activation='relu'),
                     layers.Dense(10)]) 

# input_shape 为输入层的形状参数 None 代表任意批次 28* 28 代表输入参数维度
model.build(input_shape=(None, 28*28))

# 序列模型信息打印
model.summary()

# Keras 中的一种训练方式
x_train = tf.reshape(x_train, (-1, 28*28))
x_val = tf.reshape(x_val, (-1, 28*28))
y_train = tf.one_hot(y_train, depth=10)
y_val = tf.one_hot(y_val, depth=10)

model.compile(
    optimizer = tf.keras.optimizers.Adam(lr=0.01),
    loss = tf.keras.losses.CategoricalCrossentropy(from_logits=True),
    metrics=[tf.keras.metrics.CategoricalAccuracy()])

history = model.fit(x_train,y_train,batch_size,epochs,validation_data = (x_val, y_val))

x = [i for i in range(0, epochs)]

plt.figure()
plt.plot(x, history.history["categorical_accuracy"], color='blue', label='vaildation')
plt.plot(x, history.history["val_categorical_accuracy"], color='red', label='training')
plt.xlabel('Epoch')
plt.ylabel('Accuracy')
plt.legend()
plt.show()
plt.close()

训练曲线如下：