Keras 层 API（第一部分）

Keras 中的全连接层

1import tensorflow as tf
2from tensorflow.keras import layers
3
4dense_layer = layers.Dense(8, activation='relu')

1import tensorflow as tf
2from tensorflow.keras import layers
3
4dense_layer = layers.Dense(8, activation='relu')
5
6# 从均匀分布中生成随机值
7input_tensor = tf.random.uniform(shape=(3, 5), minval=0, maxval=10, dtype=tf.int32)
8output_tensor = dense_layer(input_tensor)
9print('\n\n *** 输入张量 *** \n\n', input_tensor, 
10      '\n\n *** 输出张量 *** \n\n', output_tensor, 
11      '\n\n *** 层权重 *** \n\n', dense_layer.weights)

使用正态分布生成权重的自定义 Keras 层初始化器

1custom_initializer = tf.keras.initializers.RandomNormal(mean=0., stddev=1.)
2dense_layer = tf.keras.layers.Dense(10, kernel_initializer=custom_initializer)

Keras 激活函数

1import tensorflow as tf
2from tensorflow import keras
3from tensorflow.keras import activations
4
5input_tensor = tf.constant([[-10.0, -5.0, 0.0, 5.0, 10.0],
6                            [-1.0, -0.5, 0.0, 0.5, 1.0]], dtype=tf.float32)
7
8# 将 Keras 激活函数应用于输入张量
9sig_func = keras.activations.sigmoid(input_tensor)      
10tanh_func = tf.keras.activations.tanh(input_tensor)
11relu_func = keras.activations.relu(input_tensor, alpha=0.1)   # alpha 控制小于 0 时的斜率
12softmax_func = keras.activations.softmax(input_tensor, axis=-1)  # softmax 沿指定轴计算，-1 表示最后一个轴
13
14print('\n\n Sigmoid 输出: \n', sig_func.numpy(),
15      '\n\n Tanh 输出: \n', tanh_func.numpy(),
16      '\n\n ReLU 输出: \n', relu_func.numpy(), 
17      '\n\n Softmax 输出: \n', softmax_func.numpy())

1import tensorflow as tf
2from tensorflow import keras
3from tensorflow.keras import layers
4
5# 输入为大小为 8 的批处理，图像尺寸为 32×32×3 的 RGB 图像
6input_dimensions = (8, 32, 32, 3)
7input_tensor = tf.random.normal(input_dimensions)
8output_conv2D = tf.keras.layers.Conv2D(filters=10, kernel_size=(5, 5), 
9                                       activation='relu', 
10                                       # 取消注释下一行可使输出与输入具有相同的空间尺寸
11                                       #padding = 'same',    
12                                       input_shape=input_dimensions[1:])(input_tensor)
13print('\nConv2D 层输出的形状为', output_conv2D.shape)
14
15# 同样输入条件下
16output_dw_conv2D = tf.keras.layers.DepthwiseConv2D(kernel_size=(5, 5), 
17                                                   activation='relu', 
18                                                   # 取消注释下一行可保持输入输出尺寸一致
19                                                   #padding = 'same',   
20                                                   input_shape=input_dimensions[1:])(input_tensor)
21print('\nDepthwiseConv2D 层输出的形状为', output_dw_conv2D.shape)
22
23# 同样输入条件下
24output_sep_conv2D = tf.keras.layers.SeparableConv2D(filters=10, kernel_size=(5, 5), 
25                                                    activation='relu', 
26                                                    # 取消注释下一行可保持输入输出尺寸一致
27                                                    #padding = 'same',   
28                                                    input_shape=input_dimensions[1:])(input_tensor)
29print('\nSeparableConv2D 层输出的形状为', output_sep_conv2D.shape)
30
31# 输入为大小为 8 的批处理，体积数据尺寸为 32×32×32×1 的单通道视频帧
32input_dimensions = (8, 32, 32, 32, 1)
33input_tensor = tf.random.normal(input_dimensions)
34output_conv3D = tf.keras.layers.Conv3D(filters=10, kernel_size=(5, 5, 5), 
35                                       activation='relu',
36                                       # 取消注释下一行可使输出与输入空间尺寸一致
37                                       #padding = 'same',  
38                                       input_shape=input_dimensions[1:])(input_tensor)
39print('\nConv3D 层输出的形状为', output_conv3D.shape)