TensorFlow强化学习2D卷积Actor模型使用PReLU时出现维度不匹配错误的问题咨询

你遇到的这个维度不匹配错误，核心原因是重复使用了同一个PReLU层实例，而TensorFlow的PReLU层默认会为输入的每个空间位置+通道创建独立的可学习alpha参数。

当你第一次在Conv1中使用这个PReLU实例时，输入的形状是(None, 1000, 4000, 32)，此时PReLU会初始化一个形状为(1000, 4000, 32)的alpha参数张量。但后续经过Conv_pool_1的池化操作后，Batchnorm_1的输出形状变成了(None, 500, 2000, 32)，此时再用同一个PReLU实例处理这个输入，就会出现alpha参数的空间维度（1000×4000）和输入的空间维度（500×2000）不匹配的问题——这就是报错里提到的"Dimensions must be equal, but are 1000 and 500"的根源。

而LeakyReLU没有这个问题，因为它的alpha是固定的标量，不需要为每个输入元素/通道学习参数，所以复用实例完全没问题。

方案1：每次使用时创建新的PReLU实例 #

不要提前定义一个全局的activation变量，而是在每个Conv2D层的activation参数中直接创建新的PReLU层，这样每个层都会有自己独立的alpha参数，适配对应输入的形状：

from tensorflow.keras.layers import Conv2D, Input, MaxPool1D, concatenate, Lambda, Dense, Flatten
import tensorflow as tf

def ActorNetwork(input_shape_A,input_shape_B, n_actions):
    input_layer_A = Input(shape=input_shape_A[1:], name="input_layer_A")
    input_layer_B = Input(shape=input_shape_B[1:], name="input_layer_B")
    Rescale = Lambda(lambda x: tf.divide(tf.subtract(x, tf.reduce_max(x)), tf.subtract(tf.reduce_max(x), tf.reduce_min(x))))(input_layer_A)
    # 每个Conv层都创建新的PReLU实例
    Conv1 = Conv2D(32, 3, activation=tf.keras.layers.PReLU(alpha_initializer=tf.initializers.constant(0.25)), padding='same', name="Conv1")(Rescale)
    Conv2 = Conv2D(32, 3, activation=tf.keras.layers.PReLU(alpha_initializer=tf.initializers.constant(0.25)), padding='same', name="Conv2")(Conv1)
    Conv_pool_1 = Conv2D(32, 2, strides=2, activation='relu', padding='same', name="Conv_pool_1")(Conv2)
    Batchnorm_1 = tf.keras.layers.BatchNormalization(name='Batchnorm_1')(Conv_pool_1)
    Conv3 = Conv2D(32, 3, activation=tf.keras.layers.PReLU(alpha_initializer=tf.initializers.constant(0.25)), padding='same', name="Conv3")(Batchnorm_1)
    Conv4 = Conv2D(32, 3, activation=tf.keras.layers.PReLU(alpha_initializer=tf.initializers.constant(0.25)), padding='same', name="Conv4")(Conv3)
    Conv_pool_2 = Conv2D(32, 2, strides=2, activation='relu', padding='same', name="Conv_pool_2")(Conv4)
    Batchnorm_2 = tf.keras.layers.BatchNormalization(name='Batchnorm_2')(Conv_pool_2)
    Conv5 = Conv2D(64, 3, activation=tf.keras.layers.PReLU(alpha_initializer=tf.initializers.constant(0.25)), padding='same', name="Conv5")(Batchnorm_2)
    Conv6 = Conv2D(64, 3, activation=tf.keras.layers.PReLU(alpha_initializer=tf.initializers.constant(0.25)), padding='same', name="Conv6")(Conv5)
    Conv_pool_3 = Conv2D(64, 2, strides=2, activation='relu', padding='same', name="Conv_pool_3")(Conv6)
    Batchnorm_3 = tf.keras.layers.BatchNormalization(name='Batchnorm_3')(Conv_pool_3)
    Conv7 = Conv2D(64, 3, activation=tf.keras.layers.PReLU(alpha_initializer=tf.initializers.constant(0.25)), padding='same', name="Conv7")(Batchnorm_3)
    Conv8 = Conv2D(64, 3, activation=tf.keras.layers.PReLU(alpha_initializer=tf.initializers.constant(0.25)), padding='same', name="Conv8")(Conv7)
    Conv_pool_4 = Conv2D(64, 2, strides=2, activation='relu', padding='same', name="Conv_pool_4")(Conv8)
    Batchnorm_4 = tf.keras.layers.BatchNormalization(name='Batchnorm_4')(Conv_pool_4)
    Conv9 = Conv2D(128, 3, activation=tf.keras.layers.PReLU(alpha_initializer=tf.initializers.constant(0.25)), padding='same', name="Conv9")(Batchnorm_4)
    Conv10 = Conv2D(128, 3, activation=tf.keras.layers.PReLU(alpha_initializer=tf.initializers.constant(0.25)), padding='same', name="Conv10")(Conv9)
    Conv_pool_5 = Conv2D(128, 2, strides=2, activation='relu', padding='same', name="Conv_pool_5")(Conv10)
    Batchnorm_5 = tf.keras.layers.BatchNormalization(name='Batchnorm_5')(Conv_pool_5)
    Conv11 = Conv2D(128, 3, activation=tf.keras.layers.PReLU(alpha_initializer=tf.initializers.constant(0.25)), padding='same', name="Conv11")(Batchnorm_5)
    Conv12 = Conv2D(128, 3, activation=tf.keras.layers.PReLU(alpha_initializer=tf.initializers.constant(0.25)), padding='same', name="Conv12")(Conv11)
    Conv_pool_6 = Conv2D(128, 2, strides=2, activation='relu', padding='same', name="Conv_pool_6")(Conv12)
    Batchnorm_6 = tf.keras.layers.BatchNormalization(name='Batchnorm_6')(Conv_pool_6)
    Conv_pool_7 = Conv2D(128, 1, strides=1, activation='relu', padding='same', name="Conv_pool_7")(Batchnorm_6)
    Conv_pool_8 = Conv2D(64, 1, strides=1, activation='relu', padding='same', name="Conv_pool_8")(Conv_pool_7)
    Conv_pool_9 = Conv2D(32, 1, strides=1, activation='relu', padding='same', name="Conv_pool_9")(Conv_pool_8)
    flatten = Flatten()(Conv_pool_9)
    Concat_2 = tf.keras.layers.concatenate([flatten, input_layer_B], axis=-1,name='Concat_2')
    fc1 = Dense(8194, activation='relu', name="fc1")(Concat_2)
    fc2 = Dense(4096, activation='relu', name="fc2")(fc1)
    fc3 = Dense(n_actions, activation='softmax', name="fc3")(fc2)
    return tf.keras.models.Model(inputs=[input_layer_A,input_layer_B], outputs = fc3, name="actor_model")

方案2：设置PReLU的shared_axes参数共享空间维度的alpha #

如果你希望复用同一个PReLU实例（减少参数数量），可以设置shared_axes=[1,2]，这样alpha参数会在空间维度（高度和宽度）上共享，只针对每个通道保留一个可学习的alpha值，不管输入的空间形状怎么变化都能适配：

from tensorflow.keras.layers import Conv2D, Input, MaxPool1D, concatenate, Lambda, Dense, Flatten
import tensorflow as tf
# 设置shared_axes=[1,2]，共享空间维度的alpha参数
activation = tf.keras.layers.PReLU(alpha_initializer=tf.initializers.constant(0.25), shared_axes=[1,2])
def ActorNetwork(input_shape_A,input_shape_B, n_actions):
    # 后面的代码和你原来的完全一样即可
    input_layer_A = Input(shape=input_shape_A[1:], name="input_layer_A")
    input_layer_B = Input(shape=input_shape_B[1:], name="input_layer_B")
    Rescale = Lambda(lambda x: tf.divide(tf.subtract(x, tf.reduce_max(x)), tf.subtract(tf.reduce_max(x), tf.reduce_min(x))))(input_layer_A)
    Conv1 = Conv2D(32, 3, activation= activation, padding='same', name="Conv1")(Rescale)
    Conv2 = Conv2D(32, 3, activation=activation, padding='same', name="Conv2")(Conv1)
    Conv_pool_1 = Conv2D(32, 2, strides=2, activation='relu', padding='same', name="Conv_pool_1")(Conv2)
    Batchnorm_1 = tf.keras.layers.BatchNormalization(name='Batchnorm_1')(Conv_pool_1)
    Conv3 = Conv2D(32, 3, activation= activation, padding='same', name="Conv3")(Batchnorm_1)
    # ... 后续代码保持不变

这个方案的优势是减少了模型的参数总量，因为原本每个空间位置+通道都有一个alpha，现在每个通道只有一个alpha，更轻量化。

修改后，你可以重新运行代码，模型应该能正常初始化、编译并运行，不会再出现维度不匹配的错误。

内容的提问来源于stack exchange，提问作者AB Music Box