在 TensorFlow 2.4.1 中将 PRelu 激活实现为函数

Question

我正在尝试在此处给出的 tensorflow 2.4.1 中实现 PReLU 激活

出现以下错误

ValueError: Variable alpha already exists, disallowed. Did you mean to set reuse=True or reuse=tf.AUTO_REUSE in VarScope?

这是我的代码

def prelu(_x):
  alphas = tf.compat.v1.get_variable('alpha', _x.get_shape()[-1],
                initializer=tf.constant_initializer(0.0), dtype=tf.float32)
  pos = tf.nn.relu(_x)
  neg = alphas * (_x - abs(_x)) * 0.5
  return pos + neg

感谢任何帮助。

注：

我不想使用层接口 tf.keras.layers.PReLU，因为它不能作为参数传递给 conv2D，如下所示

Conv2D(filters, (3, 3), padding='same', activation='prelu')

Answer 1

这是 tensorflow as a function rather than layer which is available as a built-in activation layer and (I think that should be used), PRelu 中的 PRelu 实现。

def prelu_advanced(scope=None):        
    def prelu_plus(x):
        with tf.compat.v1.variable_scope(name_or_scope=scope, 
                                         default_name="prelu", reuse=True):
            alpha = tf.compat.v1.get_variable("prelu", shape=x.get_shape()[-1],
                               dtype=x.dtype, initializer=tf.constant_initializer(0.0))
            pos = tf.nn.relu(x)
            neg = alpha * (x - abs(x)) * 0.5
        return tf.maximum(0.0, x) + alpha * tf.minimum(0.0, x)
    return prelu_plus

检查

foo = tf.constant([1.0, -0.1, -1.0, 0.5, 0.5], dtype = tf.float32)

# layer (built-in)
tf.keras.layers.PReLU(alpha_initializer="zeros", alpha_regularizer=None,
                      alpha_constraint=None, shared_axes=None)(foo).numpy()
array([1. , 0. , 0. , 0.5, 0.5], dtype=float32)

# function 
x = prelu_advanced(scope='prelu1')
x(foo).numpy()
array([1. , 0. , 0. , 0.5, 0.5], dtype=float32)

假人训练

input = tf.keras.Input(shape=(28, 28, 1))
x = tf.keras.layers.Flatten()(input)
x = tf.keras.layers.Dense(128, activation=prelu_advanced(scope='prelu'))(x)
y = tf.keras.layers.Dense(units=10, activation='softmax')(x) 
func_model = tf.keras.Model(inputs=[input], outputs=[y])

func_model.compile(
          loss  = tf.keras.losses.CategoricalCrossentropy(),
          metrics = tf.keras.metrics.CategoricalAccuracy(),
          optimizer = tf.keras.optimizers.Adam())
func_model.fit(x_train, y_train)
4s 2ms/step - loss: 0.2690 - categorical_accuracy: 0.9242

在 TensorFlow 2.4.1 中将 PRelu 激活实现为函数

Implementing PRelu activation as a function in TensorFlow 2.4.1

deep-learning

keras

tensorflow

activation-function