TensorFlow自定义层创建报错：递归深度超出限制

问题详情 #

在TensorFlow 2.17.0版本中实现自定义全连接层时，出现如下递归深度超出错误：

File /opt/anaconda3/lib/python3.11/site-packages/tensorflow/python/util/lazy_loader.py:178 in __getattr__
    if item in ("_tfll_mode", "_tfll_initialized", "_tfll_name"):

RecursionError: maximum recursion depth exceeded in comparison

对应的实现代码如下：

import tensorflow as tf

class MyDenseLayer(tf.keras.layers.Layer):
    def __init__(self, num_outputs):
        super(MyDenseLayer, self).__init__()
        self.num_outputs = num_outputs

    def build(self, input_shape):
        self.kernel = self.add_weight(
            "kernel",
            shape=[int(input_shape[-1]), self.num_outputs],
            initializer='random_normal',  # Initialize with random values
            trainable=True
        )

    def call(self, inputs):
        return tf.matmul(inputs, self.kernel)


input_dim = 4  # Number of input features
output_dim = 2  # Match the target data output dimension
layer = MyDenseLayer(output_dim)

# Create a simple model
model = tf.keras.Sequential([
    layer,
    tf.keras.layers.Dense(2)  # Ensure the final output matches the target
])

# Compile the model
model.compile(optimizer='adam', loss='mse')

# Create random input data
x = tf.random.normal((5, input_dim))  # Batch size of 5
y = tf.random.normal((5, 2))  # Random target data for training

# Train the model
model.fit(x, y, epochs=10)

解决方案 #

该错误源于TensorFlow LazyLoader模块的递归引用问题，可通过以下几种方式修复：

1. 简化父类初始化调用 #

将__init__方法中的父类初始化语句改为Python 3简化写法，避免显式传递类名和self引发的递归：

def __init__(self, num_outputs):
    super().__init__()
    self.num_outputs = num_outputs

2. 避免提前实例化自定义层 #

不要先实例化自定义层再传入Sequential，直接在模型定义中声明层类：

model = tf.keras.Sequential([
    MyDenseLayer(output_dim),
    tf.keras.layers.Dense(2)
])

提前实例化的层会在模型构建阶段触发不必要的属性访问递归。

3. 显式指定dtype（可选） #

在__init__中添加dtype参数并传递给父类，避免类型推断时触发LazyLoader递归：

def __init__(self, num_outputs, dtype=None):
    super().__init__(dtype=dtype)
    self.num_outputs = num_outputs

修复后的完整代码 #

import tensorflow as tf

class MyDenseLayer(tf.keras.layers.Layer):
    def __init__(self, num_outputs):
        super().__init__()
        self.num_outputs = num_outputs

    def build(self, input_shape):
        self.kernel = self.add_weight(
            "kernel",
            shape=[int(input_shape[-1]), self.num_outputs],
            initializer='random_normal',
            trainable=True
        )

    def call(self, inputs):
        return tf.matmul(inputs, self.kernel)


input_dim = 4
output_dim = 2

model = tf.keras.Sequential([
    MyDenseLayer(output_dim),
    tf.keras.layers.Dense(2)
])

model.compile(optimizer='adam', loss='mse')

x = tf.random.normal((5, input_dim))
y = tf.random.normal((5, 2))

model.fit(x, y, epochs=10)

内容的提问来源于stack exchange，提问作者tew_evolving