Cython无法正常访问Python共享类中的权重数据问题

问题总结 #

两个Cython函数以相同方式访问Python共享类的WeightVals.Weights属性，forward_propagate能正常获取到权重数据，但back_propagate中该属性返回空列表[]，即使修改导入语句后问题仍存在。

可能原因 #

1. 类属性重复访问的优化干扰：back_propagate循环中直接通过WeightVals.Weights[layer].T访问类属性，Cython的编译优化可能缓存该属性的初始空值，导致后续访问无法获取更新后的数据。
2. 类属性初始化时机问题：WeightVals.Weights可能仅在forward_propagate调用时被赋值，但back_propagate调用时该属性的更新未被Cython正确感知；或存在其他代码路径意外将其重置为空列表。
3. Cython静态类型推断限制：Cython对全局类属性的访问，可能因静态类型推断无法正确追踪其动态变化的内容。

解决方案 #

1. 统一使用局部变量缓存类属性 #

在back_propagate中已将weights = WeightVals.Weights赋值给局部变量，后续循环应直接使用该变量，避免重复访问类属性：

cpdef np.ndarray[np.float64_t, ndim=1] back_propagate(np.ndarray[np.float64_t, ndim=1] loss,  list activation_derv_list ):
    derv = DerivativeVals.Derivatives
    neurons = NeuronVals.Neurons
    weights = WeightVals.Weights

    # 调试：先确认weights是否为空
    if not weights:
        print("Warning: Weights is empty!")
        return loss

    cdef np.ndarray[np.float64_t, ndim = 1] delta
    cdef np.ndarray[np.float64_t, ndim = 2] delta_reshape, current_reshaped
    for layer in reversed(range(len(derv))):
        delta = loss * activation_derv_list[layer](neurons[layer+1])
        delta_reshape = delta.reshape(delta.shape[0], -1).T
        current_reshaped = neurons[layer].reshape(neurons[layer].shape[0], -1)
        derv[layer] = np.dot(current_reshaped, delta_reshape)
        # 使用局部变量weights，而非直接访问WeightVals.Weights
        loss = np.dot(delta, weights[layer].T)
        print("check")

2. 显式确认类属性的初始化状态 #

在back_propagate开头添加调试代码，验证调用时WeightVals.Weights的实际值：

cpdef np.ndarray[np.float64_t, ndim=1] back_propagate(np.ndarray[np.float64_t, ndim=1] loss,  list activation_derv_list ):
    # 打印当前Weights状态，确认是否已初始化
    print("Current Weights:", WeightVals.Weights)
    derv = DerivativeVals.Derivatives
    neurons = NeuronVals.Neurons
    weights = WeightVals.Weights
    # ... 后续代码

3. 临时禁用Cython优化（调试用） #

若为编译优化导致的缓存问题，可在Cython文件头部添加指令关闭相关优化：

# cython: optimize.unpack_method_calls=False
# cython: boundscheck=True
# cython: wraparound=True

4. 确保类属性为全局共享的类变量 #

检查WeightVals类的定义，确认Weights是类变量而非实例变量，避免因实例化导致属性隔离：

# deeprai/engine/base_layer.py中的WeightVals示例
class WeightVals:
    Weights = []  # 类变量，全局共享
    @classmethod
    def set_weights(cls, new_weights):
        cls.Weights = new_weights

内容的提问来源于stack exchange，提问作者Arrows