完全传播的层代码的反向传播

Question

大家好，我正在实现神经网络库。当我尝试对全连接层的反向传播进行编码时，偶然发现一个问题-当我在前馈过程中相乘时， 我“丢失”了输入尺寸。举例来说，如果我的输入形状是矩阵形状的10x640，而权重形状是10x640x1000，则加权的输出形状将是10x1000（这应该是预期的）。但是，这意味着当我尝试在反向传播过程中计算下一层（或上一层）误差时，我将无法无错误地进行乘法运算。这是我的代码：

# Feedforward
# input.shape = 10x640
# weights.shape = 10x640x1000
# z.shape = 10x1000
dots = []
for i in range(len(input)):
    dots.append(np.dot(input[i], weights[i]))
z = np.add(dots, biases)

# Backpropagation
# thisLayerError.shape = 10x1000
# activationDerivative.shape = 10x1000
# dots.shape = 10x640
activationDerivative = activationFunction.derivative(z) 
dots = []
for i in range(len(thisLayerError)):
    dots.append(np.dot(weights[i], thisLayerError[i])) 
nextLayerError = np.multiply(dots, activationDerviative) # this will throw an exception