Rust 特征对象的输入和输出类型不匹配

所以我的问题是，我有一个具有以下输入和输出类型的层特征：

pub trait Layer {
    type Input: Dimension;
    type Output: Dimension;
    
    fn forward(&mut self, input: &ArrayBase<OwnedRepr<f32>, Self::Input>) -> ArrayBase<OwnedRepr<f32>, Self::Output>;
}

使用此转发功能：

impl<A: Activation> Layer for DenseLayer<A> {
    type Input = Ix2;
    type Output = Ix2;

    fn forward(&mut self, input: &Array2<f32>) -> Array2<f32> {
        assert_eq!(input.shape()[1], self.weights.shape()[0], "Input width must match weight height.");
        let z = input.dot(&self.weights) + &self.biases;

        self.activation.activate(&z)
    }
}

我有这些，以便我的前向或后向函数可以接受例如 2 维数组，但仍然输出只有 1 维的数组。然后我有一个这种层特征的包装器的实现，我希望通过所有层进行转发：

pub struct NeuralNetwork<'a, L>
where
    L: Layer + 'a,
{
    layers: Vec<L>,
    loss_function: &'a dyn Cost,
}

impl<'a, L> NeuralNetwork<'a, L>
where
    L: Layer + 'a,
{
    pub fn new(layers: Vec<L>, loss_function: &'a dyn Cost) -> Self {
        NeuralNetwork { layers, loss_function }
    }

    pub fn forward(&mut self, input: &ArrayBase<OwnedRepr<f32>, L::Input>) -> ArrayBase<OwnedRepr<f32>, L::Output> {
        let mut output = input.clone();

        // todo fix the layer forward changing input to output
        // causing mismatch in the input and output dimensions of forward
        for layer in &mut self.layers {
            output = layer.forward(&output);
        }

        output
    }
}

现在因为在 for 循环中我首先输入了 input 类型，然后从 layer.forward 接收输出。在下一次迭代中，它采用 output 类型，但 layer.forward 只接受 input 类型。至少我认为这是正在发生的事情。这似乎是一个非常简单的问题，但我真的不确定如何解决这个问题。

编辑1：

可重现的示例：

use ndarray::{Array, Array2, ArrayBase, Dimension, OwnedRepr};

pub trait Layer {
    type Input: Dimension;
    type Output: Dimension;

    fn forward(&mut self, input: &ArrayBase<OwnedRepr<f32>, Self::Input>) -> ArrayBase<OwnedRepr<f32>, Self::Output>;
}

// A Dense Layer struct
pub struct DenseLayer {
    weights: Array2<f32>,
    biases: Array2<f32>,
}

impl DenseLayer {
    pub fn new(input_size: usize, output_size: usize) -> Self {
        let weights = Array::random((input_size, output_size), rand::distributions::Uniform::new(-0.5, 0.5));
        let biases = Array::zeros((1, output_size));
        DenseLayer { weights, biases }
    }
}

impl Layer for DenseLayer {
    type Input = ndarray::Ix2;  // Two-dimensional input
    type Output = ndarray::Ix2; // Two-dimensional output

    fn forward(&mut self, input: &ArrayBase<OwnedRepr<f32>, Self::Input>) -> ArrayBase<OwnedRepr<f32>, Self::Output> {
        assert_eq!(input.shape()[1], self.weights.shape()[0], "Input width must match weight height.");
        let z = input.dot(&self.weights) + &self.biases;
        z // Return the output directly without activation
    }
}

// Neural Network struct
pub struct NeuralNetwork<'a, L>
where
    L: Layer + 'a,
{
    layers: Vec<L>,
}

impl<'a, L> NeuralNetwork<'a, L>
where
    L: Layer + 'a,
{
    pub fn new(layers: Vec<L>) -> Self {
        NeuralNetwork { layers }
    }

    pub fn forward(&mut self, input: &ArrayBase<OwnedRepr<f32>, L::Input>) -> ArrayBase<OwnedRepr<f32>, L::Output> {
        let mut output = input.clone();

        for layer in &mut self.layers {
            output = layer.forward(&output);
        }

        output
    }
}

fn main() {
    // Create a neural network with one Dense Layer
    let mut dense_layer = DenseLayer::new(3, 2);
    let mut nn = NeuralNetwork::new(vec![dense_layer]);

    // Create an example input (1 batch, 3 features)
    let input = Array::from_shape_vec((1, 3), vec![1.0, 2.0, 3.0]).unwrap();
    
    // Forward pass
    let output = nn.forward(&input);
    println!("Output: {:?}", output);
}