如何为包装NdArray的类型实现iter()？

huangapple
go评论101阅读模式
英文:

How to implement iter() for a type that wraps an NdArray?

问题

以下是代码部分的翻译：

  1. 作为创建基于Rust的张量库的一部分，我已经实现了一个2D的`Tensor`类型：
  3. pub struct Tensor(Rc<RefCell<TensorData>>);
  5. pub struct TensorData {
  6.     pub data: Array2<f64>,
  7.     pub grad: Array2<f64>,
  8.     // 其他字段...
  9. }
  11. impl TensorData {
  12.     fn new(data: Array2<f64>) -> TensorData {
  13.         let shape = data.raw_dim();
  14.         TensorData {
  15.             data,
  16.             grad: Array2::zeros(shape),
  17.             // 其他字段...
  18.         }
  19.     }
  20. }
  22. impl Tensor {
  23.     pub fn new(array: Array2<f64>) -> Tensor {
  24.         Tensor(Rc::new(RefCell::new(TensorData::new(array))))
  25.     }
  27.     pub fn data(&self) -> impl Deref<Target = Array2<f64>> + '_ {
  28.         Ref::map((*self.0).borrow(), |mi| &mi.data)
  29.     }
  30. }

现在我想要能够迭代张量的行（例如，用于实现随机梯度下降）。以下是我目前的实现：

  1. impl Tensor {
  2.     // 其他方法...
  4.     pub fn iter(&self) -> impl Iterator<Item = Tensor> + '_ {
  5.         self.data().outer_iter().map(|el| {
  6.             let reshaped_and_cloned_el = el.into_shape((el.shape()[0], 1)).unwrap().mapv(|el| el.clone());
  7.             reshaped_and_cloned_el
  8.         }).map(|el| Tensor::new(el))
  9.     }
  10. }

这个问题的问题是它在视觉上不友好，并且不会编译，因为迭代器是一个临时值，一旦超出作用域就会被丢弃。

  1. error[E0515]: 无法返回引用临时值
  2.    --> src/tensor/mod.rs:348:9
  3.     |
  4. 348 |           self.data().outer_iter().map(|el| {
  5.     |           ^----------
  6.     |           |
  7.     |  _________temporary value created here
  8.     | |
  9. 349 | |             let reshaped_and_cloned_el = el.into_shape((el.shape()[0], 1)).unwrap().mapv(...
  10. 350 | |             reshaped_and_cloned_el
  11. 351 | |         }).map(|el| Tensor::new(el))
  12.     | |____________________________________^ returns a value referencing data owned by the current function
  13.     |
  14.     = help: use `.collect()` to allocate the iterator

有没有一个不会有这些问题的iter()的替代实现？

英文:

As part of making a Rust-based tensor library, I have implemented a 2D Tensor type as:

  1. use ndarray::prelude::*;
  3. pub struct Tensor(Rc&lt;RefCell&lt;TensorData&gt;&gt;);
  5. pub struct TensorData {
  6.     pub data: Array2&lt;f64&gt;,
  7.     pub grad: Array2&lt;f64&gt;,
  8.     // other fields...
  9. }
  11. impl TensorData {
  12.     fn new(data: Array2&lt;f64&gt;) -&gt; TensorData {
  13.         let shape = data.raw_dim();
  14.         TensorData {
  15.             data,
  16.             grad: Array2::zeros(shape),
  17.             // other fields...
  18.         }
  19.     }
  20. }
  22. impl Tensor {
  23.     pub fn new(array: Array2&lt;f64&gt;) -&gt; Tensor {
  24.         Tensor(Rc::new(RefCell::new(TensorData::new(array))))
  25.     }
  27.     pub fn data(&amp;self) -&gt; impl Deref&lt;Target = Array2&lt;f64&gt;&gt; + &#39;_ {
  28.         Ref::map((*self.0).borrow(), |mi| &amp;mi.data)
  29.     }
  30. }

Now I want to be able to iterate over the rows of a tensor (e.g. for implementing stochastic gradient descent). This is what I have so far:

  1. impl Tensor {
  2. 	// other methods...
  3. 	
  4.     pub fn iter(&amp;self) -&gt; impl Iterator&lt;Item = Tensor&gt; + &#39;_ {
  5.         self.data().outer_iter().map(|el| {
  6.             let reshaped_and_cloned_el = el.into_shape((el.shape()[0], 1)).unwrap().mapv(|el| el.clone());
  7.             reshaped_and_cloned_el
  8.         }).map(|el| Tensor::new(el))
  9.     }
  10. }

The issue with this is that it is (1) unpleasant visually and (2) does not compile as the iterator is a temporary value that is dropped as soon as it goes out of scope:

  1. error[E0515]: cannot return value referencing temporary value
  2.    --&gt; src/tensor/mod.rs:348:9
  3.     |
  4. 348 |           self.data().outer_iter().map(|el| {
  5.     |           ^----------
  6.     |           |
  7.     |  _________temporary value created here
  8.     | |
  9. 349 | |             let reshaped_and_cloned_el = el.into_shape((el.shape()[0], 1)).unwrap().mapv(...
  10. 350 | |             reshaped_and_cloned_el
  11. 351 | |         }).map(|el| Tensor::new(el))
  12.     | |____________________________________^ returns a value referencing data owned by the current function
  13.     |
  14.     = help: use `.collect()` to allocate the iterator

What would be an alternative implementation of iter() that would not have these issues?

答案1

得分: 1

  1. pub fn iter(&amp;self) -&gt; impl Iterator&lt;Item = Tensor&gt; + &#39;_ {
  2.     let data = self.data();
  3.     (0..data.shape()[0]).map(move |i| {
  4.         let el = data.index_axis(Axis(0), i);
  5.         let reshaped_and_cloned_el = el
  6.             .into_shape((el.shape()[0], 1))
  7.             .unwrap()
  8.             .mapv(|el| el.clone());
  9.         Tensor::new(reshaped_and_cloned_el)
  10.     })
  11. }
英文:

Unfortunately, you cannot use the axis iterators, because this will make a self-referential struct (the data guard and the iterator borrowing from it). But you can index to access the axis:

  1. pub fn iter(&amp;self) -&gt; impl Iterator&lt;Item = Tensor&gt; + &#39;_ {
  2.     let data = self.data();
  3.     (0..data.shape()[0]).map(move |i| {
  4.         let el = data.index_axis(Axis(0), i);
  5.         let reshaped_and_cloned_el = el
  6.             .into_shape((el.shape()[0], 1))
  7.             .unwrap()
  8.             .mapv(|el| el.clone());
  9.         Tensor::new(reshaped_and_cloned_el)
  10.     })
  11. }

huangapple
  • 本文由 发表于 2023年7月20日 14:58:11
  • 转载请务必保留本文链接：https://go.coder-hub.com/76727378.html
  • iteration
  • rust
匿名

发表评论

匿名网友

:?: :razz: :sad: :evil: :!: :smile: :oops: :grin: :eek: :shock: :???: :cool: :lol: :mad: :twisted: :roll: :wink: :idea: :arrow: :neutral: :cry: :mrgreen:

确定