Broadcast Illustration

[bluss]

This one is really good, ndarray should adopt this (just need to change the code for each illu)

http://www.astroml.org/book_figures/appendix/fig_broadcast_visual.html

[bluss]

draft. Doesn't look good in github, fine in vim, fine in rustdoc's fonts.

array![0, 1, 2] + array![1]
shape [3] + shape [1] = shape [3]
 
┌───┬───┬───┐   ┌───┐┄┄┄┬┄┄┄╮   ┌───┬───┬───┐
│ 0 │ 1 │ 2 │ + │ 1 │ 1 ┊ 1 ┊ = │ 1 │ 2 │ 3 │
└───┴───┴───┘   └───┘┄┄┄┴┄┄┄╯   └───┴───┴───┘
 
array![[1, 1, 1],
       [1, 1, 1],
       [1, 1, 1]] + array![0, 1, 2]

shape [3, 3] + shape [3] = shape [3, 3]
 
┌───┬───┬───┐   ┌───┬───┬───┐   ┌───┬───┬───┐
│ 1 │ 1 │ 1 │   │ 0 │ 1 │ 2 │   │ 1 │ 2 │ 3 │
├───┼───┼───┤   └───┴───┴───┘   ├───┼───┼───┤
│ 1 │ 1 │ 1 │ + ┊ 0 ┊ 1 ┊ 2 ┊ = │ 1 │ 2 │ 3 │
├───┼───┼───┤   ├┄┄┄ ┄┄┄ ┄┄┄┤   ├───┼───┼───┤
│ 1 │ 1 │ 1 │   ┊ 0 ┊ 1 ┊ 2 ┊   │ 1 │ 2 │ 3 │
└───┴───┴───┘   ╰┄┄┄┴┄┄┄┴┄┄┄╯   └───┴───┴───┘
 
array![[0],
       [1],
       [2]] + array![[0, 1, 2]]

shape [3, 1] + shape [1, 3] = shape [3, 3]
 
┌───┐┄┄┄ ┄┄┄╮   ┌───┬───┬───┐   ┌───┬───┬───┐
│ 0 │ 0 ┊ 0 ┊   │ 0 │ 1 │ 2 │   │ 0 │ 1 │ 2 │
├───┤┄┄┄ ┄┄┄    └───┴───┴───┘   ├───┼───┼───┤
│ 1 │ 1 ┊ 1 ┊ + ┊ 0 ┊ 1 ┊ 2 ┊ = │ 1 │ 2 │ 3 │
├───┤┄┄┄ ┄┄┄     ┄┄┄ ┄┄┄ ┄┄┄    ├───┼───┼───┤
│ 2 │ 2 ┊ 2 ┊   ┊ 0 ┊ 1 ┊ 2 ┊   │ 1 │ 3 │ 4 │
└───┘┄┄┄ ┄┄┄╯   ╰┄┄┄ ┄┄┄ ┄┄┄╯   └───┴───┴───┘

[sglyon]

I'm a long time numpy user and am I'm having a hard time understanding how broadcasting semantics work here. In fact, if I try your third example from the previous comment I get an error:

⇒ cargo run                                                                                                                                          Spencers-MacBook-Pro
    Finished debug [unoptimized + debuginfo] target(s) in 0.0 secs
     Running `target/debug/bertola`
thread 'main' panicked at 'ndarray: could not broadcast array from shape: [1, 3] to: [3, 1]', /Users/sglyon/.cargo/registry/src/github.aaakk.us.kg-1ecc6299db9ec823/ndarray-0.7.3/src/lib.rs:573
note: Run with `RUST_BACKTRACE=1` for a backtrace.

Is this expected?

[bluss]

yes, ndarray does not implement the third case.

[sglyon]

Ok, thanks for the answer.

Are there plans to implement the third case, or is this something that will not become a feature of this library?

[bluss]

As you know arrays in ndarray have dimensionality in the type.

If we add Array<f32, D1> + Array<f32, D2> the result needs to be a Array<f32, D3> where D3 is some sort of upper bound of the dimension types D1 and D2; so D3 needs to be the bigger of the two to have room to storing the dimension information.

As it is right now, the + always leaves the result with the dimension type D1 and that makes it simple, just broadcast the right hand side to the shape of the left hand side.

I haven't attempted implementing it; I'm in general not so keen on more generic noise than absolutely necessary. Every bit makes the library hard to understand, and many apparently already think it is hard to get into.

[sglyon]

OK that makes sense -- thanks for the clarification.

[thecooldrop]

@bluss, may thisi ssue be closed due to it being practically resolved in PR #565 ?

[bluss]

We never ended up adding broadcasting illustrations for #565 so I guess we're still waiting for that.