add performance tips to docs #464
Conversation
There was a problem hiding this comment.
Its a good start, but a lot of information here is either factually wrong or just hallucinated. Its perfectly fine to use ChatGPT to improve English (as I always do myself nowadays), however, it is very dangerous to let it just generate everything. This text isn't helpful for an outsider because in its current shape its either incorrect or misleading.
On a separate note, I would also like to see some @examples-s (builtin Documenter.jl feature, you can read the documentation of Documenter.jl or see other files for reference) as well as @ref-rences to other pages in the documentation.
By implementing @example blocks you can show the proposed solutions in action and see if they actually work or not.
| @@ -33,7 +33,6 @@ makedocs(; | |||
| "User guide" => [ | |||
| "Getting started" => "manuals/getting-started.md", | |||
| "RxInfer.jl vs. Others" => "manuals/comparison.md", | |||
| "Using RxInfer from Python" => "manuals/how-to-use-rxinfer-from-python.md", | |||
|
|
||
| **Why?** | ||
| Julia uses **Just-In-Time (JIT)** compilation. | ||
| The **first time** you run a model (or even parts of it like `infer!`), Julia compiles the specialized machine code. |
There was a problem hiding this comment.
We don't have the infer! function, the extra ! is wrong
|
|
||
| ## Non-Linear Nodes | ||
|
|
||
| ### Prefer Linearization Over Unscented or CUI Methods |
There was a problem hiding this comment.
What is a CUI method? Perhaps you meant CVI?
| Thus, **Linearization is much faster** and usually good enough unless you need extremely high precision. | ||
|
|
||
| **Further Reading:** | ||
| See the “Deterministic Nodes” section of the RxInfer documentation for more on non-linear nodes. |
| **Why?** | ||
| When you use `free_energy = true`, RxInfer needs to store more detailed internal states for diagnostic purposes. | ||
| When you specify `free_energy = Float64`, it reduces tracking to **minimal scalar computations**, making the process **leaner and faster**. |
There was a problem hiding this comment.
The reason for free_energy = Float64 is faster is because by default it stores the type of the computation as Real in order to enable auto-differentiation
| ### SoftDot Node | ||
|
|
||
| **Why?** | ||
| Efficient for representing weighted sums with uncertainty — useful in regression-like structures. |
There was a problem hiding this comment.
An example with the @example block would be nice
| - **Forces premature message updates** | ||
| - **May break inference stability** | ||
|
|
||
| Use it carefully; it’s more for **debugging** deep models than regular optimization. |
| - **Cuts off legitimate computations** | ||
| - **Forces premature message updates** | ||
| - **May break inference stability** |
There was a problem hiding this comment.
these bullets are entirely hallucinated
| ### How to Convert | ||
|
|
||
| - **Smoothing models** use future observations to improve past estimates (i.e., two-pass inference). | ||
| - **Filtering models** only use past and present data, updating beliefs in a **causal** fashion. |
| **To convert:** | ||
| - Remove or disable factors that depend on future observations. | ||
| - Restrict message passing to only move forward in time. |
There was a problem hiding this comment.
There are no capabilities in RxInfer to do that...
|
Hey @Thijsie2 do you need help with it? |
First draft of performance tips page for the documentation (in reply to issue #456)