Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Update comparison section in overview.md #88

Merged
merged 1 commit into from
Jul 16, 2022
Merged

Update comparison section in overview.md #88

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
24 changes: 12 additions & 12 deletions docs/src/introduction/overview.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,25 +22,25 @@ it both compiler and autodiff friendly.

* **Neural Networks for SciML**: For SciML Applications (Neural ODEs, Deep Equilibrium
Models) solvers typically expect a monolithic parameter vector. Flux enables this via its
`destructure` mechanism, however, it often leads to
[weird bugs](https://github.com/FluxML/Flux.jl/issues?q=is%3Aissue+destructure). Lux
`destructure` mechanism, but `destructure` comes with various
[edge cases and limitations](https://fluxml.ai/Optimisers.jl/dev/api/#Optimisers.destructure). Lux
forces users to make an explicit distinction between state variables and parameter
variables to avoid these issues. Also, it comes battery-included for distributed training
using [FluxMPI.jl](https://github.com/avik-pal/FluxMPI.jl) *(I know :P the naming)*

* **Sensible display of Custom Layers** -- Ever wanted to see Pytorch like Network printouts
or wondered how to extend the pretty printing of Flux's layers. Lux handles all of that
or wondered how to extend the pretty printing of Flux's layers? Lux handles all of that
by default.

* **Less Bug-ridden Code** - *No arbitrary internal mutations* since all layers are
implemented as pure functions. *All layers are deterministic* given the parameter and
state (if the layer is supposed to be stochastic (say `Dropout`), the state must contain a
seed which is then updated after the function call).

* **Easy Parameter Manipulation** -- Wondering why Flux doesn't have `WeightNorm`,
`SpectralNorm`, etc. The implicit parameter handling makes it extremely hard to pass
parameters around without mutations which AD systems don't like. With Lux implementing
them is outright simple.
* **Truly immutable models** - No *unexpected internal mutations* since all layers are
implemented as pure functions. All layers are also *deterministic* given the parameters and
state: if a layer is supposed to be stochastic (say `Dropout`), the state must contain a
seed which is then updated after the function call.

* **Easy Parameter Manipulation** -- By separating parameter data and layer structures,
Lux makes implementing `WeightNorm`, `SpectralNorm`, etc. downright trivial.
Without this separation, it is much harder to pass such parameters
around without mutations which AD systems don't like.


# Why not use Lux?
Expand Down