Flux.jl/README.md

22 lines
1.2 KiB
Markdown
Raw Normal View History

2016-05-26 11:36:28 +00:00
# Флукс
2016-05-11 13:44:34 +00:00
## What?
2016-08-15 16:24:20 +00:00
Flux is a programming model for building neural networks, implemented in Julia.
2016-05-11 13:44:34 +00:00
## Why?
2016-04-13 11:51:55 +00:00
2016-08-15 16:24:20 +00:00
Flux is designed to be much more intuitive than traditional frameworks. For starters, that means having a simple notation for models that's as close to the mathematical description as possible (like `σ(W*x + b)`). But it's deeper than syntax; we also reuse concepts from regular programming languages (like the class/object distinction) to create principled semantics. Flux is fully declarative, so there's no more mental juggling of multiple execution paths as you read imperative graph-building code.
2016-04-13 11:51:55 +00:00
2016-08-15 16:24:20 +00:00
Flux's semantics include native support for recurrent loops, which it can automatically unroll for you never do it by hand again.
2016-05-11 13:44:34 +00:00
2016-08-15 16:24:20 +00:00
But it's also designed to be extremely flexible. Flux supports multiple backends MXNet to begin with and TensorFlow in future transparently taking advantage of all their features rather than providing a lowest common denominator. Flux's design allows for custom layer types say custom GPU kernels to be implemented in pure Julia, for backends that support it.
2016-05-11 13:44:34 +00:00
2016-05-11 16:58:10 +00:00
## How?
See [the design docs](design.md).
2016-05-11 13:44:34 +00:00
## Is it any good?
Yes.