TastyPancakes
/
Flux.jl
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

tighter

This commit is contained in:
Mike J Innes 2016-10-12 17:09:57 +01:00
parent af8001bdb8
commit b4390e6a23
1 changed files with 1 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
README.md
View File

@ -8,7 +8,7 @@ The current focus is on ANNs with TensorFlow as a backend. While it's in a very
Flux's approach is heavily declarative (as opposed to imperative graph building, which has all the drawbacks of writing one huge macro). Models can be described in high-level terms close to the mathematical notation. Model descriptions are be separated from their implementations, and changes to the way a model is used should never require changes to the model itself. Flux's approach is heavily declarative (as opposed to imperative graph building, which has all the drawbacks of writing one huge macro). Models can be described in high-level terms close to the mathematical notation. Model descriptions are be separated from their implementations, and changes to the way a model is used should never require changes to the model itself.
A core part of the approach is the idea that the same problems that come up when building ML models (composition and reuse, variable scoping, applying optimisations etc.) have already been solved by programming languages. If we think of building models as programming, we can reuse those solutions, greatly reducing the barriers to learning and using ML systems. A core idea is that the same problems that come up when building ML models (composition and reuse, variable scoping, applying optimisations etc.) have already been solved by programming languages. If we think of building models as programming, we can reuse those solutions, greatly reducing the barriers to learning and using ML systems.
There are also some unusual possibilities enabled by Julia itself. Julia's speed makes it trivial to prototype fully custom layers with reasonable performance. In future, GPU codegen may enable us to hook custom layers into TensorFlow and other backends. There are also some unusual possibilities enabled by Julia itself. Julia's speed makes it trivial to prototype fully custom layers with reasonable performance. In future, GPU codegen may enable us to hook custom layers into TensorFlow and other backends.