vseledkin / ad Goto Github PK

a simple mathematical formula language, which is auto-differentiated and compiled to http://golang.org for high performance.

see https://autodiff.info for live demo.

to get started: make sure you have latest golang.org installed (e.g. https://golang.org/dl/), or build it yourself via https://github.com/xoba/goinit

then:

git clone --recursive https://github.com/xoba/ad.git
cd ad
source goinit.sh
./install.sh
run compile -formula "f := sqrt(abs(a+b*b))"
go run compute.go

for help, you can try:

run
run compile -help
run nn -help

it runs with both scalar and slice variables; e.g.:

run compile -formula "f:= 2*x[0]+1 + a + x[1] * sin(x[2]) + z/y[0]"

to develop with emacs:

./ide.sh

to auto-generate various code:

./gogenerate.sh

to run a simple neural network example:

run nn

which produces one of these two videos (first one has 5 hidden units second one has none, and is equivalent to logistic regression):

1. https://s3.amazonaws.com/xoba-videos/test_5.mp4
2. https://s3.amazonaws.com/xoba-videos/test_0.mp4

in these movies, white pixels are correctly guessed classes, red are incorrect. class#1 is a circle embedded in a 2-d feature space, class#2 is everything else. "time" in the movie is monotonic with training iteration.