Mathe-problems

nomad

Re: mathe-problems

hi folks,

i working on little neural net with the propChips
as OBJ =
OBJ

text : "vga_text"
f : "FloatMath"
fp : "FloatString"

now i have problems with this equations:

=> outH1 = 1 / (1+exp(-1 * inH1));

1
this means: outH1 := __________
- inH1
1 + e

exp = 2.718281828 as constant
long inH1 = 0.009068429 as variable

my experiment:

x1 := f.FMul(-1.0,inH1)
x2 := f.FSqr(x1)
x3 := f.FMul(exp,x2)
x4 := f.FAdd(1.0,x3)
outH1 := f.FDiv(1.0,x4)

text.out($0D) ' newline
text.str(fp.FloatToString(outH1)) == 1

should be: 0.5022670917

with the calculater: (1/(1+e(-1*0.009068429))))


or can i make a function or a methode
for
- inH1
1 + e

have somebody tips or hints for this problem???
its possible to calculate this equation with the prop??

thanks for any help
regards nomad

simonl

Hi nomad,

Not sure if I'm correct -- haven't got a PChip near me at the moment.

I think you need to recode like: outH1 = 1 / (1+CONSTANT( exp(-1 * inH1) ));

As I say, I could be wrong! But hope that helps.

▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔
Cheers,

Simon

BTW: I type as I'm thinking, so please don't take any offense at my writing style

---

www.norfolkhelicopterclub.co.uk
You'll always have as many take-offs as landings, the trick is to be sure you can take-off again ;-)

Mike Green

You can't calculate EXP() as a constant. The compiler doesn't have the transcendental functions built-in.

There is an FEXP() function for the Propeller, but you have to use the Float32 library instead of FloatMath. Download it from the Object Exchange. There's a detailed document describing the various calls and the speed of the routines.

nomad

hi
thank you verry much
for your hints and tips
in switzerland its 18.30 PM
on monday i do some experiments with your answers
thanks
regards nomad

nomad

hi Mike Green,

Thanks for your verry god hint:
the PDF-File is verry good (documentation for Float32Full)
its runing and the output is correct
Regards
nomad

The reference

Original equations

1
output: ________________ = 0.603
-0.42
1+e

with calculater:

output := (1.0 / (1.0 + e(-0.42))) = 0.6034832499

---

{
propNN2.spin
begin : Wed. 02.May 11.30:00 GMT 2007
for : Test-spin-program for equations (exp) on a propChips
- with the Demo-Board

---

Notes:
02.05. - interduction to test the equations of (exp)

- with f.start == starting a cog its running
- compile & run OK

---

}

CON
_clkmode = xtal1 + pll16x
_xinfreq = 5_000_800
xxx = pi
exp = 2.718281828


OBJ

text : "vga_text"
f : "Float32Full"
fp : "FloatString"

VAR
long testerror
long vlong
word vword
byte vbyte

long inH1 ' this is the input-value for the equations
' static = inH1 = 0.42

long outH1 ' solutions

'intermediate results
long x1, x2, x3, x4, x5, x6

PUB Main

' start Float32Full"
f.start

'start vga-term
text.start(16)
'print a string
text.str(string("PropNeuralNet Version 0.2.",13,13))

neuralNet



PUB neuralNet

inH1 := 0.42

x1 := f.FMul(-1.0,inH1) ' positive Value convert to negative Value

x2 := f.Exp(x1) ' e(-0.42)

x3 := f.FAdd(1.0,x2) ' (1.0 + x2)
x4 := f.FDiv(1.0,x3) ' (1.0 / x3)
outH1 := x4

text.out($0D) ' newline
text.str(fp.FloatToString(outH1))

Post Edited (nomad) : 5/2/2007 10:35:48 AM GMT