def demo():
cc=[
uniform(-0.2,0.2,(3,3)),
uniform(2.0,2.0,(3,1)),
]
a=pyAnn([2,10,10,2],default=.2,rand=1)
print a
trainset=[
([.5,0.0],[0.5,0.0]),
([0.0,0.0],[0.0,0.0]),
([0.0,.5],[0.5,0.5]),
([.5,.5],[0.0,0.5]),
]
turns=10000
errth=0
for i in xrange(turns):
err=0.0
for t in trainset:
a.feed_forward(t[0])
err += a.bp(t[1],fN=0.3,r=0.000)
if err<errth:
print "in turn: ", i
break
if i%100==1:
print "error: ", err
print a
for ts in trainset:
print ts[0],a.feed_forward(ts[0])
def demo():
cc = [
uniform(-0.2, 0.2, (3, 3)),
uniform(2.0, 2.0, (3, 1)),
]
a = pyAnn([2, 10, 10, 2], default=.2, rand=1)
print a
trainset = [
([.5, 0.0], [0.5, 0.0]),
([0.0, 0.0], [0.0, 0.0]),
([0.0, .5], [0.5, 0.5]),
([.5, .5], [0.0, 0.5]),
]
turns = 10000
errth = 0
for i in xrange(turns):
err = 0.0
for t in trainset:
a.feed_forward(t[0])
err += a.bp(t[1], fN=0.3, r=0.000)
if err < errth:
print "in turn: ", i
break
if i % 100 == 1:
print "error: ", err
print a
for ts in trainset:
print ts[0], a.feed_forward(ts[0])
def randomPointInSphere(r):
"""Returns a vector drawn from a uniform distribution within
a sphere of radius |r|."""
rsq = r * r
while 1:
x = N.array([uniform(-r, r), uniform(-r, r), uniform(-r, r)])
if N.dot(x, x) < rsq: break
return Vector(x)
def randomPointInBox(a, b=None, c=None):
"""Returns a vector drawn from a uniform distribution within a
rectangular box with edge lengths |a|, |b|, |c|. If |b| and/or |c|
are omitted, they are taken to be equal to |a|."""
if b is None: b = a
if c is None: c = a
x = uniform(-0.5 * a, 0.5 * a)
y = uniform(-0.5 * b, 0.5 * b)
z = uniform(-0.5 * c, 0.5 * c)
return Vector(x, y, z)
def randomPointInSphere(r):
"""
:param r: the radius of a sphere
:type r: float
:returns: a vector drawn from a uniform distribution within
a sphere of radius r.
:rtype: Scientific.Geometry.Vector
"""
rsq = r*r
while 1:
x = N.array([uniform(-r, r), uniform(-r, r), uniform(-r, r)])
if N.dot(x, x) < rsq: break
return Vector(x)
def randomPointInSphere(r):
"""
:param r: the radius of a sphere
:type r: float
:returns: a vector drawn from a uniform distribution within
a sphere of radius r.
:rtype: Scientific.Geometry.Vector
"""
rsq = r * r
while 1:
x = N.array([uniform(-r, r), uniform(-r, r), uniform(-r, r)])
if N.dot(x, x) < rsq: break
return Vector(x)
def __bp(self, targets, sig=dsigmoid, fN=0.5, r=0.1):
_t = targets
### calculate the delta of output layer
_delta = map(lambda x, y: (x - y) * dsigmoid(y), _t, self.neurons[-1])
###print "neurons", self.neurons, "+++"
_sq_error = 0.0
for ei in range(self.layers[-1]):
_sq_error += 0.5 * (targets[ei] - self.neurons[-1][ei])**2
for n in range(self.nl - 1, 0, -1):
# calculate the strength change
_rs = r * uniform(-1.0, 1.0,
(self.layers[n - 1], self.layers[n])) * _sq_error
_change = dot(transpose([self.neurons[n - 1]]), [_delta]) + _rs
### calculate the delta of hidden layers
#### first we calculate the error of hidden layers
_error_h = dot(_delta, transpose(self.strengthmatrixes[n - 1]))
#### then we have the delta
####_delta = dot( _diagn2, _error_h)
_delta = map(lambda x, y: x * dsigmoid(y), _error_h,
self.neurons[n - 1])
self.strengthmatrixes[n -
1] = self.strengthmatrixes[n -
1] + fN * _change
return _sq_error
pass
def __bp(self, targets,sig=dsigmoid, fN=0.5,r=0.1):
_t=targets
### calculate the delta of output layer
_delta = map(lambda x,y: (x-y)*dsigmoid(y),
_t, self.neurons[-1])
###print "neurons", self.neurons, "+++"
_sq_error = 0.0
for ei in range(self.layers[-1]):
_sq_error += 0.5*(targets[ei]-self.neurons[-1][ei])**2
for n in range(self.nl-1,0,-1):
# calculate the strength change
_rs = r*uniform(-1.0,1.0,(self.layers[n-1],self.layers[n]))*_sq_error
_change = dot(transpose([self.neurons[n-1]]) ,[_delta] )+_rs
### calculate the delta of hidden layers
#### first we calculate the error of hidden layers
_error_h = dot(_delta, transpose(self.strengthmatrixes[n-1]))
#### then we have the delta
####_delta = dot( _diagn2, _error_h)
_delta = map(lambda x,y: x*dsigmoid(y), _error_h, self.neurons[n-1])
self.strengthmatrixes[n-1] = self.strengthmatrixes[n-1] + fN*_change
return _sq_error
pass
def randomPointInBox(a, b=None, c=None):
"""
:param a: the edge length of a box along the x axis
:type a: float
:param b: the edge length of a box along the y axis (default: a)
:type b: float
:param c: the edge length of a box along the z axis (default: a)
:type c: float
:returns: a vector drawn from a uniform distribution within a
rectangular box with edge lengths a, b, c.
:rtype: Scientific.Geometry.Vector
"""
if b is None: b = a
if c is None: c = a
x = uniform(-0.5 * a, 0.5 * a)
y = uniform(-0.5 * b, 0.5 * b)
z = uniform(-0.5 * c, 0.5 * c)
return Vector(x, y, z)
def randomPointInBox(a, b = None, c = None):
"""
:param a: the edge length of a box along the x axis
:type a: float
:param b: the edge length of a box along the y axis (default: a)
:type b: float
:param c: the edge length of a box along the z axis (default: a)
:type c: float
:returns: a vector drawn from a uniform distribution within a
rectangular box with edge lengths a, b, c.
:rtype: Scientific.Geometry.Vector
"""
if b is None: b = a
if c is None: c = a
x = uniform(-0.5*a, 0.5*a)
y = uniform(-0.5*b, 0.5*b)
z = uniform(-0.5*c, 0.5*c)
return Vector(x, y, z)
def randomRotation(max_angle=N.pi):
"""
:param max_angle: the upper limit for the rotation angle
:type max_angle: float
:returns: a random rotation with a uniform axis distribution
and angles drawn from a uniform distribution between
-max_angle and max_angle.
:rtype: Scientific.Geometry.Transformations.Rotation
"""
return Rotation(randomDirection(), uniform(-max_angle, max_angle))
def randomRotation(max_angle = N.pi):
"""
:param max_angle: the upper limit for the rotation angle
:type max_angle: float
:returns: a random rotation with a uniform axis distribution
and angles drawn from a uniform distribution between
-max_angle and max_angle.
:rtype: Scientific.Geometry.Transformations.Rotation
"""
return Rotation(randomDirection(), uniform(-max_angle, max_angle))
def test_check_receptors(self):
"""
The Python and C versions of the receptor site hit test should return the same result
"""
from RandomArray import uniform, seed
from TAP_ext import check_receptors
from time import time
area = 200
num_LEs = 100
num_times = 10
num_sites = 4
sites = [
array([(20, 65), (40, 35), (70, 25), (75, 45), (55, 50), (45, 75),
(20, 65)], Float)
] * num_sites
# build site bounding boxes
BBs = []
for site in sites:
max_x = site[0, 0]
min_x = site[0, 0]
max_y = site[0, 1]
min_y = site[0, 1]
max_x = max(max_x, max(site[:, 0]))
min_x = min(min_x, min(site[:, 0]))
max_y = max(max_y, max(site[:, 1]))
min_y = min(min_y, min(site[:, 1]))
BBs.append(array((max_x, min_x, max_y, min_y), Float))
LEs = uniform(0, area, (num_times, num_LEs, 2))
Hit_Table1 = zeros((num_LEs, num_sites), Int)
start = time()
hit_test(LEs, sites, BBs, Hit_Table1, 0)
print "Python version took %.3f seconds" % (time() - start)
Hit_Table2 = zeros((num_LEs, num_sites), Int)
start = time()
check_receptors.hit_test(LEs, sites, BBs, Hit_Table2, 0)
print "c version took %.3f seconds" % (time() - start)
assert alltrue(
equal(Hit_Table1,
Hit_Table2)), "Python and C version gave different results"
def resize_using_defaults(a,newsize,default=1.0,rand=0):
"""
resize a matrix to 'newsize', and fill all extra cells with 'default'
"""
oldshape=shape(a)
dif=array(newsize)-array(oldshape)
if dif[0] <= 0 :
tmpa=a[:newsize[0]]
else:
if rand==0:
_add_a=ones((dif[0],oldshape[1]))*default
else:
_add_a=uniform(-1.0,1.0,(dif[0],oldshape[1]))*default
tmpa=concatenate((a,_add_a))
if dif[1] <= 0 :
tmpb=array(tmpa)[:,:newsize[1]]
else:
if rand==0:
_add_b=ones((newsize[0],dif[1]))*default
else:
_add_b=uniform(-1.0,1.0,(newsize[0],dif[1]))*default
tmpb=concatenate((tmpa,_add_b),1)
return tmpb
def resize_using_defaults(a, newsize, default=1.0, rand=0):
"""
resize a matrix to 'newsize', and fill all extra cells with 'default'
"""
oldshape = shape(a)
dif = array(newsize) - array(oldshape)
if dif[0] <= 0:
tmpa = a[:newsize[0]]
else:
if rand == 0:
_add_a = ones((dif[0], oldshape[1])) * default
else:
_add_a = uniform(-1.0, 1.0, (dif[0], oldshape[1])) * default
tmpa = concatenate((a, _add_a))
if dif[1] <= 0:
tmpb = array(tmpa)[:, :newsize[1]]
else:
if rand == 0:
_add_b = ones((newsize[0], dif[1])) * default
else:
_add_b = uniform(-1.0, 1.0, (newsize[0], dif[1])) * default
tmpb = concatenate((tmpa, _add_b), 1)
return tmpb
def test_check_receptors(self):
"""
The Python and C versions of the receptor site hit test should return the same result
"""
from RandomArray import uniform, seed
from TAP_ext import check_receptors
from time import time
area = 200
num_LEs = 100
num_times = 10
num_sites = 4
sites = [array([(20,65),(40,35),(70,25),(75,45),(55,50),(45,75),(20,65)],Float)]*num_sites
# build site bounding boxes
BBs = []
for site in sites:
max_x = site[0,0]
min_x = site[0,0]
max_y = site[0,1]
min_y = site[0,1]
max_x = max(max_x, max(site[:,0]))
min_x = min(min_x, min(site[:,0]))
max_y = max(max_y, max(site[:,1]))
min_y = min(min_y, min(site[:,1]))
BBs.append(array((max_x,min_x,max_y,min_y),Float))
LEs = uniform(0,area,(num_times,num_LEs,2))
Hit_Table1 = zeros((num_LEs, num_sites),Int)
start = time()
hit_test(LEs,sites,BBs,Hit_Table1,0)
print "Python version took %.3f seconds"%(time()-start)
Hit_Table2 = zeros((num_LEs, num_sites),Int)
start = time()
check_receptors.hit_test(LEs,sites,BBs,Hit_Table2,0)
print "c version took %.3f seconds"%(time()-start)
assert alltrue(equal(Hit_Table1,Hit_Table2)), "Python and C version gave different results"
def makenoise(mod_depth,mod_image,side):
from RandomArray import uniform
if len(mod_image) == 0:
# no image supplied
# returns uniform white noise [0-1] as 3-d array
noise = uniform(0, mod_depth, shape=(side,side))
tmp = zeros((side,side,3), Float)
tmp[:,:,0] = noise[:]
tmp[:,:,1] = noise[:]
tmp[:,:,2] = noise[:]
noise = tmp;
else:
# image file supplied
# returns an image scaled to [0-1] as 3-d array
tmp = Sprite(fname=mod_image)
tmp = tmp.subimage(0, 0, side, side, center=1)
noise = mod_depth * (tmp.astype(Float) - 128.0) / 128.0
return noise
def makenoise(mod_depth, mod_image, side):
from RandomArray import uniform
if len(mod_image) == 0:
# no image supplied
# returns uniform white noise [0-1] as 3-d array
noise = uniform(0, mod_depth, shape=(side, side))
tmp = zeros((side, side, 3), Float)
tmp[:, :, 0] = noise[:]
tmp[:, :, 1] = noise[:]
tmp[:, :, 2] = noise[:]
noise = tmp
else:
# image file supplied
# returns an image scaled to [0-1] as 3-d array
tmp = Sprite(fname=mod_image)
tmp = tmp.subimage(0, 0, side, side, center=1)
noise = mod_depth * (tmp.astype(Float) - 128.0) / 128.0
return noise
def simple_rdp(s, dir=None, vel=None, fraction=0.25, fgcolor=(255,255,255), bgcolor=(128,128,128), rseed=None): if rseed: old_seed = get_seed() seed(rseed[0], rseed[1]) if dir is None: for n in range(3): if n == 0: m = uniform(0.0, 1.0, shape=(s.w, s.h)) mc = where(greater(m, fraction), bgcolor[n], fgcolor[n]) s.array[:,:,n] = mc[::].astype(UnsignedInt8) else: dx = -int(round(vel * math.cos(math.pi * dir / 180.0))) dy = int(round(vel * math.sin(math.pi * dir / 180.0))) a = s.array[:,:,:] a = concatenate((a[dx:,:,:],a[:dx,:,:]), axis=0) a = concatenate((a[:,dy:,:],a[:,:dy,:]), axis=1) s.array[:,:,:] = a[::] if rseed: seed(old_seed[0], old_seed[1])
def __load_strength(self, strengths, default=1.0,rand=0):
"""
strength: matrix of strengths between each layer,
default: the default stength
"""
import random
self.strengthmatrixes=[]
loadlen=len(strengths)
for i in range(0, loadlen):
fill=default
self.strengthmatrixes.append(
resize_using_defaults(strengths[i], (self.layers[i],self.layers[i+1]),fill,rand)
)
if loadlen < self.nl:
for i in range(loadlen,self.nl-1):
fill=default
if rand!=0:
self.strengthmatrixes.append(
uniform(-1.0,1.0,(self.layers[i],self.layers[i+1]))*fill
)
else:
self.strengthmatrixes.append(
ones((self.layers[i],self.layers[i+1]))*fill
)
def __load_strength(self, strengths, default=1.0, rand=0):
"""
strength: matrix of strengths between each layer,
default: the default stength
"""
import random
self.strengthmatrixes = []
loadlen = len(strengths)
for i in range(0, loadlen):
fill = default
self.strengthmatrixes.append(
resize_using_defaults(strengths[i],
(self.layers[i], self.layers[i + 1]),
fill, rand))
if loadlen < self.nl:
for i in range(loadlen, self.nl - 1):
fill = default
if rand != 0:
self.strengthmatrixes.append(
uniform(-1.0, 1.0,
(self.layers[i], self.layers[i + 1])) * fill)
else:
self.strengthmatrixes.append(
ones((self.layers[i], self.layers[i + 1])) * fill)
def randomRotation(max_angle=N.pi):
"""Returns a Rotation object describing a random rotation
with a uniform axis distribution and angles drawn from
a uniform distribution between -|max_angle| and |max_angle|."""
return Rotation(randomDirection(), uniform(-max_angle, max_angle))