"""
Old patterns not intended for new code.
These patterns are expected to be deleted eventually.
"""
import param
import numpy as np
from numpy import pi
from param.parameterized import ParamOverrides
from .patterngenerator import Constant, PatternGenerator, Composite
from . import Gaussian
from .image import FileImage, PatternSampler, ImageSampler, edge_average
[docs]class SeparatedComposite(Composite):
"""
Generalized version of the Composite PatternGenerator that
enforces spacing constraints between pattern centers.
Currently supports minimum spacing, but can be generalized to
support maximum spacing also (and both at once).
"""
min_separation = param.Number(default=0.0, bounds = (0,None),
softbounds = (0.0,1.0), doc="""
Minimum distance to enforce between all pairs of pattern centers.
Useful for ensuring that multiple randomly generated patterns
do not overlap spatially. Note that as this this value is
increased relative to the area in which locations are chosen,
the likelihood of a pattern appearing near the center of the
area will decrease. As this value approaches the available
area, the corners become far more likely to be chosen, due to
the distances being greater along the diagonals.
""")
### JABNOTE: Should provide a mechanism for collecting and
### plotting the training pattern center distribution, so that
### such issues can be checked.
max_trials = param.Integer(default = 50, bounds = (0,None),
softbounds = (0,100), precedence=-1, doc="""
Number of times to try for a new pattern location that meets
the criteria.
This is an essentially arbitrary timeout value that helps
prevent an endless loop in case the requirements cannot be
met.""")
def __distance_valid(self, g0, g1, p):
"""
Returns true if the distance between the (x,y) locations of
two generators g0 and g1 is greater than a minimum separation.
Can be extended easily to support other criteria.
"""
dist = np.sqrt((g1.x - g0.x) ** 2 +
(g1.y - g0.y) ** 2)
return dist >= p.min_separation
def _advance_pattern_generators(self,p):
"""
Advance the parameters for each generator for this
presentation.
Picks a position for each generator that is accepted by
__distance_valid for all combinations. Returns a new list of
the generators, with some potentially omitted due to failure
to meet the constraints.
"""
valid_generators = []
for g in p.generators:
for trial in range(self.max_trials):
# Generate a new position and add generator if it's ok
if np.alltrue([self.__distance_valid(g,v,p) for v in valid_generators]):
valid_generators.append(g)
break
g.force_new_dynamic_value('x')
g.force_new_dynamic_value('y')
else:
self.warning("Unable to place pattern %s subject to given constraints" %
g.name)
return valid_generators
#JABALERT: Can't this be replaced with a Composite?
[docs]class TwoRectangles(PatternGenerator):
"""Two 2D rectangle pattern generator."""
aspect_ratio = param.Number(default=1.0,bounds=(0.0,None),softbounds=(0.0,6.0),
precedence=0.31,doc=
"Ratio of width to height; size*aspect_ratio gives the width of the rectangle.")
x1 = param.Number(default=-0.15,bounds=(-1.0,1.0),softbounds=(-0.5,0.5),
doc="X center of rectangle 1.")
y1 = param.Number(default=-0.15,bounds=(-1.0,1.0),softbounds=(-0.5,0.5),
doc="Y center of rectangle 1.")
x2 = param.Number(default=0.15,bounds=(-1.0,1.0),softbounds=(-0.5,0.5),
doc="X center of rectangle 2.")
y2 = param.Number(default=0.15,bounds=(-1.0,1.0),softbounds=(-0.5,0.5),
doc="Y center of rectangle 2.")
size = param.Number(default=0.5)
# YC: Maybe this can be implemented much more cleanly by calling
# the parent's function() twice, but it's hard to see how to
# set the (x,y) offset for the parent.
def function(self,p):
height = p.size
width = p.aspect_ratio*height
return np.bitwise_or(
np.bitwise_and(np.bitwise_and(
(self.pattern_x-p.x1)<=p.x1+width/4.0,
(self.pattern_x-p.x1)>=p.x1-width/4.0),
np.bitwise_and(
(self.pattern_y-p.y1)<=p.y1+height/4.0,
(self.pattern_y-p.y1)>=p.y1-height/4.0)),
np.bitwise_and(np.bitwise_and(
(self.pattern_x-p.x2)<=p.x2+width/4.0,
(self.pattern_x-p.x2)>=p.x2-width/4.0),
np.bitwise_and(
(self.pattern_y-p.y2)<=p.y2+height/4.0,
(self.pattern_y-p.y2)>=p.y2-height/4.0)))
### JABALERT: This class should be eliminated if at all possible; it
### is just a specialized version of Composite, and should be
### implementable directly using what is already in Composite.
[docs]class GaussiansCorner(PatternGenerator):
"""
Two Gaussian pattern generators with a variable intersection point,
appearing as a corner or cross.
"""
x = param.Number(default=-0.15,bounds=(-1.0,1.0),softbounds=(-0.5,0.5),
doc="X center of the corner")
y = param.Number(default=-0.15,bounds=(-1.0,1.0),softbounds=(-0.5,0.5),
doc="Y center of the corner")
size = param.Number(default=0.5,bounds=(0,None), softbounds=(0.1,1),
doc="The size of the corner")
aspect_ratio = param.Number(default=1/0.31, bounds=(0,None), softbounds=(1,10),
doc="Ratio of the width to the height for both Gaussians")
angle = param.Number(default=0.5*pi,bounds=(0,pi), softbounds=(0.01*pi,0.99*pi),
doc="The angle of the corner")
cross = param.Number(default=0.4, bounds=(0,1), softbounds=(0,1),
doc="Where the two Gaussians cross, as a fraction of their half length")
def __call__(self,**params_to_override):
p = ParamOverrides(self,params_to_override)
g_1 = Gaussian()
g_2 = Gaussian()
x_1 = g_1(orientation = p.orientation, bounds = p.bounds, xdensity = p.xdensity,
ydensity = p.ydensity, offset = p.offset, size = p.size,
aspect_ratio = p.aspect_ratio,
x = p.x + 0.7 * np.cos(p.orientation) * p.cross * p.size * p.aspect_ratio,
y = p.y + 0.7 * np.sin(p.orientation) * p.cross * p.size * p.aspect_ratio)
x_2 = g_2(orientation = p.orientation+p.angle, bounds = p.bounds, xdensity = p.xdensity,
ydensity = p.ydensity, offset = p.offset, size = p.size,
aspect_ratio = p.aspect_ratio,
x = p.x + 0.7 * np.cos(p.orientation+p.angle) * p.cross * p.size * p.aspect_ratio,
y = p.y + 0.7 * np.sin(p.orientation+p.angle) * p.cross * p.size * p.aspect_ratio)
return np.maximum( x_1, x_2 )
[docs]class Translator(PatternGenerator):
"""
PatternGenerator that translates another PatternGenerator over
time.
This PatternGenerator will create a series of episodes, where in
each episode the underlying generator is moved in a fixed
direction at a fixed speed. To begin an episode, the Translator's
x, y, and direction are evaluated (e.g. from random
distributions), and the underlying generator is then drawn at
those values plus changes over time that are determined by the
speed. The orientation of the underlying generator should be set
to 0 to get motion perpendicular to the generator's orientation
(which is typical).
Note that at present the parameter values for x, y, and direction
cannot be passed in when the instance is called; only the values
set on the instance are used.
"""
generator = param.ClassSelector(default=Gaussian(),
class_=PatternGenerator,doc="""Pattern to be translated.""")
direction = param.Number(default=0.0,softbounds=(-pi,pi),doc="""
The direction in which the pattern should move, in radians.""")
speed = param.Number(default=0.01,bounds=(0.0,None),doc="""
The speed with which the pattern should move,
in sheet coordinates per time_fn unit.""")
reset_period = param.Number(default=1,bounds=(0.0,None),doc="""
Period between generating each new translation episode.""")
episode_interval = param.Number(default=0,doc="""
Interval between successive translation episodes.
If nonzero, the episode_separator pattern is presented for
this amount of time_fn time after each episode, e.g. to
allow processing of the previous episode to complete.""")
episode_separator = param.ClassSelector(default=Constant(scale=0.0),
class_=PatternGenerator,doc="""
Pattern to display during the episode_interval, if any.
The default is a blank pattern.""")
time_fn = param.Callable(default=param.Dynamic.time_fn,doc="""
Function to generate the time used as a base for translation.""")
def _advance_params(self):
"""
Explicitly generate new values for these parameters only
when appropriate.
"""
for p in ['x','y','direction']:
self.force_new_dynamic_value(p)
self.last_time = self.time_fn()
def __init__(self,**params):
super(Translator,self).__init__(**params)
self._advance_params()
def __call__(self,**params_to_override):
p=ParamOverrides(self,params_to_override)
if self.time_fn() >= self.last_time + p.reset_period:
## Returns early if within episode interval
if self.time_fn()<self.last_time+p.reset_period+p.episode_interval:
return p.episode_separator(xdensity=p.xdensity,
ydensity=p.ydensity,
bounds=p.bounds)
else:
self._advance_params()
# JABALERT: Does not allow x, y, or direction to be passed in
# to the call; fixing this would require implementing
# inspect_value and force_new_dynamic_value (for
# use in _advance_params) for ParamOverrides.
#
# Access parameter values without giving them new values
assert ('x' not in params_to_override and
'y' not in params_to_override and
'direction' not in params_to_override)
x = self.inspect_value('x')
y = self.inspect_value('y')
direction = self.inspect_value('direction')
# compute how much time elapsed from the last reset
# float(t) required because time could be e.g. gmpy.mpq
t = float(self.time_fn()-self.last_time)
## CEBALERT: mask gets applied twice, both for the underlying
## generator and for this one. (leads to redundant
## calculations in current lissom_oo_or usage, but will lead
## to problems/limitations in the future).
return p.generator(
xdensity=p.xdensity,ydensity=p.ydensity,bounds=p.bounds,
x=x+t*np.cos(direction)*p.speed+p.generator.x,
y=y+t*np.sin(direction)*p.speed+p.generator.y,
orientation=(direction-pi/2)+p.generator.orientation)
# Legacy Sweeper class which is used in lissom.ty, should be deleted
# once lissom.ty is deprecated
#
# CB: I removed motion_sign from this class because I think it is
# unnecessary. But maybe I misunderstood the original author's
# intention?
#
# In any case, the original implementation was incorrect - it was not
# possible to get some motion directions (directions in one whole
# quadrant were missed out).
#
# Note that to get a 2pi range of directions, one must use a 2pi range
# of orientations (there are two directions for any given
# orientation). Alternatively, we could generate a random sign, and
# use an orientation restricted to a pi range.
[docs]class OldSweeper(PatternGenerator):
"""
PatternGenerator that sweeps a supplied PatternGenerator in a
direction perpendicular to its orientation.
"""
generator = param.Parameter(default=Gaussian(),precedence=0.97, doc="Pattern to sweep.")
speed = param.Number(default=0.25,bounds=(0.0,None),doc="""
Sweep speed: number of sheet coordinate units per unit time.""")
step = param.Number(default=1,doc="""
Number of steps at the given speed to move in the sweep direction.
The distance moved is speed*step.""")
# Provide access to value needed for measuring maps
def __get_phase(self): return self.generator.phase
def __set_phase(self,new_val): self.generator.phase = new_val
phase = property(__get_phase,__set_phase)
[docs] def function(self,p):
"""Selects and returns one of the patterns in the list."""
pg = p.generator
motion_orientation=p.orientation+pi/2.0
new_x = p.x+p.size*pg.x
new_y = p.y+p.size*pg.y
image_array = pg(xdensity=p.xdensity,ydensity=p.ydensity,bounds=p.bounds,
x=new_x + p.speed*p.step*np.cos(motion_orientation),
y=new_y + p.speed*p.step*np.sin(motion_orientation),
orientation=p.orientation,
scale=pg.scale*p.scale,offset=pg.offset+p.offset)
return image_array
[docs]class NumpyFile(FileImage):
"""
For backwards compatibility.
"""
pattern_sampler = param.ClassSelector(class_=ImageSampler,
default=PatternSampler(background_value_fn=edge_average,
size_normalization='original',
whole_pattern_output_fns=[]),doc="""
The PatternSampler to use to resample/resize the image.""")
_public = list(set([_k for _k,_v in locals().items() if isinstance(_v,type) and issubclass(_v,PatternGenerator)]))
__all__ = _public