particle swarm optimization

diff --git a/filters.py b/filters.py
index 397a371..c53b6e8 100644 (file)
--- a/filters.py
+++ b/filters.py
@@ -20,7 +20,7 @@ def peaks(image):
                 - image_l[a, b] 
                     for b in range(y - 2, y + 3)]) 
                     for a in range(x - 2, x + 3)])
-                + (16 * image_l[x, y]))
+                + (17 * image_l[x, y]))
             if pix > 255:
                 pix = 255
             if pix < 0:

diff --git a/gridf.py b/gridf.py
index 0c5b68f..58229a4 100644 (file)
--- a/gridf.py
+++ b/gridf.py
@@ -1,6 +1,7 @@
 """Imago grid-fitting module"""
 
 import multiprocessing
+from functools import partial
 
 import Image, ImageDraw, ImageFilter
 
@@ -9,6 +10,7 @@ from manual import lines as g_grid, l2ad
 from intrsc import intersections_from_angl_dist
 from linef import line_from_angl_dist
 import pcf
+import pso
 
 class GridFittingFailedError(Exception):
     pass
@@ -21,31 +23,12 @@ class MyGaussianBlur(ImageFilter.Filter):
     def filter(self, image):
         return image.gaussian_blur(self.radius)
 
-def job_br1(args):
-    im_l, v1, v2, h1, h2, x, y, dv, dh, size = args
-    v1 = (v1[0] + x * dv, v1[1] + x)
-    v2 = (v2[0] + y * dv, v2[1] + y)
-    return (distance(im_l, 
-                    get_grid([v1, v2], [h1, h2], size),
-                    size), x, y) 
-
-def job_br2(args):
-    im_l, v1, v2, h1, h2, x, y, dv, dh, size = args
-    h1 = (h1[0] + x * dh, h1[1] + x)
-    h2 = (h2[0] + y * dh, h2[1] + y)
-    return (distance(im_l, 
-                    get_grid([v1, v2], [h1, h2], size),
-                    size), x, y) 
-
-def job_4(args):
-    im_l, v1, v2, h1, h2, x, y, w, z, dv, dh, size = args
+def job_4(x, y, w, z, im_l, v1, v2, h1, h2, dv, dh, size):
     v1 = (v1[0] + x * dv, v1[1] + x)
     v2 = (v2[0] + y * dv, v2[1] + y)
     h1 = (h1[0] + w * dh, h1[1] + w)
     h2 = (h2[0] + z * dh, h2[1] + z)
-    return (distance(im_l, 
-                    get_grid([v1, v2], [h1, h2], size),
-                    size), x, y, w, z) 
+    return (distance(im_l, get_grid([v1, v2], [h1, h2], size), size))
 
 def find(lines, size, l1, l2, bounds, hough, do_something, im_h):
     l1 = line_from_angl_dist(l1, size)
@@ -73,36 +56,13 @@ def find(lines, size, l1, l2, bounds, hough, do_something, im_h):
     #GaussianBlur is undocumented class, may not work in future versions of PIL
     im_l_s = im_l.tostring()
 
-    #let's try the ULTRA bruteforce aproach:
-    pool = multiprocessing.Pool(None)
+    import time
+    start = time.time()
 
-    #import time
-    #start = time.time()
+    f_dist = partial(job_4, im_l=im_l_s, v1=v1, v2=v2, h1=h1, h2=h2,
+                     dv=delta_v, dh=delta_h, size=size)
 
-    k = 30 
-    tasks = [(im_l_s, v1, v2, h1, h2, x, y, w, z, delta_v, delta_h, size) 
-             for x in xrange(-k, k, 2)
-             for y in xrange(-k, k, 2)
-             for z in xrange(-k, k, 2)
-             for w in xrange(-k, k, 2)]
-
-    opt = pool.map(job_4, tasks)
-    _, x_v, y_v, x_h, y_h = max(opt)
-
-    v1 = (v1[0] + x_v * delta_v, v1[1] + x_v)
-    v2 = (v2[0] + y_v * delta_v, v2[1] + y_v)
-    h1 = (h1[0] + x_h * delta_h, h1[1] + x_h)
-    h2 = (h2[0] + y_h * delta_h, h2[1] + y_h)
-
-    k = 5 
-    tasks = [(im_l_s, v1, v2, h1, h2, x, y, w, z, delta_v, delta_h, size) 
-             for x in xrange(-k, k)
-             for y in xrange(-k, k)
-             for z in xrange(-k, k)
-             for w in xrange(-k, k)]
-
-    opt = pool.map(job_4, tasks)
-    _, x_v, y_v, x_h, y_h = max(opt)
+    x_v, y_v, x_h, y_h = pso.optimize(4, 30, f_dist, 32, 1028)
 
     v1 = (v1[0] + x_v * delta_v, v1[1] + x_v)
     v2 = (v2[0] + y_v * delta_v, v2[1] + y_v)
@@ -112,11 +72,8 @@ def find(lines, size, l1, l2, bounds, hough, do_something, im_h):
     grid = get_grid([v1, v2], [h1, h2], size) 
     grid_lines = [[l2ad(l, size) for l in grid[0]], 
                   [l2ad(l, size) for l in grid[1]]]
-
-    pool.terminate()
-    pool.join()
     
-    #print time.time() - start
+    print time.time() - start
     
 ### Show error surface
 #
@@ -127,18 +84,18 @@ def find(lines, size, l1, l2, bounds, hough, do_something, im_h):
 
 ### Show grid over lines
 #
-#    im_t = Image.new('RGB', im_l.size, None)
-#    im_t_l = im_t.load()
-#    im_l_l = im_l.load()
-#    for x in xrange(im_t.size[0]):
-#        for y in xrange(im_t.size[1]):
-#            im_t_l[x, y] = (im_l_l[x, y], 0, 0)
-#
-#    im_t_d = ImageDraw.Draw(im_t)
-#    for l in grid[0] + grid[1]:
-#        im_t_d.line(l, width=1, fill=(0, 255, 0))
-#
-#    do_something(im_t, "lines and grid")
+    im_t = Image.new('RGB', im_l.size, None)
+    im_t_l = im_t.load()
+    im_l_l = im_l.load()
+    for x in xrange(im_t.size[0]):
+        for y in xrange(im_t.size[1]):
+            im_t_l[x, y] = (im_l_l[x, y], 0, 0)
+
+    im_t_d = ImageDraw.Draw(im_t)
+    for l in grid[0] + grid[1]:
+        im_t_d.line(l, width=1, fill=(0, 255, 0))
+
+    do_something(im_t, "lines and grid")
 ###
 
     return grid, grid_lines
@@ -165,8 +122,8 @@ def distance(im_l, grid, size):
     dr_g = ImageDraw.Draw(im_g)
     for line in grid[0] + grid[1]:
         dr_g.line(line, width=1, fill=255)
-#        if line_out(line, size):
-#            return 0
+        if line_out(line, size):
+            return 0
     #im_g = im_g.filter(MyGaussianBlur(radius=3))
     #GaussianBlur is undocumented class, may not work in future versions of PIL
     #im_d, distance = combine(im_l, im_g)

diff --git a/pso.py b/pso.py
new file mode 100644 (file)
index 0000000..9fdee2d
--- /dev/null
+++ b/pso.py
@@ -0,0 +1,39 @@
+"""Particle swarm optimization"""
+
+import random
+import multiprocessing
+from functools import partial
+
+def particle(dimension, bound, func_d):
+    position = [2 * bound * random.random() - bound for _ in xrange(dimension)]
+    velocity = [2 * bound * random.random() - bound for _ in xrange(dimension)]
+    value = func_d(*position)
+    return value, position, velocity, value, position
+
+def move(particle, omega, phi_p, phi_g, v_max, global_best, func_d):
+    _, position, velocity, best_value, best_position = particle
+    position = [p + v for (p, v) in zip(position, velocity)]
+    velocity = [omega * v 
+                + phi_p * random.random() * (b - x)
+                + phi_g * random.random() * (g - x)
+                for (v, x, b, g) in zip(velocity, position,
+                                        best_position, global_best)]
+    velocity = [min(max(v, - v_max), v_max) for v in velocity] 
+    value = func_d(*position)
+    if value > best_value:
+        best_value, best_position = value, position
+    return value, position, velocity, best_value, best_position
+
+def optimize(dimension, boundary, function_d, n_parts, n_turns):
+    pool = multiprocessing.Pool(None)
+    particles = [particle(dimension, boundary, function_d)
+                 for _ in xrange(n_parts)]
+    gl_best = max(particles)
+    for _ in xrange(n_turns):
+        move_p = partial(move, omega=0.9, phi_p=0.9, phi_g=0.2, v_max=20.,
+                         global_best=gl_best[1], func_d=function_d)
+        particles = pool.map(move_p, particles)
+        gl_best = max(max(particles), gl_best)
+    pool.terminate()
+    pool.join()
+    return gl_best[1]