Faster multi-process handling and clean up
This commit is contained in:
Sebastian Seedorf
310
__main__.py
310
__main__.py
@@ -1,178 +1,168 @@
|
|||||||
import numpy as np
|
import numpy as np
|
||||||
import threading
|
|
||||||
import queue
|
|
||||||
import time
|
|
||||||
|
|
||||||
INPUT_FILE = "data/medium.in"
|
INPUT_FILE = "data/medium.in"
|
||||||
POPULATION = 50
|
POPULATION = 50
|
||||||
MUTATION_AMOUNT = 2000
|
MUTATION_AMOUNT = 200#1000000
|
||||||
ITERATIONS = 30
|
ITERATIONS = 100
|
||||||
THREAD_COUNT = 20
|
|
||||||
|
|
||||||
data = [line for line in open(INPUT_FILE)]
|
def mutation(entries):
|
||||||
params = list(map(int, data[0].split(" ")))
|
|
||||||
data = [[0 if x=="T" else 1 for x in line] for line in data[1:]]
|
|
||||||
data = np.array(data)[:, :-1]
|
|
||||||
clusters = np.arange(params[0]*params[1]).reshape((1, params[0], params[1]))
|
|
||||||
clusters = np.repeat(clusters, POPULATION, axis=0)+1
|
|
||||||
print(params)
|
|
||||||
print(data)
|
|
||||||
print(clusters[0])
|
|
||||||
|
|
||||||
values = {}
|
def mutation_entry(entry, args):
|
||||||
first = True
|
HEIGHT, WIDTH, MIN_PIECES, MAX_SIZE, _ = args
|
||||||
|
|
||||||
def get_fitness(vals, clean=False):
|
def expand_hztl(entry, x, y, direction):
|
||||||
fit = 0
|
val = entry[y, x]
|
||||||
for key, val in vals.items():
|
ynew = y+direction
|
||||||
if key == 0:
|
oval = entry[ynew, x]
|
||||||
continue
|
# left
|
||||||
size = sum(val)
|
nx = x
|
||||||
if size <= params[3] and min(val) >= params[2]:
|
while nx > 0 and val == entry[y, nx-1]:
|
||||||
fit += size
|
entry[ynew, nx] = val
|
||||||
if clean:
|
nx -= 1
|
||||||
continue
|
nval = entry[ynew, nx] if nx != x else oval
|
||||||
size_diff = params[3]-size
|
entry[ynew, nx] = val
|
||||||
if size_diff < 0:
|
if nval != 0:
|
||||||
fit += 1-size_diff**2
|
while nx > 0 and nval == entry[ynew, nx-1]:
|
||||||
elif size_diff > 0:
|
nx -= 1
|
||||||
fit += np.exp(-abs(size-params[3]))
|
entry[ynew, nx] = 0
|
||||||
return fit
|
# right
|
||||||
|
nx = x
|
||||||
|
while nx < WIDTH-1 and val == entry[y, nx+1]:
|
||||||
|
entry[ynew, nx] = val
|
||||||
|
nx += 1
|
||||||
|
nval = entry[ynew, nx] if nx != x else oval
|
||||||
|
entry[ynew, nx] = val
|
||||||
|
if nval != 0:
|
||||||
|
while nx < WIDTH-1 and nval == entry[ynew, nx+1]:
|
||||||
|
nx += 1
|
||||||
|
entry[ynew, nx] = 0
|
||||||
|
|
||||||
|
def expand_vert(entry, x, y, direction):
|
||||||
|
val = entry[y, x]
|
||||||
|
xnew = x+direction
|
||||||
|
oval = entry[y, xnew]
|
||||||
|
# left
|
||||||
|
ny = y
|
||||||
|
while ny > 0 and val == entry[ny-1, x]:
|
||||||
|
entry[ny, xnew] = val
|
||||||
|
ny -= 1
|
||||||
|
nval = entry[ny, xnew] if ny != y else oval
|
||||||
|
entry[ny, xnew] = val
|
||||||
|
if nval != 0:
|
||||||
|
while ny > 0 and nval == entry[ny-1, xnew]:
|
||||||
|
ny -= 1
|
||||||
|
entry[ny, xnew] = 0
|
||||||
|
# right
|
||||||
|
ny = y
|
||||||
|
while ny < HEIGHT-1 and val == entry[ny+1, x]:
|
||||||
|
entry[ny, xnew] = val
|
||||||
|
ny += 1
|
||||||
|
nval = entry[ny, xnew] if ny != y else oval
|
||||||
|
entry[ny, xnew] = val
|
||||||
|
if nval != 0:
|
||||||
|
while ny < HEIGHT-1 and nval == entry[ny+1, xnew]:
|
||||||
|
ny += 1
|
||||||
|
entry[ny, xnew] = 0
|
||||||
|
|
||||||
def get_left_bound(clust, y, x):
|
for _ in range(np.random.random_integers(MUTATION_AMOUNT)):
|
||||||
val = clust[y, x]
|
y = np.random.random_integers(HEIGHT)-1
|
||||||
while x > 0 and val == clust[y, x-1]:
|
x = np.random.random_integers(WIDTH)-1
|
||||||
x -= 1
|
if entry[y, x] == 0: # create new cluser
|
||||||
return (y, x)
|
entry[y, x] = np.amax(entry)+1
|
||||||
|
z = np.random.random()
|
||||||
|
if z < 0.25: # expand to top
|
||||||
|
if y > 0:
|
||||||
|
expand_hztl(entry, x, y, -1)
|
||||||
|
elif z < 0.5: # expand to left
|
||||||
|
if x > 0:
|
||||||
|
expand_vert(entry, x, y, -1)
|
||||||
|
elif z < 0.75: # expand to bottom
|
||||||
|
if y < HEIGHT-1:
|
||||||
|
expand_hztl(entry, x, y, 1)
|
||||||
|
else: # expand to right
|
||||||
|
if x < WIDTH-1:
|
||||||
|
expand_vert(entry, x, y, 1)
|
||||||
|
return entry
|
||||||
|
|
||||||
|
sub_arr, args = entries
|
||||||
|
for idx in range(sub_arr.shape[0]):
|
||||||
|
sub_arr[idx,:,:] = mutation_entry(sub_arr[idx,:,:], args)
|
||||||
|
return sub_arr
|
||||||
|
|
||||||
def get_right_bound(clust, y, x):
|
def get_fitnesses(entries):
|
||||||
val = clust[y, x]
|
def get_fitness(entry, args):
|
||||||
while x+1 < clust.shape[1] and val == clust[y, x+1]:
|
HEIGHT, WIDTH, MIN_PIECES, MAX_SIZE, (data, data_inv) = args
|
||||||
x += 1
|
|
||||||
return (y, x)
|
|
||||||
|
|
||||||
|
def get_fitness_per_cluster(cluster):
|
||||||
|
fit = 0
|
||||||
|
size = cluster[0]
|
||||||
|
a = 0
|
||||||
|
b = 0
|
||||||
|
if size <= MAX_SIZE and cluster[1] >= MIN_PIECES:
|
||||||
|
a = size
|
||||||
|
size_diff = MAX_SIZE-size
|
||||||
|
if size_diff < 0:
|
||||||
|
b = MAX_SIZE/8+size_diff
|
||||||
|
elif size_diff > 0:
|
||||||
|
b = np.power(3, -size_diff)*MAX_SIZE/8
|
||||||
|
else:
|
||||||
|
b = a+1
|
||||||
|
return [a, b] # a = clean score ; b = fitness score
|
||||||
|
|
||||||
def get_top_bound(clust, y, x):
|
mname, mcount = np.unique(entry*data, return_counts=True)
|
||||||
val = clust[y, x]
|
mdict = dict(zip(mname, mcount))
|
||||||
while y > 0 and val == clust[y-1, x]:
|
tname, tcount = np.unique(entry*data_inv, return_counts=True)
|
||||||
y -= 1
|
tdict = dict(zip(tname, tcount))
|
||||||
return (y, x)
|
c = np.array([[mdict.get(key) or 0, tdict.get(key) or 0] for key in (mdict.keys() | tdict.keys()) if key != 0])
|
||||||
|
c = np.vstack((np.sum(c, axis=1), np.min(c, axis=1))).T
|
||||||
|
return sum(np.apply_along_axis(get_fitness_per_cluster, 1, c))
|
||||||
|
|
||||||
|
sub_arr, args = entries
|
||||||
|
fitnesses = np.zeros((sub_arr.shape[0], 2))
|
||||||
|
for idx in range(sub_arr.shape[0]):
|
||||||
|
x = get_fitness(sub_arr[idx,:,:], args)
|
||||||
|
#print(x)
|
||||||
|
fitnesses[idx] = get_fitness(sub_arr[idx,:,:], args)
|
||||||
|
return fitnesses
|
||||||
|
|
||||||
def get_bottom_bound(clust, y, x):
|
if __name__ == '__main__':
|
||||||
val = clust[y, x]
|
import numpy as np
|
||||||
while y+1 < clust.shape[0] and val == clust[y+1, x]:
|
import multiprocessing
|
||||||
y += 1
|
|
||||||
return (y, x)
|
|
||||||
|
|
||||||
|
def thread_map(func, array, data=None):
|
||||||
|
chunks = [(sub_arr, (HEIGHT, WIDTH, MIN_PIECES, MAX_SIZE, data)) for sub_arr in np.array_split(array, min(multiprocessing.cpu_count(), len(array)))]
|
||||||
|
pool = multiprocessing.Pool()
|
||||||
|
individual_results = pool.map(func, chunks)
|
||||||
|
pool.close()
|
||||||
|
pool.join()
|
||||||
|
return np.concatenate(individual_results)
|
||||||
|
|
||||||
def set_area(clust, y1, x1, y2, x2, value):
|
data = [line for line in open(INPUT_FILE)]
|
||||||
for y in range(y1, y2+1):
|
params = list(map(int, data[0].split(" ")))
|
||||||
for x in range(x1, x2+1):
|
HEIGHT = params[0]
|
||||||
clust[y, x] = value
|
WIDTH = params[1]
|
||||||
return clust
|
MIN_PIECES = params[2]
|
||||||
|
MAX_SIZE = params[3]
|
||||||
|
data = [[0 if x=="T" else 1 for x in line] for line in data[1:]]
|
||||||
|
data = np.array(data)[:, :-1]
|
||||||
|
data_inv = 1-data
|
||||||
|
clusters = np.arange(HEIGHT*WIDTH).reshape((1, HEIGHT, WIDTH))
|
||||||
|
clusters = np.repeat(clusters, POPULATION, axis=0)+1
|
||||||
|
|
||||||
|
for iteration in range(ITERATIONS):
|
||||||
def mutation(clust):
|
# mutation
|
||||||
for _ in range(np.random.random_integers(MUTATION_AMOUNT)):
|
print(iteration, "Mutation")
|
||||||
y = np.random.random_integers(params[0])-1
|
clusters = thread_map(mutation, clusters)
|
||||||
x = np.random.random_integers(params[1])-1
|
# get fitness
|
||||||
z = np.random.random()
|
print(iteration, "Get Fitness")
|
||||||
if z < 0.2:
|
fitnesses = thread_map(get_fitnesses, clusters, (data, data_inv))
|
||||||
if y > 0: # expand to top
|
max_idx = np.argmax(fitnesses[:, 0])
|
||||||
yn = y-1
|
print(clusters[max_idx])
|
||||||
_, inner_left = get_left_bound(clust, y, x)
|
print(iteration, max(fitnesses[:, 0]))
|
||||||
_, outer_left = get_left_bound(clust, yn, inner_left)
|
# selection
|
||||||
_, inner_right = get_right_bound(clust, y, x)
|
print(iteration, "Selection")
|
||||||
_, outer_right = get_right_bound(clust, yn, inner_right)
|
z_exp = [np.exp(i) for i in fitnesses[:, 1]]
|
||||||
clust = set_area(clust, yn, outer_left, yn, outer_right, 0)
|
sum_z_exp = sum(z_exp)
|
||||||
clust = set_area(clust, yn, inner_left, yn, inner_right, clust[y, x])
|
softmax = [i / sum_z_exp for i in z_exp]
|
||||||
elif z < 0.4:
|
idx = np.random.choice(POPULATION, POPULATION, p=softmax)
|
||||||
if x > 0: # expand to left
|
clusters = clusters[idx, :, :]
|
||||||
xn = x-1
|
|
||||||
inner_top, _ = get_top_bound(clust, y, x)
|
|
||||||
outer_top, _ = get_top_bound(clust, inner_top, xn)
|
|
||||||
inner_bot, _ = get_bottom_bound(clust, y, x)
|
|
||||||
outer_bot, _ = get_bottom_bound(clust, inner_bot, xn)
|
|
||||||
clust = set_area(clust, outer_top, xn, outer_bot, xn, 0)
|
|
||||||
clust = set_area(clust, inner_top, xn, inner_bot, xn, clust[y, x])
|
|
||||||
elif z < 0.6:
|
|
||||||
if y < params[0]-1: # expand to bottom
|
|
||||||
yn = y+1
|
|
||||||
_, inner_left = get_left_bound(clust, y, x)
|
|
||||||
_, outer_left = get_left_bound(clust, yn, inner_left)
|
|
||||||
_, inner_right = get_right_bound(clust, y, x)
|
|
||||||
_, outer_right = get_right_bound(clust, yn, inner_right)
|
|
||||||
clust = set_area(clust, yn, outer_left, yn, outer_right, 0)
|
|
||||||
clust = set_area(clust, yn, inner_left, yn, inner_right, clust[y, x])
|
|
||||||
elif z < 0.8:
|
|
||||||
if x < params[1]-1: # expand to right
|
|
||||||
xn = x+1
|
|
||||||
inner_top, _ = get_top_bound(clust, y, x)
|
|
||||||
outer_top, _ = get_top_bound(clust, inner_top, xn)
|
|
||||||
inner_bot, _ = get_bottom_bound(clust, y, x)
|
|
||||||
outer_bot, _ = get_bottom_bound(clust, inner_bot, xn)
|
|
||||||
clust = set_area(clust, outer_top, xn, outer_bot, xn, 0)
|
|
||||||
clust = set_area(clust, inner_top, xn, inner_bot, xn, clust[y, x])
|
|
||||||
else:
|
|
||||||
pass#clust[y, x] = np.amax(clust)+1
|
|
||||||
return clust
|
|
||||||
|
|
||||||
|
|
||||||
|
|
||||||
|
|
||||||
|
|
||||||
|
|
||||||
def myfunc(a, b):
|
|
||||||
global first, values
|
|
||||||
if first:
|
|
||||||
first = False
|
|
||||||
return
|
|
||||||
if a not in values:
|
|
||||||
values[a] = [0, 0]
|
|
||||||
values[a][b] += 1
|
|
||||||
vfunc = np.vectorize(myfunc)
|
|
||||||
|
|
||||||
# mutation
|
|
||||||
for i in range(POPULATION):
|
|
||||||
if i % 20 == 0:
|
|
||||||
print("mutation", i)
|
|
||||||
clusters[i] = mutation(clusters[i])
|
|
||||||
|
|
||||||
for iteration in range(ITERATIONS):
|
|
||||||
# calc fitness
|
|
||||||
fitnesses = np.zeros((POPULATION, ))
|
|
||||||
for i, cluster in enumerate(clusters):
|
|
||||||
if i % 20 == 0:
|
|
||||||
print("fitness", i, iteration)
|
|
||||||
values = {}
|
|
||||||
first = True
|
|
||||||
vfunc(cluster, data)
|
|
||||||
fitnesses[i] = get_fitness(values)
|
|
||||||
# select
|
|
||||||
z_exp = [np.exp(i) for i in fitnesses]
|
|
||||||
sum_z_exp = sum(z_exp)
|
|
||||||
softmax = [i / sum_z_exp for i in z_exp]
|
|
||||||
idx = np.random.choice(POPULATION, POPULATION, p=softmax)
|
|
||||||
clusters = clusters[idx, :, :]
|
|
||||||
|
|
||||||
# print best
|
|
||||||
max_idx = np.argmax(fitnesses)
|
|
||||||
print(clusters[max_idx])
|
|
||||||
print(iteration, max(fitnesses))
|
|
||||||
|
|
||||||
# mutation
|
|
||||||
for i in range(POPULATION):
|
|
||||||
clusters[i] = mutation(clusters[i])
|
|
||||||
|
|
||||||
fitnesses = np.zeros((POPULATION, ))
|
|
||||||
for i, cluster in enumerate(clusters):
|
|
||||||
values = {}
|
|
||||||
first = True
|
|
||||||
vfunc(cluster, data)
|
|
||||||
fitnesses[i] = get_fitness(values, clean=True)
|
|
||||||
max_idx = np.argmax(fitnesses)
|
|
||||||
print(clusters[max_idx])
|
|
||||||
print(max(fitnesses))
|
|
||||||
|
|||||||
Reference in New Issue
Block a user