def simple_reduction(puzzle):
"""
simple_reduction returns a solution to <puzzle>.
It works by reducing the number of greenhouses one by one until it has the
lowest cost and meets the max constraint.
"""
max, field = puzzle
# figure out the current number of greenhouses
greenhouses = common.ids(field)
# we need to keep a copy of the previous field and it's cost in order
# to return it once we've realized we've done one reduction too many
prev_field, prev_cost = None, sys.maxint
if len(greenhouses) <= max:
prev_field, prev_cost = copy.deepcopy(field), common.cost(field)
# join greenhouses until when run out of them or until max constraint
# is met *and* cost increases from one reduction to the next
while len(greenhouses) > 1:
j1, j2, js = 0, 0, sys.maxint
# try each combination of greenhouses
for g1, g2 in itertools.combinations(greenhouses, 2):
# find outer bounds (left, right, top and bottom) for a greenhouse made
# up of g1 and g2
size3, p31, p32 = common.outer_bounds([g1, g2], field)
if size3 is not None:
size1, p11, p12 = common.outer_bounds(g1, field)
size2, p21, p22 = common.outer_bounds(g2, field)
diff = size3 - size2 - size1
if diff < js:
j1, j2, js = g1, g2, diff
# if we run out of combinations to try
# we must either surrender (return None)
# or if len(greenhouses) <= max return
# the best solution we have.
if j1 == 0:
if len(greenhouses) <= max:
return max, prev_field
else:
return max, None
# join j1 and j2, remove j2 from greenhouses
field = common.join(j1, j2, field)
greenhouses.remove(j2)
# decide if we should exit this loop or keep on reducing
curr_cost = common.cost(field)
if len(greenhouses) < max:
if prev_cost < curr_cost:
return max, prev_field
prev_field, prev_cost = copy.deepcopy(field), curr_cost
# if we end up here, we've come down to 1 greenhouse
return max, field
def find_template_exhaustive(test, ref):
st = time.time()
then = st
d_thresh = ref.shape[0] * ref.shape[1] * 0.001
#Each pixel was an integer from 0 to 255 before we normalized them. So minimum difference between
#two pixels is 1/255 = 0.0039.. becomes 0.000015... if squared. And we are allowing average squared difference
#of 0.001
print('Acceptable cost value:', d_thresh)
coord = (0, 0)
d = float('inf')
max_h, max_w = test.shape[0] - ref.shape[0], test.shape[1] - ref.shape[1]
after_print = 0
for i in range(0, max_h + 1): # from 0 to test_h - ref_h
#log---------
if after_print > 1000:
now = time.time()
if now - then > 1:
print('Scanning in row', i, '... d=', d)
then = now
after_print = 0
###logend-----
for j in range(0, max_w + 1): # from 0 to test_w - ref_w
d_cur = cost(test, ref, i, j)
if d_cur < d:
d = d_cur
coord = (j, i)
if d < d_thresh:
return coord, False, time.time() - st
after_print += max_w
return coord, True, time.time() - st
def heirarchical_recursive(test, ref, level):
if level == 0:
coord, _, _ = find_template_exhaustive(test/255.0, ref/255.0)
return coord
test2 = cv2.resize(test, None, fx=0.5, fy=0.5)
ref2 = cv2.resize(ref, None, fx=0.5, fy=0.5)
coord2 = heirarchical_recursive(test2, ref2, level - 1)
coord = center = (coord2[0]*2, coord2[1]*2)
#print(coord)
search_area = (0, 0, test.shape[1] - ref.shape[1] - 1, test.shape[0] - ref.shape[0] - 1) # top left and bottom right
d = float('inf')
for dx in [0, 1, -1]:
for dy in [0, 1, -1]:
j, i = center[0] + dx , center[1] + dy
if j < search_area[0] or j > search_area[2] or i < search_area[1] or i > search_area[3]:
continue
# print(j, i, dwidth, dheight, d)
d_cur = cost(test/255.0, ref/255.0, i, j)
if d_cur < d:
d = d_cur
coord = (j, i)
return coord
def solve(filename):
"""
solve prints out solutions to each of the fields described in <filename>.
"""
tstart = time.time()
count, total = 0, 0
for puzzle in common.parse_file(filename):
pstart = time.time()
max, field = variant_reduction(s0.join_vertically(s0.join_horizontally(s0.identify(puzzle))))
count += 1
total += common.cost(field)
print "cost:", common.cost(field), "time:", time.strftime("%H:%M:%S", time.gmtime(time.time() - pstart))
print common.format(field)
print "%s field(s). Total cost is $%s" % (count, total)
print time.strftime("Total time is %H:%M:%S", time.gmtime(time.time() - tstart))
def variant_reduction(puzzle):
"""
variant_reduction flips and rotates <puzzle> to find a better solution.
"""
max, field = puzzle
# there are 8 variations to each puzzle, these 4 + 4 rotated ones
variants = [lambda f: f, lambda f: flip_hz(f), lambda f: flip_vt(f), lambda f: flip_vt(flip_hz(f))]
rotated_field = rotate(copy.deepcopy(field))
best_cost, best_field = sys.maxint, None
for vi in xrange(4):
_, _field = simple_reduction((max, variants[vi](copy.deepcopy(field))))
if common.cost(_field) < best_cost:
best_cost, best_field = common.cost(_field), variants[vi](_field)
_, _field = simple_reduction((max, variants[vi](copy.deepcopy(rotated_field))))
if common.cost(_field) < best_cost:
best_cost, best_field = common.cost(_field), unrotate(variants[vi](_field))
return max, best_field
def logsearch(test, ref, search_area):
d_thresh = ref.shape[0] * ref.shape[1] * 0.001
# Each pixel was an integer from 0 to 255 before we normalized them. So minimum difference between
# two pixels is 1/255 = 0.0039.. becomes 0.000015... if squared. And we are allowing average squared difference
# of 0.001
print('Acceptable cost value:', d_thresh)
#print(search_area)
# search width = p, height = q
p = search_area[2] - search_area[0]
q = search_area[3] - search_area[1]
dwidth = 2**(math.ceil(math.log2(p / 2)) - 1)
dheight = 2**(math.ceil(math.log2(q / 2)) - 1)
coord = center = (p // 2, q // 2)
# print(dwidth, dheight)
d = cost(test, ref, center[1], center[0])
while dwidth >= 1 and dheight >= 1:
for dx in [0, 1, -1]:
for dy in [0, 1, -1]:
j, i = center[0] + dx * dwidth, center[1] + dy * dheight
if (dx == 0
and dy == 0) or j < search_area[0] or j > search_area[
2] or i < search_area[1] or i > search_area[3]:
continue
#print(j, i, dwidth, dheight, d)
d_cur = cost(test, ref, i, j)
if d_cur < d:
d = d_cur
coord = (j, i)
if d < d_thresh:
return coord, False
dwidth //= 2
dheight //= 2
center = coord
return coord, True
def simplified_bruteForce(R, ax=None, aCap=0.20, beta=1.5, anim=False):
"""
a simplified algorithm. Still a little bit complex dock, but this is as simple as it gets I think.
Does not add any edges, only removal. No stochastic elements are involved.
anim=True creates a movie.
"""
R.beta=beta
inf = 1e15
eps=1e-9
lastAdded=None
origin=R.origin
for e in R.edges(data=True): assert e[2]['weight']>=0
if not origin: raise Exception('need info about origin')
#now, start for real and save away all kind of info about the paths.
for node in R.nodes(data=True):
p1,p2=get_shortest_and_second(R,node)
node[1]['shortest_path']=p1
if len(p1)==0: #origin
assert node[0]==origin
node[1]['second_shortest']=[]
else:
node[1]['second_shortest']=p2
for p in p1,p2: #now, update edge info for all visited edges
for edge in R.edges_from_path_gen(p):
d=R.get_edge_data(*edge)
if not node in d['visited_from_node']: d['visited_from_node'].append(node) #in order to later see...
remList=[]
for e in R.edges(data=False): #add to remove list and calc. costs
e_data=R.get_edge_data(*e)
e_data['c']=cost(R, e, storeData=True)
remList.append(e)
remList=sortRemList(R,remList)
while len(remList)>0: #the loop where edges are removed..
"""
now, choose the edge with the lowest cost.
We have a problem since the cost saved is not necessary updated to the correct value.
Actually, the only way to update it correctly is to scan through all edges for every removal. One could think that it would be possible to only update the ones affected by the removal, i.e. the ones connected by the shortest roads. The problem is that we also need to account for the ones that would take this road if some other arbitrary road was removed. Storing that variable would be possible but very memory inefficient. It could be tried and developed further, but would only be necessary if we add edges because the below alg. works pretty good.
We use the fact that the cost from removing edges can only be bigger, i.e when removing edges it only gets worse.
Thus, if the updated cost is still the smallest in the list, we know that this road is cheapest to remove.
"""
print "start"
if anim: #movietime
R.movieFlush()
while True: #two possibilities to break out below..
e=remList[0]
e_data=R.get_edge_data(*e)
e_data['c']=cost(R,e,storeData=False)
if len(remList)==1:
break
#now, look for the other one..
e2=remList[1]
e2_data=R.get_edge_data(*e2)
if e_data['c']<=e2_data['c']:
break #e is our candidate, we know it's best
e2_data['c']=cost(R, e2, storeData=False) #update, can only get worse
for e in e,e2: #check for infinite costs, remove in that case.
d=R.get_edge_data(*e) #this c has just been updated.
if d['c']>=inf:
remList.remove(e)
remList=sortRemList(R,remList) #sort it, try again..
if e_data['c']>=inf: break #if last in remList and infinite cost..
remList.remove(e)
e_data['c']=cost(R,e,storeData=True) #in order to store the new path..
assert e_data['c']<inf
#we are outside... will we go over the areaLimit if we remove e?
if e_data['c']>eps and R.areaCover-go.singleRoadSegmentCoverage(e, R, remove=True)*R.Ainv<aCap:
assert abs(R.areaCover-go.roadAreaCoverage(R))<eps #compare internal and actual.
break #we are finished
assert R.degree(e[0])>2 and R.degree(e[1])>2 #cost func should have given c=inf.
print "removes edge ",e, e_data['c']
remove_edge(e, R) #remove from R.
if anim:
R.movieFlush(final=True)
R.cost=cf.totalCost(R)
print "construction finished."
print "road area coverage:", R.areaCover
print "total area:", R.A
print "total cost:", R.cost
return R
def search(puzzle, breakpoint = 2):
"""
search produces a solution to <puzzle>.
>>> solve("p3.text") # doctest: +ELLIPSIS
71
...
"""
max, field = puzzle
solution = (common.cost(field), field)
paths = [solution, ]
while len(paths) > 0:
curr_cost, field = paths.pop(0)
# Figure out the current number of greenhouses
greenhouses = common.ids(field)
if len(greenhouses) > 1:
diffs = {}
# Try each combination of greenhouses
for g1, g2 in itertools.combinations(greenhouses, 2):
# Find outer bounds (left, right, top and bottom) for a greenhouse made
# up of g1 and g2
size3, p31, p32 = common.outer_bounds([g1, g2], field)
if size3 is not None:
size1, p11, p12 = common.outer_bounds(g1, field)
size2, p21, p22 = common.outer_bounds(g2, field)
diff = size3 - size2 - size1
if diff not in diffs.keys():
diffs[diff] = [(g1, g2),]
else:
diffs[diff].append((g1, g2))
# Find the list of joins which has the lowest diff and select the joins
# of the most frequent greenhouse.
if len(diffs.keys()) > 0:
freqs = {}
for (g1, g2) in diffs[sorted(diffs.keys())[0]]:
if g1 not in freqs.keys():
freqs[g1] = [(g1, g2),]
else:
freqs[g1].append((g1, g2))
if g2 not in freqs.keys():
freqs[g2] = [(g1, g2),]
else:
freqs[g2].append((g1, g2))
# Perform each join in a fresh copy of field and add it to paths if
# cost is lower than current cost, otherwise compare cost to solution
# and either discard this path or add it as best-so-far.
joins = freqs[sorted(freqs.keys(), key = lambda k: len(freqs[k]), reverse = True)[0]]
if len(joins) <= breakpoint:
(g1, g2) = joins[0]
_, _field = s1.simple_reduction((max, common.join(g1, g2, copy.deepcopy(field))))
cf = common.cost(_field)
if cf < solution[0] and \
len(common.ids(_field)) <= max:
solution = (cf, _field)
else:
for (g1, g2) in joins:
_field = common.join(g1, g2, copy.deepcopy(field))
cf = common.cost(_field)
if cf < curr_cost:
paths.append((cf, _field))
if cf < solution[0] and \
len(common.ids(_field)) <= max:
solution = (cf, _field)
return max, solution[1]
def stochastic(R, ax=None, aCap=0.20, beta=1.5, anim=False, probListGen=None):
"""
a simplified algorithm. Still a little bit complex dock, but this is as simple as it gets I think.
Does not add any edges, only removal. No stochastic elements are involved.
anim=True creates a movie.
probListGen can be given. That is a class of type ProbListGen that defines a specific
distribution that is used by the stochastic parts.
"""
if probListGen==None: #use default, 0.5^i
probListGen=ProbListGen(0.5,15)
R.beta=beta
inf = 1e15
eps=1e-9
choiceMax=15 #we will not randomly choose something bigger than this.
lastAdded=None
origin=R.origin
for e in R.edges(data=True): assert e[2]['weight']>=0
if not origin: raise Exception('need info about origin')
#now, start for real and save away all kind of info about the paths.
for node in R.nodes(data=True):
p1,p2=get_shortest_and_second(R,node)
node[1]['shortest_path']=p1
if len(p1)==0: #origin
assert node[0]==origin
node[1]['second_shortest']=[]
else:
node[1]['second_shortest']=p2
for p in p1,p2: #now, update edge info for all visited edges
for edge in R.edges_from_path_gen(p):
d=R.get_edge_data(*edge)
if not node in d['visited_from_node']:
d['visited_from_node'].append(node) #in order to later see...
remList=[]
for e in R.edges(data=False): #add to remove list and calc. costs
e_data=R.get_edge_data(*e)
e_data['c']=cost(R, e, storeData=True)
remList.append(e)
remList=sortRemList(R,remList)
choices=[0]*15 #for statistics areound the stochastics.
while len(remList)>0: #the loop where edges are removed..
"""
We have a problem since the cost saved is not necessary updated to the correct value.
Actually, the only way to update it correctly is to scan through all edges for every removal. One could think that it would be possible to only update the ones affected by the removal, i.e. the ones connected by the shortest roads. The problem is that we also need to account for the ones that would take this road if some other arbitrary road was removed. Storing that variable would be possible but very memory inefficient. It could be tried and developed further, but would only be necessary if we add edges because the below alg. works pretty good.
We use the fact that the cost from removing edges can only be bigger, i.e when removing edges it only gets worse.
Thus, if the updated cost is still the smallest in the list, we know that this road is cheapest to remove.
"""
print "start"
probList=probListGen.getList(N=len(remList))
if anim: #it's showtime..
R.movieFlush()
r=random.uniform(0,1)
choice=choiceMax
for i,p in enumerate(probList): #time to make the choice..
if r<p:
choice=i #usually 0.. 50% prob of that.
break
updated=[] #store edges that we have updated the cost for..
done=False
while not done: #two possibilities to break out below..
done=True
for i in range(choice+2): #loop over all necessary edges
try:
e=remList[i]
except IndexError: #remList is too short.. we are done here...
done=True
break #breaks out of for-loop, not while loop
e_data=R.get_edge_data(*e)
if not e in updated: #saves us some calculations
done=False #have to iterate once more.
e_data['c']=cost(R,e,storeData=False)
updated.append(e)
if e_data['c']>=inf:
remList.remove(e)
break #otherwise we "jump over" one in the list. new for loop
if e_data['c']>=inf:
remList.remove(e)
remList=sortRemList(R,remList) #sort it..
remList=sortRemList(R,remList) #last time.. may be a double sort but who cares?
#sorting remList and updating cost procedure is now done.
if len(remList)==0: break #we are done
if choice>=len(remList): #may happen if edges have been removed due to inf. cost
choice=int(floor(random.uniform(0,len(remList)))) #take a random one..
print "choice:", choice
e=remList[choice]
e_data=R.get_edge_data(*e)
remList.remove(e)
e_data['c']=cost(R,e,storeData=True) #in order to store the new path..
assert e_data['c']<inf
#following lines just to check, not used... remove later..
if __debug__ and choice != -1 and len(remList)>choice+1:
e2=remList[choice+1]
e2_data=R.get_edge_data(*e2)
assert e_data['c']<=e2_data['c']
#we are outside... will we go over the areaLimit if we remove e?
if e_data['c']>eps and R.areaCover-go.singleRoadSegmentCoverage(e, R, remove=True)*R.Ainv<aCap:
assert abs(R.areaCover-go.roadAreaCoverage(R))<eps #compare internal and actual.
break #we are finished
assert R.degree(e[0])>2 and R.degree(e[1])>2 #cost func should have given c=inf.
print "removes edge ",e, e_data['c']
remove_edge(e, R) #remove from R.
choices[choice]+=1
if anim:
R.movieFlush(final=True)
print "construction finished."
print "road area coverage:", R.areaCover
print "total area:", R.A
print "choices:", choices
R.cost=cf.totalCost(R)
print "total cost:", R.cost
return R
