1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
|
import sync
import sys
import thread
import time
from numpy import arange, shape, zeros
import remote_exec
import population
######
#I've got to clean up evaluate and initial in population so that
#the incorporation of the parallel stuff is smoother.
######
def array_round(x):
y = zeros(shape(x))
for i in range(len(x.flat)):
y[i] = int(round(x[i]))
return y
def divide_list(l,sections):
Ntot = len(l)
Nsec = float(sections)
Neach = Ntot/Nsec
div_points = array_round(arange(0,Ntot,Neach)).tolist()
if div_points[-1] != Ntot: div_points.append(Ntot)
sub_pops = []
st = div_points[0]
for end in div_points[1:]:
sub_pops.append(l[st:end])
st = end
return sub_pops
class parallel_pop_initializer:
def evaluate(self,pop,settings = None):
#only send the individuals out that need evaluation
if len(pop):
Nserv = len(pop.server_list)
groups = divide_list(pop,Nserv)
sys.setcheckinterval(10)
finished = sync.event()
bar = sync.barrier(Nserv)
print '************',len(groups), len(pop.server_list), len(pop)
for i in range(len(groups)):
inputs = {'sub_pop':groups[i],'settings':settings, 'initializer':pop.initializer}
returns = ('sub_pop',)
code = 'initializer.evaluate(sub_pop,settings)'
data_pack = (inputs,returns,code)
server = pop.server_list[i]
thread.start_new_thread(remote_thread_init,(bar,finished,server,data_pack))
finished.wait()
sys.setcheckinterval(10)
#what is this? for ind in pop: ind.evaluate(force)
import cPickle
def plen(obj): return len(cPickle.dumps(obj,1))
class parallel_pop_evaluator:
def evaluate(self,pop,force = 0):
#import tree
#print '1',tree.ref()
#only send the individuals out that need evaluation
if force:
_eval_list = pop.data
else:
_eval_list = filter(lambda x: not x.evaluated,pop)
#print '2',tree.ref()
eval_list = pop.clone()
#print '3',tree.ref()
eval_list.data = _eval_list
if len(eval_list):
Nserv = len(pop.server_list)
groups = divide_list(eval_list,Nserv)
#print '4',tree.ref()
sys.setcheckinterval(10)
finished = sync.event()
bar = sync.barrier(Nserv)
#print "EVAL LENGTH!!!", plen(pop.evaluator)
gr = groups[0]
print "GROUP LENGTH!!!", plen(groups[0]), len(gr),
#print "IND!!!", plen(gr[0]),plen(gr[0].root)
#print '4.5',tree.ref()
for i in range(len(groups)):
inputs = {'sub_pop':groups[i], 'evaluator':pop.evaluator, 'force':force}
returns = ('sub_pop',)
code = 'evaluator.evaluate(sub_pop,force)'
data_pack = (inputs,returns,code)
server = pop.server_list[i]
thread.start_new_thread(remote_thread_eval,(bar,finished,server,data_pack))
#print '7',tree.ref()
finished.wait()
sys.setcheckinterval(10)
#what is this? for ind in pop: ind.evaluate(force)
"""
def evaluate(self,pop,force = 0):
#only send the individuals out that need evaluation
_eval_list = filter(lambda x: not x.evaluated,pop)
eval_list = pop.clone()
eval_list.data = _eval_list
if len(eval_list):
#finest grain possible
groups = divide_list(eval_list,len(eval_list))
finished = sync.event()
bar = sync.barrier(groups)
sys.setcheckinterval(10)
Nserv = len(pop.server_list)
idx = 0
while idx < len(groups):
inputs = {'sub_pop':groups[idx], 'evaluator':pop.evaluator}
returns = ('sub_pop',)
code = 'evaluator.evaluate(sub_pop)'
data_pack = (inputs,returns,code)
server = pop.server_list[i]
thread.start_new_thread(remote_thread_eval,(bar,finished,server,data_pack))
#for i in range(len(groups)):
# inputs = {'sub_pop':groups[i], 'evaluator':pop.evaluator}
# returns = ('sub_pop',)
# code = 'evaluator.evaluate(sub_pop)'
# data_pack = (inputs,returns,code)
# server = pop.server_list[i]
# thread.start_new_thread(remote_thread,(bar,finished,server,data_pack))
finished.wait()
sys.setcheckinterval(10)
#what is this? for ind in pop: ind.evaluate(force)
"""
def remote_thread_init(bar,finished,server,data_pack):
try:
remote = remote_exec.remote_exec(server[0],server[1],0,1)
results = remote.run(data_pack)
#assign the results from the returned data to the local individuals
inputs = data_pack[0]
old = inputs['sub_pop']
new = results['sub_pop']
for i in range(len(old)):
old[i].__dict__.update(new[i].__dict__)
except IndexError:
print 'error in %s,%d' % server
bar.enter()
finished.post()
def remote_thread_eval(bar,finished,server,data_pack):
#import tree
try:
#print '5',tree.ref()
remote = remote_exec.remote_exec(server[0],server[1],0,1)
results = remote.run(data_pack)
#print '6',tree.ref()
#assign the results from the returned data to the local individuals
inputs = data_pack[0]
old = inputs['sub_pop']
new = results['sub_pop']
for gnm in new:
gnm.root.delete_circulars()
del gnm.root
#print '6.25',tree.ref()
for i in range(len(old)):
old[i].__dict__.update(new[i].__dict__)
#print '6.5',tree.ref()
except IndexError:
print 'error in %s,%d' % server
"""
import sys
#r = new[0].root
#print 'ref count',sys.getrefcount(r)
#print '6.75',tree.ref()
#Huh??? Why do I need to delete the new genomes
#individually here? Why aren't they garbage collected?
indices = range(len(new))
indices.reverse()
for i in indices:
del new[i]
#print 'ref count',sys.getrefcount(r)
#print '6.8',tree.ref()
#r.delete_circulars()
#print 'ref count',sys.getrefcount(r)
#print '6.9',tree.ref()
#del r
#print '6.95',tree.ref()
"""
bar.enter()
finished.post()
class ga_parallel_pop(population.population):
parallel_evaluator = parallel_pop_evaluator()
parallel_initializer = parallel_pop_initializer()
def __init__(self,genome,size=1,server_list=None):
"""Arguments:
genome -- a genome object.
size -- number. The population size. The genome will be
replicated size times to fill the population.
server_list -- a list of tuple pairs with machine names and
ports listed for the available servers
ex: [(ee.duke.edu,8000),('elsie.ee.duke.edu',8000)]
"""
population.population.__init__(self,genome,size)
assert(server_list)
self.server_list = server_list
def initialize(self,settings = None):
"""This method **must** be called before a genetic algorithm
begins evolving the population. It takes care of initializing
the individual genomes, evaluating them, and scaling the population.
It also clears and intializes the statistics for the population.
Arguments:
settings -- dictionary of genetic algorithm parameters. These
are passed on to the genomes for initialization.
"""
self.stats = {'current':{},'initial':{},'overall':{}}
self.stats['ind_evals'] = 0
print "beigninning genome generation"
b = time.clock()
self.parallel_initializer.evaluate(self,settings)
e = time.clock()
print "finished generation: ", e-b
self.touch();
b = time.clock()
self.evaluate()
e = time.clock()
print "evaluation time: ", e-b
self.scale()
self.update_stats()
self.stats['initial']['avg'] = self.stats['current']['avg']
self.stats['initial']['max'] = self.stats['current']['max']
self.stats['initial']['min'] = self.stats['current']['min']
self.stats['initial']['dev'] = self.stats['current']['dev']
def evaluate(self, force = 0):
""" call the parallel_evaluator instead of the evaluator directly
"""
self.selector.clear()
self.parallel_evaluator.evaluate(self,force)
#self.post_evaluate()
#all of the remaining should be put in post eval...
self.sort()
#this is a cluge to get eval count to work correctly
preval = self.stats['ind_evals']
for ind in self:
self.stats['ind_evals'] = self.stats['ind_evals'] + ind.evals
ind.evals = 0
print 'evals: ', self.stats['ind_evals'] - preval
self.touch()
self.evaluated = 1
########################## test stuff ############################
#import genome
#import gene
#import time
#
#import socket
#
#class objective:
# def __init__(self,wait=.01):
# self.wait = wait
# def evaluate(self,genome):
# time.sleep(self.wait)
# return sum(genome.array(),axis=0)
#
#def test_pop(server_list,size=100,wait=.01):
# obj = objective(wait)
# the_gene = gene.float_gene((0,2.5))
# genome_ = genome.list_genome(the_gene.replicate(5))
# genome_.evaluator = obj
# pop = ga_parallel_pop(genome_,size,server_list)
# print '########### awaiting evaluation#############'
# pop.initialize()
# print ' evaluation done!'
# print 'best:', pop.best()
# print 'worst',pop.worst()
#
#
#def gen_pop():
# genome.list_genome.evaluator = objective()
# gene = gene.float_gene((0,2.5))
# genome_ = genome.list_genome(gene.replicate(5))
# pop = ga_parallel_pop(genome_,100,[(host,port),])
# return pop
#
#import os
#
#import parallel_pop
#
#
#def test_pop2(server_list,size=100,wait=.01):
# import hmm_gnm,os
# genome = hmm_gnm.make_genome()
# #pop = ga_parallel_pop(genome,4,server_list)
# global galg
# #genome.target = targets[0]
# pop = ga_parallel_pop(genome,1,server_list)
# galg = hmm_gnm.class_ga(pop)
# galg.settings.update({ 'pop_size':6,'gens':2,'p_mutate':.03,
# 'dbase':os.environ['HOME'] + '/all_lift3', 'p_cross':0.9, 'p_replace':.6,
# 'p_deviation': -.001})
# galg.evolve()
#
# print '########### awaiting evaluation#############'
# pop.initialize()
# print ' evaluation done!'
# print 'best:', pop.best()
# print 'worst',pop.worst()
#
#import thread
#def test():
# host = socket.gethostname()
# port = 8000
# server_list = [(host,port),(host,port+1)]
# for server in server_list:
# host,port = server
# thread.start_new_thread(remote_exec.server,(host,port))
# thread.start_new_thread(test_pop2,(server_list,))
#
#def test2(machines=32,size=100,wait=.01):
# import time
# t1 = time.time()
# #requires that servers are started on beowulf 1 and 2.
# import beowulf
# server_list = beowulf.beowulf.servers[:machines]
# thread.start_new_thread(test_pop,(server_list,size,wait))
# print 'total time:', time.time()-t1
|
