-
Notifications
You must be signed in to change notification settings - Fork 0
/
playout_eval.py
313 lines (257 loc) · 10.8 KB
/
playout_eval.py
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
from cnf_util import *
from Game import ProverAdversaryGame
import numpy as np
from query_model import *
from gen_fmlas import get_unsat_randkcnf, get_unsat_sr, get_unsat_src
from Z3 import *
from types import SimpleNamespace
import ray
import numpy as np
from util import *
# from playout_eval import Match
def softmax_sample(action_logit_list, tau=3): # TODO(jesse): test
logits = [x[1] for x in action_logit_list]
actions = [x[0] for x in action_logit_list]
ps = tf.nn.softmax(np.array(logits, dtype="float32"))
selected_action = random.choices(actions, ps)[0]
return selected_action
class NeuroDRATProver():
def __init__(self, model_cfg):
self.oracle = NeuroCuberQuery(model_cfg)
def get_action(self, image, legal_actions):
drat_logits = self.oracle.get_logits(image)[0]
action_logit_list = [(i, drat_logits[i]) for i in legal_actions]
return softmax_sample(action_logit_list)
class NeuroCoreVarProver():
def __init__(self, model_cfg):
self.oracle = NeuroCuberQuery(model_cfg)
def get_action(self, image, legal_actions):
drat_logits = self.oracle.get_logits(image)[1]
action_logit_list = [(i, drat_logits[i]) for i in legal_actions]
return softmax_sample(action_logit_list)
class RandomProver():
def __init__(self):
pass
def get_action(self, image, legal_actions):
choice = int(np.random.choice(legal_actions))
return choice
class RandomAdversary():
def __init__(self):
pass
def __call__(self, image, legal_actions):
choice = bool(np.random.choice(legal_actions))
return choice
class Match:
def __init__(self, game, prover, adversary):
self.game = game
self.prover = prover
self.adversary = adversary
self.start_fmla = self.game.fmla
self.move_history = []
self.terminal_value = None
self.unit_prop_counts = []
def statistics(self):
return np.array([np.mean(self.unit_prop_counts), self.terminal_value])
def play_round(self):
assert not self.game.DONE_FLAG
image = self.game.make_image(0, True)
legal_actions = self.game.legal_actions()
if self.game.PROVER_TURN:
action = self.prover.get_action(image, legal_actions) # call remote prover
else:
action = self.adversary(image, legal_actions)
if not self.game.PROVER_TURN:
old_trail_length = len(self.game.trail) + 1
self.game.apply(action)
if self.game.PROVER_TURN:
self.unit_prop_counts.append(len(self.game.trail) - old_trail_length)
def play_game(self):
while not self.game.DONE_FLAG:
self.play_round()
# self.move_history = [(x[1], x[2]) for x in self.game.history]
self.terminal_value = self.game.terminal_value
return self.statistics()
@ray.remote
class NeuroProverService:
def __init__(self, model_cfg):
self.oracle = NeuroCuberQuery(model_cfg)
@ray.method(num_return_vals=1)
def get_action(self, image, legal_actions):
drat_logits = self.oracle.get_logits(image)[0]
action_logit_list = [(i, drat_logits[i]) for i in legal_actions]
return softmax_sample(action_logit_list)
class RemoteMatch: # initialized with actor handle of a NeuroProverService
def __init__(self, game, prover, adversary):
self.game = game
self.prover = prover
self.adversary = adversary
self.start_fmla = self.game.fmla
self.move_history = []
self.terminal_value = None
self.unit_prop_counts = []
def statistics(self):
return [np.mean(self.unit_prop_counts), self.terminal_value]
def play_round(self):
assert not self.game.DONE_FLAG
image = self.game.make_image(0, True)
legal_actions = self.game.legal_actions()
if self.game.PROVER_TURN:
action = ray.get(self.prover.get_action.remote(image, legal_actions)) # call remote prover
else:
action = self.adversary(image, legal_actions) # assume adversary is local
if not self.game.PROVER_TURN:
old_trail_length = len(self.game.trail) + 1
self.game.apply(action)
if self.game.PROVER_TURN:
self.unit_prop_counts.append(len(self.game.trail) - old_trail_length)
def play_game(self):
while not self.game.DONE_FLAG:
self.play_round()
self.terminal_value = self.game.terminal_value
return self.statistics()
@ray.remote
def play_match_remote(fmla, prover, adversary): # accepts reference to a prover service instead of instantiating own prover
match = RemoteMatch(ProverAdversaryGame(fmla),prover,adversary)
return match.play_game()
@ray.remote
def play_match(fmla, model_cfg, adversary): # TODO(jesse): make more efficient
prover = NeuroDRATProver(model_cfg)
match = Match(ProverAdversaryGame(fmla),prover,adversary)
return match.play_game()
def play_match_single(fmla, model_cfg, adversary): # TODO(jesse): make more efficient
prover = NeuroDRATProver(model_cfg)
match = Match(ProverAdversaryGame(fmla),prover,adversary)
return match.play_game()
def ParallelFormulaPlayout(fmla, model_cfg, adversary, num_matches):
jobs = []
for _ in range(num_matches):
jobs.append(play_match.remote(fmla, model_cfg, adversary))
return jobs
def SequentialFormulaPlayout(fmla, model_cfg, adversary, num_matches):
results = []
for _ in range(num_matches):
results.append(play_match_single(fmla, model_cfg, adversary))
return results
def test_running_time():
ray.init()
fmla = get_unsat_randkcnf(3,70)
NUM_MATCHES = 100
start = time.time()
jobs = ParallelFormulaPlayout(fmla, "res_models/res_grid_0_randkcnf.json", RandomAdversary(), NUM_MATCHES)
stats = np.mean(ray.get(jobs), axis=0)
elapsed1 = time.time() - start
print(stats)
ray.shutdown()
start = time.time()
results = SequentialFormulaPlayout(fmla, "res_models/res_grid_0_randkcnf.json", RandomAdversary(), NUM_MATCHES)
stats = np.mean(results, axis=0)
print(stats)
print("elapsed time", elapsed1)
# print("elapsed time", elapsed2)
print("elapsed time", time.time() - start)
@ray.remote
class FormulaWorker:
def __init__(self, prover_type, model_cfg, adversary, index):
if prover_type == "z3":
self.prover = Z3Prover()
elif prover_type == "random":
self.prover = RandomProver()
elif prover_type == "drat":
assert model_cfg is not None
self.prover = NeuroDRATProver(model_cfg)
elif prover_type == "core":
assert model_cfg is not None
self.prover = NeuroCoreVarProver(model_cfg)
else:
raise Exception("unsupported prover type")
self.adversary = RandomAdversary()
self.index = index
@ray.method(num_return_vals=1)
def play_match(self, fmla):
match = Match(ProverAdversaryGame(fmla), self.prover, self.adversary)
return match.play_game(), self.index
def ParallelPlayout(fmla, num_workers, num_matches, prover_type, model_cfg, adversary=RandomAdversary()):
assert num_matches > num_workers
fmla = ray.put(fmla)
worker_pool = [FormulaWorker.remote(prover_type, model_cfg, adversary, i) for i in range(num_workers)]
active_jobs = [worker.play_match.remote(fmla) for worker in worker_pool]
results = []
while len(results) < num_matches - num_workers:
ready_ids, active_jobs = ray.wait(active_jobs)
for x in ready_ids:
result, index = ray.get(x)
results.append(result)
active_jobs.append(worker_pool[index].play_match.remote(fmla))
for job in active_jobs:
result, _ = ray.get(job)
results.append(result)
return np.mean(results, axis=0)
def playout_eval(experiment_name, prover_type, data_dir, model_cfg, n_cpus, n_matches, test): # play 50 matches per formula
ray.init(num_cpus=n_cpus)
cnf_files = files_with_extension(data_dir, "cnf")
step_count = 0
if model_cfg is not None:
name = prover_type + "_" + os.path.basename(os.path.splitext(model_cfg)[0])
else:
name = prover_type
log_dir = os.path.join("playout_eval/", experiment_name, name +"/")
writer = tf.summary.create_file_writer(log_dir)
for f in cnf_files:
fmla = CNF(from_file=f)
avg_unit_props, avg_terminal_value = ParallelPlayout(fmla, n_cpus, n_matches, prover_type, model_cfg)
with writer.as_default():
tf.summary.scalar("avg unit props", avg_unit_props, step=(step_count+1))
tf.summary.scalar("avg terminal value", avg_terminal_value, step=(step_count+1))
print("finished step", step_count)
step_count += 1
def test_playout_eval(experiment_name): # play 100 matches per formula
ray.init(num_cpus=4)
for prover_type in ["random", "drat", "core"]:
for model_cfg in ["res_models/res_grid_2_sr.json", "res_models/res_grid_2_src.json"]:
step_count = 0
if model_cfg is not None:
name = prover_type + "_" + os.path.basename(os.path.splitext(model_cfg)[0])
else:
name = prover_type
log_dir = os.path.join("test_playout_eval/", experiment_name, name +"/")
writer = tf.summary.create_file_writer(log_dir)
for _ in range(3):
fmla = get_unsat_randkcnf(3,40)
avg_unit_props, avg_terminal_value = ParallelPlayout(fmla, 4, 10, prover_type, model_cfg)
with writer.as_default():
tf.summary.scalar("avg unit props", avg_unit_props, step=(step_count+1))
tf.summary.scalar("avg terminal value", avg_terminal_value, step=(step_count+1))
step_count += 1
print(f"{name} ok")
# example usage:
# # python playout_eval.py EXPERIMENT_NAME drat --model=res_models/res_grid_2_sr.json --cpus=16 --matches=100 --data="cnf_data/ramsey/test/"
# # python playout_eval.py EXPERIMENT_NAME core --model=res_models/res_grid_2_src.json --cpus=16 --matches=100 --data="cnf_data/ramsey/test/"
# # python playout_eval.py EXPERIMENT_NAME z3 --cpus=16 --matches=50 --data="cnf_data/ramsey/test/"
# # python playout_eval.py EXPERIMENT_NAME random --cpus=16 --matches=100 --data="cnf_data/ramsey/test/"
# # python playout_eval.py test1 random --cpus=4 --matches=50 --data="cnf_data/ramsey/test/" --test
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("experiment_name")
parser.add_argument("prover_type")
parser.add_argument("--data", action="store", dest="data_dir")
parser.add_argument("--model", action="store", dest="model_cfg", default=None)
parser.add_argument("--cpus", action="store", dest="n_cpus", type=int)
parser.add_argument("--matches", action="store", dest="n_matches", type=int)
parser.add_argument("--test", action="store_true", dest="test")
opts = parser.parse_args()
if opts.test:
test_playout_eval(opts.experiment_name)
else:
playout_eval(**vars(opts))
# for key in vars(opts):
# print(vars(opts)[key])
# test_playout_eval(**vars(opts))
# ramsey_playout_eval(**vars(opts))
# if __name__ == "__main__":
# ray.init()
# start = time.time()
# result0, result1 = ParallelPlayout(get_unsat_randkcnf(3,100), 4, 100, "drat", "res_models/res_grid_2_randkcnf.json")
# # result0, result1 = ParallelPlayout(gen_ramsey_fragment(4,4,18,35), 4, 100, "drat", "res_models/res_grid_0_randkcnf.json")
# print("elapsed", time.time() - start)
# print(result0, result1)