-
Notifications
You must be signed in to change notification settings - Fork 0
/
cadet_env.py
440 lines (391 loc) · 14.6 KB
/
cadet_env.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
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
import ipdb
import select
import gym
import time
from subprocess import Popen, PIPE, STDOUT
from collections import deque
from collections import namedtuple
from gym import spaces
from IPython.core.debugger import Tracer
from settings import *
from qbf_data import *
from cadet_utils import *
from envbase import *
from rl_utils import *
DEF_GREEDY_ALPHA = 0.01
MAX_EPISODE_LENGTH = 200
DQN_DEF_COST = 1e-4
# DQN_DEF_COST = 0.
BINARY_SUCCESS = 1.
LOG_SIZE = 100
def require_init(f, *args, **kwargs):
def inner(instance, *args, **kwargs):
assert(instance.cadet_proc != None)
return f(instance,*args,**kwargs)
return inner
class CadetSpace(gym.Space):
def contains(self, x):
return True
@property
def shape(self):
return ()
@shape.setter
def shape(self, value):
pass
# Cadet actions are 1-based. The CadetEnv exposes 0-based actions
class CadetEnv(EnvBase):
def __init__(self, provider=None, **kwargs):
self.EnvObservation = namedtuple('EnvObservation',
['state', 'vars_add', 'vars_remove', 'activities', 'decision', 'clause',
'reward', 'vars_set', 'done'])
self.settings = CnfSettings()
self.provider = provider
self.cadet_binary = self.settings['cadet_binary']
self.debug = self.settings['debug']
self.qbf = QbfBase(**kwargs)
self.observation_space = CadetSpace()
self.action_space = spaces.Discrete(self.settings['max_variables'])
self.greedy_rewards = self.settings['greedy_rewards']
self.clause_learning = self.settings['clause_learning']
self.vars_set = self.settings['vars_set']
self.fresh_seed = self.settings['fresh_seed']
self.use_old_rewards = self.settings['use_old_rewards']
self.use_vsids_rewards = self.settings['use_vsids_rewards']
self.slim_state = self.settings['slim_state']
self.def_step_cost = self.settings['def_step_cost']
self.cadet_completion_reward = self.settings['cadet_completion_reward']
self.max_step = self.settings['max_step']
self.restart_solver_every = int(self.settings['restart_solver_every'])
# self.logger = logger
self.greedy_alpha = DEF_GREEDY_ALPHA if self.greedy_rewards else 0.
self.tail = deque([],LOG_SIZE)
self.reset_counter = 0
self.use_activities = self.settings['cadet_use_activities']
self.last_processed_obs = None
self.start_cadet()
def start_cadet(self):
cadet_params = ['--rl', '--cegar', '--sat_by_qbf', '--rl_reward_per_decision', '{}'.format(self.def_step_cost),
'--rl_completion_reward', '{}'.format(self.cadet_completion_reward)]
if not self.use_old_rewards:
if self.debug:
print('Using new rewards!')
cadet_params.append('--rl_advanced_rewards')
if self.fresh_seed:
if self.debug:
print('Using fresh seed!')
cadet_params.append('--fresh_seed')
if self.slim_state:
if self.debug:
print('Using slim_state!')
cadet_params.append('--rl_slim_state')
if self.use_vsids_rewards:
if self.debug:
print('Using vsids rewards!!')
cadet_params.append('--rl_vsids_rewards')
if self.debug:
print(' '.join([self.cadet_binary, *cadet_params]))
self.cadet_proc = Popen([self.cadet_binary, *cadet_params], stdout=PIPE, stdin=PIPE, stderr=STDOUT, universal_newlines=True)
self.poll_obj = select.poll()
self.poll_obj.register(self.cadet_proc.stdout, select.POLLIN)
self.finished = False
self.done = True
self.current_fname = None
def stop_cadet(self, timeout):
assert(self.cadet_proc != None)
self.cadet_proc.terminate()
time.sleep(timeout)
if self.cadet_proc.poll() != None:
self.cadet_proc.kill()
self.cadet_proc = None
self.poll_obj = None
def restart_env(self, timeout=5):
if self.debug:
print('Stopping cadet...')
self.stop_cadet(timeout)
if self.debug:
print('Restarting cadet...')
self.start_cadet()
def eat_initial_output(self):
self.read_line_with_timeout()
self.read_line_with_timeout()
def write(self, val):
self.cadet_proc.stdin.write(val)
self.cadet_proc.stdin.flush()
def terminate(self):
if not self.done:
if self.debug:
print('interrupting mid-episode!')
self.write_action(-1)
a = self.read_line_with_timeout()
if self.debug:
print(a)
self.done = True
rewards = np.asarray(list(map(float,a.split()[1:])))
return rewards
def exit(self):
print('exit()')
self.terminate()
self.stop_cadet(5)
def reset(self):
self.reset_counter += 1
if self.restart_solver_every > 0 and (self.settings['restart_in_test'] or (self.reset_counter % self.restart_solver_every == 0)):
self.restart_env(timeout=0)
self.terminate()
fname = self.provider.get_next()
if self.debug:
print('Starting Env {}'.format(fname))
# if fname == 'data/huge_gen1/small-bug1-fixpoint-3.qdimacs':
# Tracer()()
self.qbf.reload_qdimacs(fname) # This holds our own representation of the qbf graph
self.vars_deterministic = np.zeros(self.qbf.num_vars)
self.total_vars_deterministic = np.zeros(self.qbf.num_vars)
self.activities = np.zeros(self.qbf.num_vars)
self.max_rewards = self.qbf.num_existential
self.timestep = 0
self.finished = False
self.running_reward = []
self.rewards = None
self.write(fname+'\n')
self.done = False
self.current_fname = fname
rc = self.read_state_update() # Initial state
self.last_processed_obs = self.process_observation(None, rc)
return self.last_processed_obs
def read_line_with_timeout(self, timeout=10.):
return self.cadet_proc.stdout.readline()
# entry = time.time()
# curr = entry
# line = ''
# while curr < entry + timeout:
# p = self.poll_obj.poll(0)
# if p:
# c = self.cadet_proc.stdout.read(1)
# if c == '\n':
# # Tracer()()
# return line
# line += c
# else:
# # print('Poll negative..')
# pass
# curr = time.time()
# return None
# This is where we go from 0-based to 1-based
def write_action(self, a):
if self.debug:
print('Writing action {}'.format(a))
if a in ['?','r']:
self.write('{}\n'.format(a))
return
if type(a) is tuple:
cadet_action = int(a[0]) + 1
if a[1]:
cadet_action = -cadet_action
else:
cadet_action = int(a)+1
self.write('%d\n' % cadet_action)
'''
Returns:
state - A bunch of numbers describing general solver state
candidates - an array of (0-based) available actions in the current state
done - Is this the end of the episode?
'''
def read_state_update(self):
self.vars_deterministic.fill(0)
self.activities.fill(0)
clause = None
reward = None
decision = None
vars_set = []
while True:
a = self.read_line_with_timeout()
if not a or a == '\n': continue
self.tail.append(a)
if self.debug:
print(a)
if False and a == 'UNSAT\n':
if self.cadet_binary != './cadet':
a = self.read_line_with_timeout() # refutation line
a = self.read_line_with_timeout() # rewards
self.rewards = np.asarray(list(map(float,a.split()[1:])))
if np.isnan(self.rewards).any():
if np.isnan(self.rewards[:-1]).any():
Tracer()()
else:
self.rewards[-1]=BINARY_SUCCESS
self.done = True
state = None
break
elif False and a == 'SAT\n':
Tracer()()
a = self.read_line_with_timeout() # rewards
self.rewards = np.asarray(list(map(float,a.split()[1:])))
if np.isnan(self.rewards).any():
if np.isnan(self.rewards[:-1]).any():
Tracer()()
else:
self.rewards[-1]=BINARY_SUCCESS
self.done = True
state = None
break
elif a.startswith('rewards') or a.startswith('SATrewards') or a.startswith('UNSATrewards'):
self.rewards = np.asarray(list(map(float,a.split()[1:])))
# if self.debug:
# Tracer()()
if np.isnan(self.rewards).any():
if np.isnan(self.rewards[:-1]).any():
Tracer()()
else:
self.rewards[-1]=BINARY_SUCCESS
self.done = True
self.finished = True
state = None
if self.debug:
print('Successfuly finished episode in {} steps!'.format(self.timestep))
break
elif a[0] == 'u' and a[1] != 'c': # New cadet has 'uc'
update = int(a[3:])-1 # Here we go from 1-based to 0-based
if a[1] == '+':
self.vars_deterministic[update] = 1
self.total_vars_deterministic[update] = 1
elif a[1] == '-':
self.vars_deterministic[update] = -1
self.total_vars_deterministic[update] = 0
elif a.startswith('delete_clause'):
if not self.clause_learning:
continue
cid = int(a.split()[1])
# print('Removing clause id {}'.format(cid))
# print(a)
self.qbf.remove_clause(cid)
clause = True
elif a[0] == 'd':
decision = [int(x) for x in a[2:].split(',')]
decision[0] -= 1
elif a[0] == 's':
state = np.array([float(x) for x in a[2:].split(',')])
if self.debug:
print('Got state!')
break
elif a[0] == 'a':
b = a[2:].split(',')
update = int(b[0])-1
activity = float(b[1])
self.activities[update] = activity
elif a[0] == 'v' and self.vars_set:
v, pol = a.split(' ')[1:]
vars_set.append((int(v)-1,int(pol)))
elif self.timestep > 0 and a[0] == 'c':
if a.startswith('conflict'):
continue
elif a.startswith('clause'): # new cadet version
if not self.clause_learning:
continue
c = a.split()
cid = int(c[1])
b = [int(x) for x in c[4:]]
self.qbf.add_clause(b,cid)
# print('Adding clause id {}'.format(cid))
# print(a)
else:
print('This version is too old')
Tracer()()
b = [int(x) for x in a[2:].split()]
# clause = (np.array([abs(x)-1 for x in b if x > 0]), np.array([abs(x)-1 for x in b if x < 0]))
clause = True
elif self.debug:
print('Got unprocessed line: %s' % a)
if a.startswith('Error'):
return
if self.timestep > 0:
greedy_reward = np.count_nonzero(self.total_vars_deterministic) - self.last_total_determinized
self.running_reward.append(greedy_reward)
reward = self.def_step_cost + (self.cadet_completion_reward if self.done else 0.)
if self.greedy_rewards:
reward += self.greedy_alpha*self.running_reward[-1]
self.last_total_determinized = np.count_nonzero(self.total_vars_deterministic)
# if self.timestep > 200:
# self.debug = True
if sum(self.running_reward) < -1:
Tracer()()
if self.done:
self.rewards = self.rewards + self.greedy_alpha*np.asarray(self.running_reward)
pos_vars = np.where(self.vars_deterministic>0)[0]
neg_vars = np.where(self.vars_deterministic<0)[0]
return self.EnvObservation(state, pos_vars, neg_vars, self.activities, decision, clause, reward, np.array(vars_set), self.done)
# This gets a tuple from stepping the environment:
# state, vars_add, vars_remove, activities, decision, clause, reward, done = env.step(action)
# And it returns the next observation.
def process_observation(self, last_obs, env_obs):
# import ipdb
# ipdb.set_trace()
if not env_obs or env_obs.done:
return EmptyDenseState
if env_obs.clause or not last_obs:
# Tracer()()
cmat = get_input_from_qbf(self.qbf, self.settings, False) # Do not split
clabels = Variable(torch.from_numpy(self.qbf.get_clabels()).float().unsqueeze(0)).t()
else:
last_obs = undensify_obs(last_obs)
cmat, clabels = last_obs.cmat, last_obs.clabels
if last_obs:
ground_embs = np.copy(last_obs.ground.data.numpy().squeeze())
vmask = last_obs.vmask
cmask = last_obs.cmask
else:
ground_embs = self.qbf.get_base_embeddings()
vmask = None
cmask = None
if env_obs.decision:
ground_embs[env_obs.decision[0]][IDX_VAR_POLARITY_POS+1-env_obs.decision[1]] = True
if len(env_obs.vars_add):
ground_embs[:,IDX_VAR_DETERMINIZED][env_obs.vars_add] = True
if len(env_obs.vars_remove):
ground_embs[:,IDX_VAR_DETERMINIZED][env_obs.vars_remove] = False
ground_embs[:,IDX_VAR_POLARITY_POS:IDX_VAR_POLARITY_NEG][env_obs.vars_remove] = False
if self.use_activities:
ground_embs[:,IDX_VAR_ACTIVITY] = env_obs.activities
if len(env_obs.vars_set):
a = env_obs.vars_set
idx = a[:,0][np.where(a[:,1]==1)[0]]
ground_embs[:,IDX_VAR_SET_POS][idx] = True
idx = a[:,0][np.where(a[:,1]==-1)[0]]
ground_embs[:,IDX_VAR_SET_NEG][idx] = True
idx = a[:,0][np.where(a[:,1]==0)[0]]
ground_embs[:,IDX_VAR_SET_POS:IDX_VAR_SET_NEG][idx] = False
state = Variable(torch.from_numpy(env_obs.state).float().unsqueeze(0))
ground_embs = Variable(torch.from_numpy(ground_embs).float().unsqueeze(0))
# ipdb.set_trace()
return densify_obs(State(state,cmat, ground_embs, clabels, vmask, cmask, None))
# This returns already a State (higher-level) observation, not EnvObs.
def new_episode(self, fname, settings=None, **kwargs):
try:
obs = self.reset(fname)
# state, vars_add, vars_remove, activities, _, _ , _, vars_set, _ = self.reset(fname)
if obs.state is not None: # Env solved in 0 steps
return obs
except:
print('Error reseting with file {}'.format(fname))
def check_break(self):
return self.timestep >= self.max_step
def translate_action(self, action):
try:
if action in ['?']:
return action
except:
pass
return (int(action/2),int(action%2))
def step(self, action):
assert(not self.done)
# if self.greedy_rewards and self.timestep > MAX_EPISODE_LENGTH:
# rewards = self.terminate() + self.greedy_alpha*np.asarray(self.running_reward)
# self.rewards = np.concatenate([rewards, [DQN_DEF_COST]]) # Average action
# return None, None, None, None, None, None, DQN_DEF_COST, None, True
self.timestep += 1
# print('Got action {}'.format(action))
# ipdb.set_trace()
self.write_action(self.translate_action(action))
env_obs = self.read_state_update()
obs = self.process_observation(self.last_processed_obs, env_obs)
self.last_processed_obs = obs
rc = obs, env_obs.reward, env_obs.done or self.check_break(), {'fname': self.current_fname}
return rc