add cyclic groups; fix random seeds (should use self.np_rng, otherwise the seed is not fixed)
Browse files- automata.py +102 -48
automata.py
CHANGED
@@ -103,60 +103,69 @@ class SyntheticAutomataDataset(datasets.GeneratorBasedBuilder):
|
|
103 |
|
104 |
|
105 |
class AutomatonSampler:
|
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 |
print(self.__info__)
|
141 |
|
142 |
|
|
|
|
|
143 |
class BinaryInputSampler(AutomatonSampler):
|
144 |
-
|
145 |
-
|
|
|
|
|
|
|
|
|
146 |
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
|
153 |
-
|
154 |
-
|
155 |
|
156 |
-
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
|
161 |
class ParitySampler(BinaryInputSampler):
|
162 |
def __init__(self, data_config):
|
@@ -220,6 +229,7 @@ class GridworldSampler(BinaryInputSampler):
|
|
220 |
states = states[1:] # remove the 1st entry with is the (meaningless) initial value 0
|
221 |
return np.array(states).astype(np.int64)
|
222 |
|
|
|
223 |
class ABABSampler(BinaryInputSampler):
|
224 |
def __init__(self, data_config):
|
225 |
super().__init__(data_config)
|
@@ -265,7 +275,7 @@ class ABABSampler(BinaryInputSampler):
|
|
265 |
return labels
|
266 |
|
267 |
def sample(self):
|
268 |
-
pos_sample =
|
269 |
if pos_sample:
|
270 |
T = self.sample_length()
|
271 |
x = [0,1,0,1] * (T//4)
|
@@ -299,23 +309,25 @@ class FlipFlopSampler(AutomatonSampler):
|
|
299 |
|
300 |
def f(self, x):
|
301 |
state, states = 0, []
|
302 |
-
for
|
303 |
-
state = self.transition[state,
|
304 |
states += state,
|
305 |
return np.array(states)
|
306 |
|
307 |
def sample(self):
|
308 |
T = self.sample_length()
|
309 |
-
rand =
|
310 |
nonzero_pos = (rand < 0.5).astype(np.int64)
|
311 |
-
writes =
|
312 |
x = writes * nonzero_pos
|
313 |
return x, self.f(x)
|
314 |
|
315 |
|
316 |
|
|
|
317 |
class PermutationSampler(AutomatonSampler):
|
318 |
"""
|
|
|
319 |
Subclasses: SymmetricSampler, AlternatingSampler
|
320 |
"""
|
321 |
def __init__(self, data_config):
|
@@ -344,8 +356,8 @@ class PermutationSampler(AutomatonSampler):
|
|
344 |
def f(self, x):
|
345 |
curr_state = np.arange(self.n)
|
346 |
labels = []
|
347 |
-
for
|
348 |
-
curr_state = self.actions[
|
349 |
|
350 |
if self.label_type == 'state':
|
351 |
labels += self.get_state_label(curr_state),
|
@@ -356,7 +368,7 @@ class PermutationSampler(AutomatonSampler):
|
|
356 |
|
357 |
def sample(self):
|
358 |
T = self.sample_length()
|
359 |
-
x =
|
360 |
|
361 |
return x, self.f(x)
|
362 |
|
@@ -462,13 +474,55 @@ class AlternatingSampler(PermutationSampler):
|
|
462 |
|
463 |
|
464 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
465 |
|
466 |
dataset_map = {
|
467 |
'abab': ABABSampler,
|
468 |
'alternating': AlternatingSampler,
|
469 |
-
'
|
470 |
'flipflop': FlipFlopSampler,
|
|
|
471 |
'parity': ParitySampler,
|
472 |
'symmetric': SymmetricSampler,
|
473 |
# TODO: more datasets
|
474 |
}
|
|
|
|
103 |
|
104 |
|
105 |
class AutomatonSampler:
|
106 |
+
"""
|
107 |
+
This is a parent class that must be inherited.
|
108 |
+
"""
|
109 |
+
def __init__(self, data_config):
|
110 |
+
self.data_config = data_config
|
111 |
|
112 |
+
if 'seed' in self.data_config:
|
113 |
+
self.np_rng = np.random.default_rng(self.data_config['seed'])
|
114 |
+
else:
|
115 |
+
self.np_rng = np.random.default_rng()
|
116 |
|
117 |
+
if 'length' not in data_config: # sequence length
|
118 |
+
data_config['length'] = 20
|
119 |
+
self.T = self.data_config['length']
|
120 |
|
121 |
+
if 'random_length' not in data_config:
|
122 |
+
data_config['random_length'] = 0
|
123 |
+
self.random_length = data_config['random_length']
|
124 |
|
125 |
+
self.__info__ = " - T (int): sequence length.\n" \
|
126 |
+
+ " - random_length (int in {0, 1}): whether to randomly sample a length per sample.\n"
|
127 |
|
128 |
+
def f(self, x):
|
129 |
+
"""
|
130 |
+
Get output sequence given an input seq
|
131 |
+
"""
|
132 |
+
raise NotImplementedError()
|
133 |
|
134 |
+
def sample(self):
|
135 |
+
raise NotImplementedError()
|
136 |
|
137 |
+
def sample_length(self):
|
138 |
+
if self.random_length:
|
139 |
+
return self.np_rng.choice(range(1, self.T+1))
|
140 |
+
return self.T
|
141 |
|
142 |
+
def help(self):
|
143 |
print(self.__info__)
|
144 |
|
145 |
|
146 |
+
|
147 |
+
|
148 |
class BinaryInputSampler(AutomatonSampler):
|
149 |
+
"""
|
150 |
+
This is a parent class that must be inherited.
|
151 |
+
Subclasses: ParitySampler, GridworldSampler, ABABSampler
|
152 |
+
"""
|
153 |
+
def __init__(self, data_config):
|
154 |
+
super().__init__(data_config)
|
155 |
|
156 |
+
if 'prob1' not in data_config:
|
157 |
+
data_config['prob1'] = 0.5
|
158 |
+
self.prob1 = data_config['prob1']
|
159 |
+
self.__info__ = " - prob1 (float in [0,1]): probability of token 1\n" \
|
160 |
+
+ self.__info__
|
161 |
|
162 |
+
def f(self, x):
|
163 |
+
raise NotImplementedError()
|
164 |
|
165 |
+
def sample(self):
|
166 |
+
T = self.sample_length()
|
167 |
+
x = self.np_rng.binomial(1, self.prob1, size=T)
|
168 |
+
return x, self.f(x)
|
169 |
|
170 |
class ParitySampler(BinaryInputSampler):
|
171 |
def __init__(self, data_config):
|
|
|
229 |
states = states[1:] # remove the 1st entry with is the (meaningless) initial value 0
|
230 |
return np.array(states).astype(np.int64)
|
231 |
|
232 |
+
|
233 |
class ABABSampler(BinaryInputSampler):
|
234 |
def __init__(self, data_config):
|
235 |
super().__init__(data_config)
|
|
|
275 |
return labels
|
276 |
|
277 |
def sample(self):
|
278 |
+
pos_sample = self.np_rng.random() < self.prob_abab_pos_sample
|
279 |
if pos_sample:
|
280 |
T = self.sample_length()
|
281 |
x = [0,1,0,1] * (T//4)
|
|
|
309 |
|
310 |
def f(self, x):
|
311 |
state, states = 0, []
|
312 |
+
for action_id in x:
|
313 |
+
state = self.transition[state, action_id]
|
314 |
states += state,
|
315 |
return np.array(states)
|
316 |
|
317 |
def sample(self):
|
318 |
T = self.sample_length()
|
319 |
+
rand = self.np_rng.uniform(size=T)
|
320 |
nonzero_pos = (rand < 0.5).astype(np.int64)
|
321 |
+
writes = self.np_rng.choice(range(1, self.n_states+1), size=T)
|
322 |
x = writes * nonzero_pos
|
323 |
return x, self.f(x)
|
324 |
|
325 |
|
326 |
|
327 |
+
|
328 |
class PermutationSampler(AutomatonSampler):
|
329 |
"""
|
330 |
+
This is a parent class that must be inherited.
|
331 |
Subclasses: SymmetricSampler, AlternatingSampler
|
332 |
"""
|
333 |
def __init__(self, data_config):
|
|
|
356 |
def f(self, x):
|
357 |
curr_state = np.arange(self.n)
|
358 |
labels = []
|
359 |
+
for action_id in x:
|
360 |
+
curr_state = self.actions[action_id].dot(curr_state)
|
361 |
|
362 |
if self.label_type == 'state':
|
363 |
labels += self.get_state_label(curr_state),
|
|
|
368 |
|
369 |
def sample(self):
|
370 |
T = self.sample_length()
|
371 |
+
x = self.np_rng.choice(range(self.n_actions), replace=True, size=T)
|
372 |
|
373 |
return x, self.f(x)
|
374 |
|
|
|
474 |
|
475 |
|
476 |
|
477 |
+
class CyclicSampler(AutomatonSampler):
|
478 |
+
def __init__(self, data_config):
|
479 |
+
super().__init__(data_config)
|
480 |
+
|
481 |
+
if 'n' not in data_config:
|
482 |
+
data_config['n'] = 5
|
483 |
+
self.n = data_config['n']
|
484 |
+
|
485 |
+
"""
|
486 |
+
Get actions: shift by i positions, for i = 0 to n_actions-1
|
487 |
+
"""
|
488 |
+
if 'n_actions' not in data_config:
|
489 |
+
data_config['n_actions'] = 2
|
490 |
+
self.n_actions = data_config['n_actions']
|
491 |
+
shift_idx = list(range(1, self.n)) + [0]
|
492 |
+
self.actions = {}
|
493 |
+
for i in range(self.n_actions):
|
494 |
+
shift_idx = list(range(i, self.n)) + list(range(0, i))
|
495 |
+
self.actions[i] = np.eye(self.n)[shift_idx]
|
496 |
+
|
497 |
+
|
498 |
+
def f(self, x):
|
499 |
+
if OLD_PY_VERSION:
|
500 |
+
# NOTE: for Python 3.7 or below, accumulate doesn't have the 'initial' argument.
|
501 |
+
x_padded = np.concatenate([np.array([0]), x]).astype(np.int64)
|
502 |
+
states = list(itertools.accumulate(x_padded, lambda a,b: (a+b)%self.n ))
|
503 |
+
states = states[1:]
|
504 |
+
else:
|
505 |
+
states = list(itertools.accumulate(x, lambda a,b: (a+b)%self.n, initial=0))
|
506 |
+
states = states[1:] # remove the 1st entry with is the (meaningless) initial value 0
|
507 |
+
|
508 |
+
return np.array(states).astype(np.int64)
|
509 |
+
|
510 |
+
def sample(self):
|
511 |
+
T = self.sample_length()
|
512 |
+
x = self.np_rng.choice(range(self.n_actions), replace=True, size=T)
|
513 |
+
|
514 |
+
return x, self.f(x)
|
515 |
+
|
516 |
+
|
517 |
|
518 |
dataset_map = {
|
519 |
'abab': ABABSampler,
|
520 |
'alternating': AlternatingSampler,
|
521 |
+
'cyclic': CyclicSampler,
|
522 |
'flipflop': FlipFlopSampler,
|
523 |
+
'gridworld': GridworldSampler,
|
524 |
'parity': ParitySampler,
|
525 |
'symmetric': SymmetricSampler,
|
526 |
# TODO: more datasets
|
527 |
}
|
528 |
+
|