cyrilzhang commited on
Commit
b0d21d5
1 Parent(s): dbf2014

unify indentation

Browse files
Files changed (1) hide show
  1. automata.py +572 -580
automata.py CHANGED
@@ -109,660 +109,652 @@ class AutomatonDataset(datasets.GeneratorBasedBuilder):
109
  }
110
 
111
  class Automaton:
112
- """
113
- This is a parent class that must be inherited.
114
- """
115
- def __init__(self, data_config):
116
- self.data_config = data_config
117
-
118
- if 'seed' in self.data_config:
119
- self.np_rng = np.random.default_rng(self.data_config['seed'])
120
- else:
121
- self.np_rng = np.random.default_rng()
122
-
123
- if 'length' not in data_config: # sequence length
124
- data_config['length'] = 20
125
- self.T = self.data_config['length']
126
-
127
- if 'random_length' not in data_config:
128
- data_config['random_length'] = 0
129
- self.random_length = data_config['random_length']
130
-
131
- self.__info__ = \
132
- " - T (int): sequence length.\n" \
133
- + " - random_length (int in {0, 1}): whether to randomly sample a length per sample.\n"
134
-
135
- def f(self, x):
136
- """
137
- Get output sequence given an input seq
138
- """
139
- raise NotImplementedError()
140
-
141
- def sample(self):
142
- raise NotImplementedError()
143
-
144
- def sample_length(self):
145
- if self.random_length:
146
- return self.np_rng.choice(range(1, self.T+1))
147
- return self.T
148
-
149
- def help(self):
150
- print(self.__info__)
151
 
152
- class BinaryInputAutomaton(Automaton):
153
- """
154
- This is a parent class that must be inherited.
155
- Subclasses: ParityAutomaton, GridworldAutomaton, ABABAutomaton
156
 
157
- TODO: sample sequences with a given number of 1s
158
- """
159
- def __init__(self, data_config):
160
- super().__init__(data_config)
161
 
162
- if 'prob1' not in data_config:
163
- data_config['prob1'] = 0.5
164
- self.prob1 = data_config['prob1']
165
- self.__info__ = " - prob1 (float in [0,1]): probability of token 1\n" \
166
- + self.__info__
167
 
168
- def f(self, x):
169
- raise NotImplementedError()
 
170
 
171
- def sample(self):
172
- T = self.sample_length()
173
- x = self.np_rng.binomial(1, self.prob1, size=T)
174
- return x, self.f(x)
 
175
 
176
- class ParityAutomaton(BinaryInputAutomaton):
177
- def __init__(self, data_config):
178
- super().__init__(data_config)
179
- self.name = 'parity'
180
 
181
- self.__info__ = "Parity machine with 2 states: \n" \
182
- + "- Inputs: binary strings\n" \
183
- + "- Labels: binary strings of the partial parity\n" \
184
- + "- Config: \n" \
185
- + self.__info__
186
 
187
- def f(self, x):
188
- return np.cumsum(x) % 2
189
 
190
- class GridworldAutomaton(BinaryInputAutomaton):
191
- """
192
- Note: gridworld currently doesn't include a no-op.
193
- """
194
- def __init__(self, data_config):
195
- super().__init__(data_config)
196
-
197
- if 'n' not in data_config:
198
- data_config['n'] = 9
199
- """
200
- NOTE: n is the number of states, and S is the id (0-indexing) of the rightmost state.
201
- i.e. the states are 0,1,2,...,S, where S=n-1.
202
  """
203
- self.n = data_config['n']
204
- self.S = self.n - 1
205
-
206
- if 'label_type' not in data_config:
207
- # Options: state, parity, boundary
208
- data_config['label_type'] = 'state'
209
- self.label_type = data_config['label_type']
210
-
211
- self.name = f'Grid{self.n}'
212
-
213
- self.__info__ = f"1d Gridworld of n={self.n} states:\n" \
214
- + "- Inputs: binary strings, i.e. move left(0) or right(1)\n" \
215
- + "- Labels: depending on 'label_type'. \n" \
216
- + "- Config: \n" \
217
- + " - n (int): number of states; i.e. the states are 0,1,2,...,n-1.\n" \
218
- + " - label_type (str): choosing from the following options:\n" \
219
- + " - 'state' (default): the state id, i.e. 0 to n-1.\n" \
220
- + " - 'parity': the state id mod 2.\n" \
221
- + " - 'boundary': whether the current state is in {0, n-1} or not.\n" \
222
- + self.__info__
223
-
224
-
225
- def f(self, x):
226
- x = copy(x)
227
- x[x == 0] = -1
228
- if OLD_PY_VERSION:
229
- # NOTE: for Python 3.7 or below, accumulate doesn't have the 'initial' argument.
230
- x = np.concatenate([np.array([0]), x]).astype(np.int64)
231
- states = list(itertools.accumulate(x, lambda a,b: max(min(a+b, self.S), 0)))
232
- states = states[1:]
233
- else:
234
- states = list(itertools.accumulate(x, lambda a,b: max(min(a+b, self.S), 0), initial=0))
235
- states = states[1:] # remove the 1st entry with is the (meaningless) initial value 0
236
- return np.array(states).astype(np.int64)
237
 
 
 
 
 
238
 
239
- class ABABAutomaton(BinaryInputAutomaton):
240
- def __init__(self, data_config):
241
- super().__init__(data_config)
242
- self.name = 'abab'
243
-
244
- if 'prob_abab_pos_sample' not in data_config:
245
- # The probability of having a positive sequence, i.e. 010101010101...
246
- data_config['prob_abab_pos_sample'] = 0.25
247
- if 'label_type' not in data_config:
248
- # Options: 'state', 'boundary'
249
- data_config['label_type'] = 'state'
250
-
251
- self.prob_abab_pos_sample = data_config['prob_abab_pos_sample']
252
- self.label_type = data_config['label_type']
253
-
254
- self.transition = np.array(
255
- [[4, 1], # state 0
256
- [2, 4], # state 1
257
- [4, 3], # state 2
258
- [0, 4], # state 3
259
- [4, 4], # state 4
260
- ])
261
-
262
- self.__info__ = "abab: an automaton with 4 states + 1 absorbing state:\n" \
263
- + "- Inputs: binary strings\n" \
264
- + "- Labels: depending on 'label_type'.\n" \
265
- + "- Config:\n" \
266
- + " - prob_abab_pos_sample (float in [0,1]): probability of having a 'positive' sequence, i.e. 01010101010...\n" \
267
- + " - label_type (str): choosing from the following options:\n" \
268
- + " - 'state' (default): the state id.\n" \
269
- + " - 'boundary': whether the state is in state 3 (the states are 0,1,2,3).\n" \
270
- + self.__info__
271
-
272
- def f(self, x):
273
- labels = []
274
- curr_state = 3
275
- for each in x:
276
- curr_state = self.transition[curr_state, each]
277
- labels += curr_state,
278
- labels = np.array(labels).astype(np.int64)
279
- if self.label_type == 'boundary':
280
- labels = (labels == 3).astype(np.int64)
281
- return labels
282
-
283
- def sample(self):
284
- pos_sample = self.np_rng.random() < self.prob_abab_pos_sample
285
- if pos_sample:
286
- T = self.sample_length()
287
- x = [0,1,0,1] * (T//4)
288
- x += [0,1,0,1][:(T%4)]
289
- x = np.array(x)
290
- return x, self.f(x)
291
- else:
292
- return super().sample()
293
 
 
 
294
 
295
- class AdderAutomaton(BinaryInputAutomaton):
296
- def __init__(self, data_config):
297
- super().__init__(data_config)
298
- self.name = 'addition'
299
-
300
- if 'n_addends' not in data_config:
301
- data_config['n_addends'] = 2
302
- self.n_addends = data_config['n_addends']
303
- self.addend_scales = np.array([2**i for i in range(self.n_addends)]).reshape(-1, 1)
304
-
305
- if 'label_type' not in data_config:
306
- data_config['label_type'] = 'state'
307
- self.label_type = data_config['label_type']
308
-
309
- self.__info__ = f'Adder of n={self.n_addends} binary numbers:\n' \
310
- +f"- Inputs: {self.n_addends} binary numbers, encoded as the int for the {self.n_addends}-bit binary number.\n" \
311
- + "- Labels: depending on the label_type.\n" \
312
- + "- Config:\n" \
313
- + " - n_addends (int): number of binary numbers to be added; default as 2.\n" \
314
- + " - label_type (str): choosing from the following options: \n" \
315
- +f" - 'state': the state id, i.e. the int for the base-{self.n_addends} int corresponding to the number (carry, digit). \n" \
316
- +f" - 'digit': the current output base-{self.n_addends} digit, without the carry. \n" \
317
- + " - 'position': the current carry bit.\n" \
318
- + self.__info__
319
-
320
-
321
- def f(self, x):
322
- outputs, carries = [], []
323
- carry = 0
324
- T = x.shape[-1]
325
- for i in range(T):
326
- curr_sum = x[:, i].sum() + carry
327
- # NOTE: 'mod n_addends' makes sure the carry is binary
328
- output, carry = curr_sum % self.n_addends, curr_sum // self.n_addends
329
- outputs += output,
330
- carries += carry,
331
- outputs = np.array(outputs).astype(np.int64)
332
- carries = np.array(carries).astype(np.int64)
333
-
334
- if self.label_type == 'state':
335
- return outputs + self.n_addends*carries
336
- elif self.label_type == 'digit':
337
- return outputs
338
- elif self.label_type == 'carry':
339
- return carries
340
-
341
- def sample_addend(self, T):
342
- a = self.np_rng.binomial(1, self.prob1, size=T)
343
- return a
344
-
345
- def sample(self):
346
- T = self.sample_length()
347
- x = np.stack([self.sample_addend(T) for _ in range(self.n_addends)])
348
- # Pad the most significant bit (rightmost position, i.e. we're reversing the number) with 0 to handle the potential carry
349
- pad = np.zeros((self.n_addends, 1))
350
- x = np.concatenate([x, pad], 1)
351
-
352
- x_encode = (self.addend_scales * x).sum(0)
353
- return x_encode, self.f(x)
354
 
 
 
 
 
355
 
356
- class FlipFlopAutomaton(Automaton):
357
- def __init__(self, data_config):
358
- super().__init__(data_config)
359
- self.name = 'flipflop'
360
-
361
- if 'n' not in data_config:
362
- data_config['n'] = 2
363
-
364
- self.n_states = data_config['n']
365
- self.n_actions = self.n_states + 1
366
- self.transition = np.array([list(range(self.n_actions))] + [[i+1]*self.n_actions for i in range(self.n_states)]).T
367
 
368
- self.__info__ = f"Flipflop with n={self.n_states} states:\n" \
369
- +f"- Inputs: tokens are either 0 (read) or 1:{self.n} (write).\n" \
370
- + "- Labels: the state id.\n" \
371
- + "- Config:\n" \
372
- + " - n (int): number of write states; i.e. the states are 1,2,...,n, plus a default start state 0.\n" \
373
- + self.__info__
374
 
375
- def f(self, x):
376
- state, states = 0, []
377
- for action_id in x:
378
- state = self.transition[state, action_id]
379
- states += state,
380
- return np.array(states)
381
 
382
- def sample(self):
383
- T = self.sample_length()
384
- rand = self.np_rng.uniform(size=T)
385
- nonzero_pos = (rand < 0.5).astype(np.int64)
386
- writes = self.np_rng.choice(range(1, self.n_states+1), size=T)
387
- x = writes * nonzero_pos
388
- return x, self.f(x)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
389
 
390
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
391
 
392
 
393
- class PermutationAutomaton(Automaton):
394
- """
395
- This is a parent class that must be inherited.
396
- Subclasses: SymmetricAutomaton, AlternatingAutomaton
397
- """
398
- def __init__(self, data_config):
399
- super().__init__(data_config)
400
-
401
- if 'n' not in data_config:
402
- data_config['n'] = 5
403
- if 'label_type' not in data_config:
404
- # Options: 'state', 'first_chair'
405
- data_config['label_type'] = 'state'
406
-
407
- self.n = data_config['n'] # the symmetric group Sn
408
- self.label_type = data_config['label_type']
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
409
 
410
- self.__info__ = \
411
- " - label_type (str): choosing from the following options:\n" \
412
- + " - 'state' (default): the state id.\n" \
413
- + " - 'first_chair': the element in the first position of the permutation.\n" \
414
- + " e.g. if the current permutation is [2,1,4,3], then 'first_chair' is 2.\n" \
415
- + self.__info__
416
 
417
- def get_state_label(self, state):
418
- enc = self.state_encode(state)
419
- return self.state_label_map[enc]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
420
 
421
- def f(self, x):
422
- curr_state = np.arange(self.n)
423
- labels = []
424
- for action_id in x:
425
- curr_state = self.actions[action_id].dot(curr_state)
426
 
427
- if self.label_type == 'state':
428
- labels += self.get_state_label(curr_state),
429
- elif self.label_type == 'first_chair':
430
- labels += curr_state[0],
 
 
 
 
 
 
 
 
 
 
 
 
431
 
432
- return np.array(labels)
 
 
 
 
 
433
 
434
- def sample(self):
435
- T = self.sample_length()
436
- x = self.np_rng.choice(range(self.n_actions), replace=True, size=T)
437
 
438
- return x, self.f(x)
 
 
 
 
439
 
 
 
 
 
 
440
 
441
- class SymmetricAutomaton(PermutationAutomaton):
442
- """
443
- TODO: add options for labels as functions of states
444
- - parity (whether a state is even): this may need packages (e.g. Permutation from sympy)
445
- - position / toggle: for S3 ~ D6, we can add labels for substructures as in Dihedral groups.
446
- """
447
- def __init__(self, data_config):
448
- super().__init__(data_config)
449
 
450
- self.name = f'S{self.n}'
451
 
452
- """
453
- Get states
454
- """
455
- self.state_encode = lambda state: ''.join([str(int(each)) for each in state])
456
- self.state_label_map = {}
457
- for si, state in enumerate(itertools.permutations(range(self.n))):
458
- enc = self.state_encode(state)
459
- self.state_label_map[enc] = si
460
 
 
461
  """
462
- Get actions (3 defaults: id, shift-by-1, swap-first-two)
 
 
463
  """
464
- if 'n_actions' not in data_config:
465
- data_config['n_actions'] = 3
466
- self.n_actions = data_config['n_actions']
467
- self.actions = {0: np.eye(self.n)}
468
- # shift all elements to the right by 1
469
- shift_idx = list(range(1, self.n)) + [0]
470
- self.actions[1] = np.eye(self.n)[shift_idx]
471
- # swap the first 2 elements
472
- shift_idx = [1, 0] + list(range(2, self.n))
473
- self.actions[2] = np.eye(self.n)[shift_idx]
474
-
475
- if self.n_actions > 3:
476
- # add permutations in the order given by itertools.permutations
477
- self.all_permutations = list(itertools.permutations(range(self.n)))[1:]
478
- cnt = 2
479
- for each in self.all_permutations:
480
- action = np.eye(self.n)[list(each)]
481
- if np.linalg.norm(action - self.actions[0]) == 0:
482
- continue
483
- elif np.linalg.norm(action - self.actions[1]) == 0:
484
- continue
485
- self.actions[cnt] = action
486
- cnt += 1
487
- if cnt == self.n_actions: break
488
-
489
- self.__info__ = f"Symmetric group on n={self.n} objects:\n" \
490
- +f"- Inputs: tokens are either 0 (no-op), or 1:{self.n_actions} (corresponding to {self.n_actions} permutations).\n" \
491
- + "- Labels: depending on 'label_type'.\n" \
492
- + "- Config:\n" \
493
- + " - n (int): number of objects, i.e. there are n! states.\n" \
494
- + " - n_actions (int): number of permutations to include in the generator set;\n" \
495
- + " the ordering is given by itertools.permutations, and the first 'n_actions' permutations will be included.\n" \
496
- + self.__info__
497
 
 
498
 
499
- class AlternatingAutomaton(PermutationAutomaton):
500
- """
501
- TODO: other choices of generators (currently using (12x))?
502
- """
503
- def __init__(self, data_config):
504
- super().__init__(data_config)
 
 
505
 
506
- self.name = f'A{self.n}'
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
507
 
508
- """
509
- Get states
510
- """
511
- self.state_label_map = {}
512
- self.state_encode = lambda state: ''.join([str(int(each)) for each in state])
513
- cnt = 0
514
- for si, state in enumerate(itertools.permutations(range(self.n))):
515
- if not Permutation(state).is_even:
516
- continue
517
- enc = self.state_encode(state)
518
- self.state_label_map[enc] = cnt
519
- cnt += 1
520
 
 
521
  """
522
- Get actions: all 3 cycles of the form (12x)
523
  """
524
- self.actions = {0: np.eye(self.n)}
525
- for idx in range(2, self.n):
526
- # (1, 2, idx)
527
- shift_idx = list(range(self.n))
528
- shift_idx[0],shift_idx[1], shift_idx[idx] = shift_idx[1], shift_idx[idx], shift_idx[0]
529
- self.actions[idx-1] = np.eye(self.n)[shift_idx]
530
- self.n_actions = len(self.actions)
531
 
532
- self.__info__ = f"Alternating group on n={self.n} objects:\n" \
533
- +f"- Inputs: tokens from 0 to n-3, corresponding to all 3-cycles of the form (12x).\n" \
534
- + "- Labels: depending on 'label_type'.\n" \
535
- + "- Config:\n" \
536
- + " - n (int): number of objects, i.e. there are n!/2 states.\n" \
537
- + self.__info__
538
 
 
 
 
 
 
 
 
 
 
 
 
 
539
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
540
 
541
 
542
  class CyclicAutomaton(Automaton):
543
- def __init__(self, data_config):
544
- super().__init__(data_config)
545
 
546
- if 'n' not in data_config:
547
- data_config['n'] = 5
548
- self.n = data_config['n']
549
 
550
- """
551
- Get actions: shift by i positions, for i = 0 to n_actions-1
552
- """
553
- if 'n_actions' not in data_config:
554
- data_config['n_actions'] = 2
555
- self.n_actions = data_config['n_actions']
556
- shift_idx = list(range(1, self.n)) + [0]
557
- self.actions = {}
558
- for i in range(self.n_actions):
559
- shift_idx = list(range(i, self.n)) + list(range(0, i))
560
- self.actions[i] = np.eye(self.n)[shift_idx]
561
 
562
- self.__info__ = 'Cyclic group of n={self.n} states:\n' \
563
- +f"- Inputs: tokens from 0 to n_actions-1\n" \
564
- + "- Labels: the current state.\n" \
565
- + "- Config:\n" \
566
- + " - n (int): number of states.\n" \
567
- + " - n_actions (int): number of actions/generators, which are 0, 1, ..., n_actions-1.\n" \
568
- + self.__info__
569
 
570
- def f(self, x):
571
- return np.cumsum(x) % self.n
572
 
573
- def sample(self):
574
- T = self.sample_length()
575
- x = self.np_rng.choice(range(self.n_actions), replace=True, size=T)
576
 
577
- return x, self.f(x)
578
 
579
 
580
  class DihedralAutomaton(Automaton):
581
- def __init__(self, data_config):
582
- super().__init__(data_config)
583
 
584
- if 'n' not in data_config:
585
- data_config['n'] = 4
586
- self.n = data_config['n']
587
 
588
- if 'label_type' not in data_config:
589
- # Options: 'state', 'toggle', 'position'
590
- data_config['label_type'] = 'state'
591
- self.label_type = data_config['label_type']
592
 
593
- """
594
- 2 actions: toggle, or shift by 1 position (direction determined by the toggle).
595
- """
596
- self.n_actions = 2
597
- self.actions = {}
598
- # shift all elements to the left (counter-clockwise) by 1
599
- shift_idx = list(range(1, self.n)) + [0]
600
- self.actions[0] = np.eye(self.n)[shift_idx]
601
- # shift all elements to the right (closewise) by 1
602
- shift_idx = [self.n-1] + list(range(self.n-1))
603
- self.actions[1] = np.eye(self.n)[shift_idx]
604
-
605
- self.__info__ = 'Dihedral group of order 2n, where n={self.n}:\n' \
606
- +f"- Inputs: binary tokens:\n" \
607
- + " 0 for toggle, i.e. change direction in the n-cycle;\n" \
608
- + " 1 for drive, i.e. move forward 1 step on the n-cycle.\n" \
609
- + "- Labels: depending on the label_type.\n" \
610
- + "- Config:\n" \
611
- + " - n (int): size of the 'cycle'; i.e. there are 2n states considering also the toggle bit.\n" \
612
- + " - label_type (str): choosing from the following options: \n" \
613
- + " - 'state': the state id, i.e. considering both toggle and position. \n" \
614
- + " - 'toggle': the toggle bit (in {0, 1}). \n" \
615
- + " - 'position': the position on the n-cycle (in [n]).\n" \
616
- + self.__info__
617
-
618
- def f_sequential(self, x):
619
- # sanity check: sequential solution
620
- position = np.arange(self.n)
621
- states = []
622
- toggle = 0 # i.e. parity
623
- for action in x:
624
- if action == 0:
625
- # toggle direction
626
- toggle = 1 - toggle
627
- else:
628
- # drive by 1
629
- position = self.actions[toggle].dot(position)
630
- states += (toggle, position[0]),
631
- return states
632
-
633
- def f(self, x):
634
- # Parallel solution
635
-
636
- # Get toggles: a parity task on the toggle bit
637
- toggles = (x == 0).astype(np.int64)
638
- toggle_status = np.cumsum(toggles) % 2
639
-
640
- # Get positions: a directed modular counter
641
- directions = (-1)**toggle_status
642
- directed_drives = (x != 0).astype(np.int64) * directions
643
- positions = np.cumsum(directed_drives) % self.n
644
-
645
- if self.label_type == 'state':
646
- labels = [self.get_state_label(each) for each in zip(toggle_status, positions)]
647
- return np.array(labels).astype(np.int64)
648
- elif self.label_type == 'toggle':
649
- return toggle_status
650
- elif self.label_type == 'positions':
651
- return positions
652
-
653
- def get_state_label(self, state):
654
- """
655
- toggle in {0,1}
656
- position in [k]
657
- """
658
- toggle, position = state
659
- label = self.n*toggle + position
660
- return label
661
-
662
- def sample(self):
663
- T = self.sample_length()
664
- x = self.np_rng.choice(range(self.n_actions), replace=True, size=T)
665
 
666
- return x, self.f(x)
 
 
667
 
 
668
 
669
 
670
  class QuaternionAutomaton(Automaton):
671
- def __init__(self, data_config):
672
- super().__init__(data_config)
673
-
674
- self.vocab_size = 8 # {-1, 1} x {1, i, j, k}
675
- self.n_actions = 4 # {1, i, j, k}
676
- self.transition_pos = [
677
- 0, 1, 2, 3,
678
- 1, 4, 3, 6,
679
- 2, 7, 4, 1,
680
- 3, 2, 5, 4,]
681
- self.transition_neg = [(each+4)%8 for each in self.transition_pos]
682
- self.transition = np.array(self.transition_pos + self.transition_neg)
683
- self.transition = self.transition.reshape(-1, 4)
684
-
685
- self.__info__ = "Quaternion group:\n" \
686
- + "- Inputs: tokens in {0,1,2,3}, corresponding to 1,i,j,k.\n" \
687
- + "- Labels: the state id; 8 states in total: 2 signs ({-1,1}) x 4 values ({1,i,j,k}).\n" \
688
- + "- Config:\n" \
689
- + self.__info__
690
-
691
- def f(self, x):
692
- curr_state = 0
693
- states = []
694
- for action_id in x:
695
- curr_state = self.transition[curr_state, action_id]
696
- states += curr_state,
697
- return np.array(states).astype(np.int64)
698
-
699
- def sample(self):
700
- T = self.sample_length()
701
- x = self.np_rng.choice(range(self.n_actions), size=T)
702
- return x, self.f(x)
 
703
 
704
 
705
  class PermutationResetAutomaton(Automaton):
706
- def __init__(self, data_config):
707
- super().__init__(data_config)
708
 
709
- self.n = data_config['n']
710
- self.generators = data_config['generators']
711
- self.perm_probs = data_config['perm_probs']
712
- if type(self.generators[0]) is str:
713
- self.generators = [ np.array(list(map(int, list(g)))) for g in self.generators ]
714
 
715
- self.vocab_size = math.factorial(self.n) # states = permutations; maybe rename
716
- self.n_generators = len(self.generators) # actions = generators
717
- self.n_actions = self.vocab_size + self.n_generators # 1 reset symbol per state, 1 apply symbol per generator
718
 
719
- self.init_state = np.arange(self.n) # identity permutation
720
-
721
- # lookup tables
722
- self.int2perm = list(map(np.array, itertools.permutations(range(self.n))))
723
- self.perm2int = {tuple(p):i for i,p in enumerate(self.int2perm)}
724
-
725
- # interval lengths
726
- T = self.sample_length()
727
- self.lags = [1]
728
- while self.lags[-1]*2 < T:
729
- self.lags.append(self.lags[-1]*2)
730
-
731
- def f(self, x):
732
- curr_state = self.init_state
733
- states = []
734
- for action_id in x:
735
- if action_id >= self.vocab_size:
736
- curr_state = self.generators[action_id - self.vocab_size][curr_state]
737
- else:
738
- curr_state = self.int2perm[action_id]
739
- states.append(self.perm2int[tuple(curr_state)])
740
- return np.array(states, dtype=np.int64)
741
-
742
- def sample(self):
743
- T = self.sample_length()
744
- x = self.np_rng.choice(range(self.n_generators), p=self.perm_probs, size=T) + self.vocab_size
745
-
746
- i = 0
747
- while i < T:
748
- x[i] = self.np_rng.choice(range(self.vocab_size))
749
- i += self.np_rng.choice(self.lags)
750
-
751
- return x, self.f(x)
752
 
753
 
754
 
755
  dataset_map = {
756
- 'abab': ABABAutomaton,
757
- 'add': AdderAutomaton,
758
- 'alternating': AlternatingAutomaton,
759
- 'cyclic': CyclicAutomaton,
760
- 'dihedral': DihedralAutomaton,
761
- 'flipflop': FlipFlopAutomaton,
762
- 'gridworld': GridworldAutomaton,
763
- 'parity': ParityAutomaton,
764
- 'quaternion': QuaternionAutomaton,
765
- 'symmetric': SymmetricAutomaton,
766
- 'permutation_reset': PermutationResetAutomaton
767
- # TODO: add Dyck
768
- }
 
109
  }
110
 
111
  class Automaton:
112
+ """
113
+ This is a parent class that must be inherited.
114
+ """
115
+ def __init__(self, data_config):
116
+ self.data_config = data_config
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
117
 
118
+ if 'seed' in self.data_config:
119
+ self.np_rng = np.random.default_rng(self.data_config['seed'])
120
+ else:
121
+ self.np_rng = np.random.default_rng()
122
 
123
+ if 'length' not in data_config: # sequence length
124
+ data_config['length'] = 20
125
+ self.T = self.data_config['length']
 
126
 
127
+ if 'random_length' not in data_config:
128
+ data_config['random_length'] = 0
129
+ self.random_length = data_config['random_length']
 
 
130
 
131
+ self.__info__ = \
132
+ " - T (int): sequence length.\n" \
133
+ + " - random_length (int in {0, 1}): whether to randomly sample a length per sample.\n"
134
 
135
+ def f(self, x):
136
+ """
137
+ Get output sequence given an input seq
138
+ """
139
+ raise NotImplementedError()
140
 
141
+ def sample(self):
142
+ raise NotImplementedError()
 
 
143
 
144
+ def sample_length(self):
145
+ if self.random_length:
146
+ return self.np_rng.choice(range(1, self.T+1))
147
+ return self.T
 
148
 
149
+ def help(self):
150
+ print(self.__info__)
151
 
152
+ class BinaryInputAutomaton(Automaton):
 
 
 
 
 
 
 
 
 
 
 
153
  """
154
+ This is a parent class that must be inherited.
155
+ Subclasses: ParityAutomaton, GridworldAutomaton, ABABAutomaton
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
156
 
157
+ TODO: sample sequences with a given number of 1s
158
+ """
159
+ def __init__(self, data_config):
160
+ super().__init__(data_config)
161
 
162
+ if 'prob1' not in data_config:
163
+ data_config['prob1'] = 0.5
164
+ self.prob1 = data_config['prob1']
165
+ self.__info__ = " - prob1 (float in [0,1]): probability of token 1\n" \
166
+ + self.__info__
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
167
 
168
+ def f(self, x):
169
+ raise NotImplementedError()
170
 
171
+ def sample(self):
172
+ T = self.sample_length()
173
+ x = self.np_rng.binomial(1, self.prob1, size=T)
174
+ return x, self.f(x)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
175
 
176
+ class ParityAutomaton(BinaryInputAutomaton):
177
+ def __init__(self, data_config):
178
+ super().__init__(data_config)
179
+ self.name = 'parity'
180
 
181
+ self.__info__ = "Parity machine with 2 states: \n" \
182
+ + "- Inputs: binary strings\n" \
183
+ + "- Labels: binary strings of the partial parity\n" \
184
+ + "- Config: \n" \
185
+ + self.__info__
 
 
 
 
 
 
186
 
187
+ def f(self, x):
188
+ return np.cumsum(x) % 2
 
 
 
 
189
 
190
+ class GridworldAutomaton(BinaryInputAutomaton):
191
+ """
192
+ Note: gridworld currently doesn't include a no-op.
193
+ """
194
+ def __init__(self, data_config):
195
+ super().__init__(data_config)
196
 
197
+ if 'n' not in data_config:
198
+ data_config['n'] = 9
199
+ """
200
+ NOTE: n is the number of states, and S is the id (0-indexing) of the rightmost state.
201
+ i.e. the states are 0,1,2,...,S, where S=n-1.
202
+ """
203
+ self.n = data_config['n']
204
+ self.S = self.n - 1
205
+
206
+ if 'label_type' not in data_config:
207
+ # Options: state, parity, boundary
208
+ data_config['label_type'] = 'state'
209
+ self.label_type = data_config['label_type']
210
+
211
+ self.name = f'Grid{self.n}'
212
+
213
+ self.__info__ = f"1d Gridworld of n={self.n} states:\n" \
214
+ + "- Inputs: binary strings, i.e. move left(0) or right(1)\n" \
215
+ + "- Labels: depending on 'label_type'. \n" \
216
+ + "- Config: \n" \
217
+ + " - n (int): number of states; i.e. the states are 0,1,2,...,n-1.\n" \
218
+ + " - label_type (str): choosing from the following options:\n" \
219
+ + " - 'state' (default): the state id, i.e. 0 to n-1.\n" \
220
+ + " - 'parity': the state id mod 2.\n" \
221
+ + " - 'boundary': whether the current state is in {0, n-1} or not.\n" \
222
+ + self.__info__
223
+
224
+ def f(self, x):
225
+ x = copy(x)
226
+ x[x == 0] = -1
227
+ if OLD_PY_VERSION:
228
+ # NOTE: for Python 3.7 or below, accumulate doesn't have the 'initial' argument.
229
+ x = np.concatenate([np.array([0]), x]).astype(np.int64)
230
+ states = list(itertools.accumulate(x, lambda a,b: max(min(a+b, self.S), 0)))
231
+ states = states[1:]
232
+ else:
233
+ states = list(itertools.accumulate(x, lambda a,b: max(min(a+b, self.S), 0), initial=0))
234
+ states = states[1:] # remove the 1st entry with is the (meaningless) initial value 0
235
+ return np.array(states).astype(np.int64)
236
 
237
 
238
+ class ABABAutomaton(BinaryInputAutomaton):
239
+ def __init__(self, data_config):
240
+ super().__init__(data_config)
241
+ self.name = 'abab'
242
+
243
+ if 'prob_abab_pos_sample' not in data_config:
244
+ # The probability of having a positive sequence, i.e. 010101010101...
245
+ data_config['prob_abab_pos_sample'] = 0.25
246
+ if 'label_type' not in data_config:
247
+ # Options: 'state', 'boundary'
248
+ data_config['label_type'] = 'state'
249
+
250
+ self.prob_abab_pos_sample = data_config['prob_abab_pos_sample']
251
+ self.label_type = data_config['label_type']
252
+
253
+ self.transition = np.array(
254
+ [[4, 1], # state 0
255
+ [2, 4], # state 1
256
+ [4, 3], # state 2
257
+ [0, 4], # state 3
258
+ [4, 4], # state 4
259
+ ])
260
+
261
+ self.__info__ = "abab: an automaton with 4 states + 1 absorbing state:\n" \
262
+ + "- Inputs: binary strings\n" \
263
+ + "- Labels: depending on 'label_type'.\n" \
264
+ + "- Config:\n" \
265
+ + " - prob_abab_pos_sample (float in [0,1]): probability of having a 'positive' sequence, i.e. 01010101010...\n" \
266
+ + " - label_type (str): choosing from the following options:\n" \
267
+ + " - 'state' (default): the state id.\n" \
268
+ + " - 'boundary': whether the state is in state 3 (the states are 0,1,2,3).\n" \
269
+ + self.__info__
270
+
271
+ def f(self, x):
272
+ labels = []
273
+ curr_state = 3
274
+ for each in x:
275
+ curr_state = self.transition[curr_state, each]
276
+ labels += curr_state,
277
+ labels = np.array(labels).astype(np.int64)
278
+ if self.label_type == 'boundary':
279
+ labels = (labels == 3).astype(np.int64)
280
+ return labels
281
+
282
+ def sample(self):
283
+ pos_sample = self.np_rng.random() < self.prob_abab_pos_sample
284
+ if pos_sample:
285
+ T = self.sample_length()
286
+ x = [0,1,0,1] * (T//4)
287
+ x += [0,1,0,1][:(T%4)]
288
+ x = np.array(x)
289
+ return x, self.f(x)
290
+ else:
291
+ return super().sample()
292
 
293
 
294
+ class AdderAutomaton(BinaryInputAutomaton):
295
+ def __init__(self, data_config):
296
+ super().__init__(data_config)
297
+ self.name = 'addition'
298
+
299
+ if 'n_addends' not in data_config:
300
+ data_config['n_addends'] = 2
301
+ self.n_addends = data_config['n_addends']
302
+ self.addend_scales = np.array([2**i for i in range(self.n_addends)]).reshape(-1, 1)
303
+
304
+ if 'label_type' not in data_config:
305
+ data_config['label_type'] = 'state'
306
+ self.label_type = data_config['label_type']
307
+
308
+ self.__info__ = f'Adder of n={self.n_addends} binary numbers:\n' \
309
+ +f"- Inputs: {self.n_addends} binary numbers, encoded as the int for the {self.n_addends}-bit binary number.\n" \
310
+ + "- Labels: depending on the label_type.\n" \
311
+ + "- Config:\n" \
312
+ + " - n_addends (int): number of binary numbers to be added; default as 2.\n" \
313
+ + " - label_type (str): choosing from the following options: \n" \
314
+ +f" - 'state': the state id, i.e. the int for the base-{self.n_addends} int corresponding to the number (carry, digit). \n" \
315
+ +f" - 'digit': the current output base-{self.n_addends} digit, without the carry. \n" \
316
+ + " - 'position': the current carry bit.\n" \
317
+ + self.__info__
318
+
319
+ def f(self, x):
320
+ outputs, carries = [], []
321
+ carry = 0
322
+ T = x.shape[-1]
323
+ for i in range(T):
324
+ curr_sum = x[:, i].sum() + carry
325
+ # NOTE: 'mod n_addends' makes sure the carry is binary
326
+ output, carry = curr_sum % self.n_addends, curr_sum // self.n_addends
327
+ outputs += output,
328
+ carries += carry,
329
+ outputs = np.array(outputs).astype(np.int64)
330
+ carries = np.array(carries).astype(np.int64)
331
+
332
+ if self.label_type == 'state':
333
+ return outputs + self.n_addends*carries
334
+ elif self.label_type == 'digit':
335
+ return outputs
336
+ elif self.label_type == 'carry':
337
+ return carries
338
+
339
+ def sample_addend(self, T):
340
+ a = self.np_rng.binomial(1, self.prob1, size=T)
341
+ return a
342
+
343
+ def sample(self):
344
+ T = self.sample_length()
345
+ x = np.stack([self.sample_addend(T) for _ in range(self.n_addends)])
346
+ # Pad the most significant bit (rightmost position, i.e. we're reversing the number) with 0 to handle the potential carry
347
+ pad = np.zeros((self.n_addends, 1))
348
+ x = np.concatenate([x, pad], 1)
349
+
350
+ x_encode = (self.addend_scales * x).sum(0)
351
+ return x_encode, self.f(x)
352
 
353
+ class FlipFlopAutomaton(Automaton):
354
+ def __init__(self, data_config):
355
+ super().__init__(data_config)
356
+ self.name = 'flipflop'
 
 
357
 
358
+ if 'n' not in data_config:
359
+ data_config['n'] = 2
360
+
361
+ self.n_states = data_config['n']
362
+ self.n_actions = self.n_states + 1
363
+ self.transition = np.array([list(range(self.n_actions))] + [[i+1]*self.n_actions for i in range(self.n_states)]).T
364
+
365
+ self.__info__ = f"Flipflop with n={self.n_states} states:\n" \
366
+ +f"- Inputs: tokens are either 0 (read) or 1:{self.n} (write).\n" \
367
+ + "- Labels: the state id.\n" \
368
+ + "- Config:\n" \
369
+ + " - n (int): number of write states; i.e. the states are 1,2,...,n, plus a default start state 0.\n" \
370
+ + self.__info__
371
+
372
+ def f(self, x):
373
+ state, states = 0, []
374
+ for action_id in x:
375
+ state = self.transition[state, action_id]
376
+ states += state,
377
+ return np.array(states)
378
+
379
+ def sample(self):
380
+ T = self.sample_length()
381
+ rand = self.np_rng.uniform(size=T)
382
+ nonzero_pos = (rand < 0.5).astype(np.int64)
383
+ writes = self.np_rng.choice(range(1, self.n_states+1), size=T)
384
+ x = writes * nonzero_pos
385
+ return x, self.f(x)
386
 
 
 
 
 
 
387
 
388
+ class PermutationAutomaton(Automaton):
389
+ """
390
+ This is a parent class that must be inherited.
391
+ Subclasses: SymmetricAutomaton, AlternatingAutomaton
392
+ """
393
+ def __init__(self, data_config):
394
+ super().__init__(data_config)
395
+
396
+ if 'n' not in data_config:
397
+ data_config['n'] = 5
398
+ if 'label_type' not in data_config:
399
+ # Options: 'state', 'first_chair'
400
+ data_config['label_type'] = 'state'
401
+
402
+ self.n = data_config['n'] # the symmetric group Sn
403
+ self.label_type = data_config['label_type']
404
 
405
+ self.__info__ = \
406
+ " - label_type (str): choosing from the following options:\n" \
407
+ + " - 'state' (default): the state id.\n" \
408
+ + " - 'first_chair': the element in the first position of the permutation.\n" \
409
+ + " e.g. if the current permutation is [2,1,4,3], then 'first_chair' is 2.\n" \
410
+ + self.__info__
411
 
412
+ def get_state_label(self, state):
413
+ enc = self.state_encode(state)
414
+ return self.state_label_map[enc]
415
 
416
+ def f(self, x):
417
+ curr_state = np.arange(self.n)
418
+ labels = []
419
+ for action_id in x:
420
+ curr_state = self.actions[action_id].dot(curr_state)
421
 
422
+ if self.label_type == 'state':
423
+ labels += self.get_state_label(curr_state),
424
+ elif self.label_type == 'first_chair':
425
+ labels += curr_state[0],
426
+ return np.array(labels)
427
 
428
+ def sample(self):
429
+ T = self.sample_length()
430
+ x = self.np_rng.choice(range(self.n_actions), replace=True, size=T)
 
 
 
 
 
431
 
432
+ return x, self.f(x)
433
 
 
 
 
 
 
 
 
 
434
 
435
+ class SymmetricAutomaton(PermutationAutomaton):
436
  """
437
+ TODO: add options for labels as functions of states
438
+ - parity (whether a state is even): this may need packages (e.g. Permutation from sympy)
439
+ - position / toggle: for S3 ~ D6, we can add labels for substructures as in Dihedral groups.
440
  """
441
+ def __init__(self, data_config):
442
+ super().__init__(data_config)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
443
 
444
+ self.name = f'S{self.n}'
445
 
446
+ """
447
+ Get states
448
+ """
449
+ self.state_encode = lambda state: ''.join([str(int(each)) for each in state])
450
+ self.state_label_map = {}
451
+ for si, state in enumerate(itertools.permutations(range(self.n))):
452
+ enc = self.state_encode(state)
453
+ self.state_label_map[enc] = si
454
 
455
+ """
456
+ Get actions (3 defaults: id, shift-by-1, swap-first-two)
457
+ """
458
+ if 'n_actions' not in data_config:
459
+ data_config['n_actions'] = 3
460
+ self.n_actions = data_config['n_actions']
461
+ self.actions = {0: np.eye(self.n)}
462
+ # shift all elements to the right by 1
463
+ shift_idx = list(range(1, self.n)) + [0]
464
+ self.actions[1] = np.eye(self.n)[shift_idx]
465
+ # swap the first 2 elements
466
+ shift_idx = [1, 0] + list(range(2, self.n))
467
+ self.actions[2] = np.eye(self.n)[shift_idx]
468
+
469
+ if self.n_actions > 3:
470
+ # add permutations in the order given by itertools.permutations
471
+ self.all_permutations = list(itertools.permutations(range(self.n)))[1:]
472
+ cnt = 2
473
+ for each in self.all_permutations:
474
+ action = np.eye(self.n)[list(each)]
475
+ if np.linalg.norm(action - self.actions[0]) == 0:
476
+ continue
477
+ elif np.linalg.norm(action - self.actions[1]) == 0:
478
+ continue
479
+ self.actions[cnt] = action
480
+ cnt += 1
481
+ if cnt == self.n_actions: break
482
+
483
+ self.__info__ = f"Symmetric group on n={self.n} objects:\n" \
484
+ +f"- Inputs: tokens are either 0 (no-op), or 1:{self.n_actions} (corresponding to {self.n_actions} permutations).\n" \
485
+ + "- Labels: depending on 'label_type'.\n" \
486
+ + "- Config:\n" \
487
+ + " - n (int): number of objects, i.e. there are n! states.\n" \
488
+ + " - n_actions (int): number of permutations to include in the generator set;\n" \
489
+ + " the ordering is given by itertools.permutations, and the first 'n_actions' permutations will be included.\n" \
490
+ + self.__info__
491
 
 
 
 
 
 
 
 
 
 
 
 
 
492
 
493
+ class AlternatingAutomaton(PermutationAutomaton):
494
  """
495
+ TODO: other choices of generators (currently using (12x))?
496
  """
497
+ def __init__(self, data_config):
498
+ super().__init__(data_config)
 
 
 
 
 
499
 
500
+ self.name = f'A{self.n}'
 
 
 
 
 
501
 
502
+ """
503
+ Get states
504
+ """
505
+ self.state_label_map = {}
506
+ self.state_encode = lambda state: ''.join([str(int(each)) for each in state])
507
+ cnt = 0
508
+ for si, state in enumerate(itertools.permutations(range(self.n))):
509
+ if not Permutation(state).is_even:
510
+ continue
511
+ enc = self.state_encode(state)
512
+ self.state_label_map[enc] = cnt
513
+ cnt += 1
514
 
515
+ """
516
+ Get actions: all 3 cycles of the form (12x)
517
+ """
518
+ self.actions = {0: np.eye(self.n)}
519
+ for idx in range(2, self.n):
520
+ # (1, 2, idx)
521
+ shift_idx = list(range(self.n))
522
+ shift_idx[0],shift_idx[1], shift_idx[idx] = shift_idx[1], shift_idx[idx], shift_idx[0]
523
+ self.actions[idx-1] = np.eye(self.n)[shift_idx]
524
+ self.n_actions = len(self.actions)
525
+
526
+ self.__info__ = f"Alternating group on n={self.n} objects:\n" \
527
+ +f"- Inputs: tokens from 0 to n-3, corresponding to all 3-cycles of the form (12x).\n" \
528
+ + "- Labels: depending on 'label_type'.\n" \
529
+ + "- Config:\n" \
530
+ + " - n (int): number of objects, i.e. there are n!/2 states.\n" \
531
+ + self.__info__
532
 
533
 
534
  class CyclicAutomaton(Automaton):
535
+ def __init__(self, data_config):
536
+ super().__init__(data_config)
537
 
538
+ if 'n' not in data_config:
539
+ data_config['n'] = 5
540
+ self.n = data_config['n']
541
 
542
+ """
543
+ Get actions: shift by i positions, for i = 0 to n_actions-1
544
+ """
545
+ if 'n_actions' not in data_config:
546
+ data_config['n_actions'] = 2
547
+ self.n_actions = data_config['n_actions']
548
+ shift_idx = list(range(1, self.n)) + [0]
549
+ self.actions = {}
550
+ for i in range(self.n_actions):
551
+ shift_idx = list(range(i, self.n)) + list(range(0, i))
552
+ self.actions[i] = np.eye(self.n)[shift_idx]
553
 
554
+ self.__info__ = 'Cyclic group of n={self.n} states:\n' \
555
+ +f"- Inputs: tokens from 0 to n_actions-1\n" \
556
+ + "- Labels: the current state.\n" \
557
+ + "- Config:\n" \
558
+ + " - n (int): number of states.\n" \
559
+ + " - n_actions (int): number of actions/generators, which are 0, 1, ..., n_actions-1.\n" \
560
+ + self.__info__
561
 
562
+ def f(self, x):
563
+ return np.cumsum(x) % self.n
564
 
565
+ def sample(self):
566
+ T = self.sample_length()
567
+ x = self.np_rng.choice(range(self.n_actions), replace=True, size=T)
568
 
569
+ return x, self.f(x)
570
 
571
 
572
  class DihedralAutomaton(Automaton):
573
+ def __init__(self, data_config):
574
+ super().__init__(data_config)
575
 
576
+ if 'n' not in data_config:
577
+ data_config['n'] = 4
578
+ self.n = data_config['n']
579
 
580
+ if 'label_type' not in data_config:
581
+ # Options: 'state', 'toggle', 'position'
582
+ data_config['label_type'] = 'state'
583
+ self.label_type = data_config['label_type']
584
 
585
+ """
586
+ 2 actions: toggle, or shift by 1 position (direction determined by the toggle).
587
+ """
588
+ self.n_actions = 2
589
+ self.actions = {}
590
+ # shift all elements to the left (counter-clockwise) by 1
591
+ shift_idx = list(range(1, self.n)) + [0]
592
+ self.actions[0] = np.eye(self.n)[shift_idx]
593
+ # shift all elements to the right (closewise) by 1
594
+ shift_idx = [self.n-1] + list(range(self.n-1))
595
+ self.actions[1] = np.eye(self.n)[shift_idx]
596
+
597
+ self.__info__ = 'Dihedral group of order 2n, where n={self.n}:\n' \
598
+ +f"- Inputs: binary tokens:\n" \
599
+ + " 0 for toggle, i.e. change direction in the n-cycle;\n" \
600
+ + " 1 for drive, i.e. move forward 1 step on the n-cycle.\n" \
601
+ + "- Labels: depending on the label_type.\n" \
602
+ + "- Config:\n" \
603
+ + " - n (int): size of the 'cycle'; i.e. there are 2n states considering also the toggle bit.\n" \
604
+ + " - label_type (str): choosing from the following options: \n" \
605
+ + " - 'state': the state id, i.e. considering both toggle and position. \n" \
606
+ + " - 'toggle': the toggle bit (in {0, 1}). \n" \
607
+ + " - 'position': the position on the n-cycle (in [n]).\n" \
608
+ + self.__info__
609
+
610
+ def f_sequential(self, x):
611
+ # sanity check: sequential solution
612
+ position = np.arange(self.n)
613
+ states = []
614
+ toggle = 0 # i.e. parity
615
+ for action in x:
616
+ if action == 0:
617
+ # toggle direction
618
+ toggle = 1 - toggle
619
+ else:
620
+ # drive by 1
621
+ position = self.actions[toggle].dot(position)
622
+ states += (toggle, position[0]),
623
+ return states
624
+
625
+ def f(self, x):
626
+ # Parallel solution
627
+
628
+ # Get toggles: a parity task on the toggle bit
629
+ toggles = (x == 0).astype(np.int64)
630
+ toggle_status = np.cumsum(toggles) % 2
631
+
632
+ # Get positions: a directed modular counter
633
+ directions = (-1)**toggle_status
634
+ directed_drives = (x != 0).astype(np.int64) * directions
635
+ positions = np.cumsum(directed_drives) % self.n
636
+
637
+ if self.label_type == 'state':
638
+ labels = [self.get_state_label(each) for each in zip(toggle_status, positions)]
639
+ return np.array(labels).astype(np.int64)
640
+ elif self.label_type == 'toggle':
641
+ return toggle_status
642
+ elif self.label_type == 'positions':
643
+ return positions
644
+
645
+ def get_state_label(self, state):
646
+ """
647
+ toggle in {0,1}
648
+ position in [k]
649
+ """
650
+ toggle, position = state
651
+ label = self.n*toggle + position
652
+ return label
 
 
 
 
653
 
654
+ def sample(self):
655
+ T = self.sample_length()
656
+ x = self.np_rng.choice(range(self.n_actions), replace=True, size=T)
657
 
658
+ return x, self.f(x)
659
 
660
 
661
  class QuaternionAutomaton(Automaton):
662
+ def __init__(self, data_config):
663
+ super().__init__(data_config)
664
+
665
+ self.vocab_size = 8 # {-1, 1} x {1, i, j, k}
666
+ self.n_actions = 4 # {1, i, j, k}
667
+ self.transition_pos = [
668
+ 0, 1, 2, 3,
669
+ 1, 4, 3, 6,
670
+ 2, 7, 4, 1,
671
+ 3, 2, 5, 4,
672
+ ]
673
+ self.transition_neg = [(each+4)%8 for each in self.transition_pos]
674
+ self.transition = np.array(self.transition_pos + self.transition_neg)
675
+ self.transition = self.transition.reshape(-1, 4)
676
+
677
+ self.__info__ = "Quaternion group:\n" \
678
+ + "- Inputs: tokens in {0,1,2,3}, corresponding to 1,i,j,k.\n" \
679
+ + "- Labels: the state id; 8 states in total: 2 signs ({-1,1}) x 4 values ({1,i,j,k}).\n" \
680
+ + "- Config:\n" \
681
+ + self.__info__
682
+
683
+ def f(self, x):
684
+ curr_state = 0
685
+ states = []
686
+ for action_id in x:
687
+ curr_state = self.transition[curr_state, action_id]
688
+ states += curr_state,
689
+ return np.array(states).astype(np.int64)
690
+
691
+ def sample(self):
692
+ T = self.sample_length()
693
+ x = self.np_rng.choice(range(self.n_actions), size=T)
694
+ return x, self.f(x)
695
 
696
 
697
  class PermutationResetAutomaton(Automaton):
698
+ def __init__(self, data_config):
699
+ super().__init__(data_config)
700
 
701
+ self.n = data_config['n']
702
+ self.generators = data_config['generators']
703
+ self.perm_probs = data_config['perm_probs']
704
+ if type(self.generators[0]) is str:
705
+ self.generators = [ np.array(list(map(int, list(g)))) for g in self.generators ]
706
 
707
+ self.vocab_size = math.factorial(self.n) # states = permutations; maybe rename
708
+ self.n_generators = len(self.generators) # actions = generators
709
+ self.n_actions = self.vocab_size + self.n_generators # 1 reset symbol per state, 1 apply symbol per generator
710
 
711
+ self.init_state = np.arange(self.n) # identity permutation
712
+
713
+ # lookup tables
714
+ self.int2perm = list(map(np.array, itertools.permutations(range(self.n))))
715
+ self.perm2int = {tuple(p):i for i,p in enumerate(self.int2perm)}
716
+
717
+ # interval lengths
718
+ T = self.sample_length()
719
+ self.lags = [1]
720
+ while self.lags[-1]*2 < T:
721
+ self.lags.append(self.lags[-1]*2)
722
+
723
+ def f(self, x):
724
+ curr_state = self.init_state
725
+ states = []
726
+ for action_id in x:
727
+ if action_id >= self.vocab_size:
728
+ curr_state = self.generators[action_id - self.vocab_size][curr_state]
729
+ else:
730
+ curr_state = self.int2perm[action_id]
731
+ states.append(self.perm2int[tuple(curr_state)])
732
+ return np.array(states, dtype=np.int64)
733
+
734
+ def sample(self):
735
+ T = self.sample_length()
736
+ x = self.np_rng.choice(range(self.n_generators), p=self.perm_probs, size=T) + self.vocab_size
737
+
738
+ i = 0
739
+ while i < T:
740
+ x[i] = self.np_rng.choice(range(self.vocab_size))
741
+ i += self.np_rng.choice(self.lags)
742
+
743
+ return x, self.f(x)
744
 
745
 
746
 
747
  dataset_map = {
748
+ 'abab': ABABAutomaton,
749
+ 'add': AdderAutomaton,
750
+ 'alternating': AlternatingAutomaton,
751
+ 'cyclic': CyclicAutomaton,
752
+ 'dihedral': DihedralAutomaton,
753
+ 'flipflop': FlipFlopAutomaton,
754
+ 'gridworld': GridworldAutomaton,
755
+ 'parity': ParityAutomaton,
756
+ 'quaternion': QuaternionAutomaton,
757
+ 'symmetric': SymmetricAutomaton,
758
+ 'permutation_reset': PermutationResetAutomaton
759
+ # TODO: add Dyck
760
+ }