cyrilzhang
commited on
Commit
•
0c45ea8
1
Parent(s):
2a180d8
add PermutationResetSampler
Browse files- automata.py +49 -0
automata.py
CHANGED
@@ -699,7 +699,55 @@ class QuaternionSampler(AutomatonSampler):
|
|
699 |
return x, self.f(x)
|
700 |
|
701 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
702 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
703 |
|
704 |
|
705 |
|
@@ -714,6 +762,7 @@ dataset_map = {
|
|
714 |
'parity': ParitySampler,
|
715 |
'quaternion': QuaternionSampler,
|
716 |
'symmetric': SymmetricSampler,
|
|
|
717 |
# TODO: add Dyck
|
718 |
}
|
719 |
|
|
|
699 |
return x, self.f(x)
|
700 |
|
701 |
|
702 |
+
class PermutationResetSampler(AutomatonSampler):
|
703 |
+
def __init__(self, data_config):
|
704 |
+
super().__init__(data_config)
|
705 |
+
|
706 |
+
self.n = data_config['n']
|
707 |
+
self.generators = data_config['generators']
|
708 |
+
self.perm_probs = data_config['perm_probs']
|
709 |
+
if type(self.generators[0]) is str:
|
710 |
+
self.generators = [ np.array(list(map(int, list(g)))) for g in self.generators ]
|
711 |
+
|
712 |
+
self.vocab_size = math.factorial(self.n) # states = permutations; maybe rename
|
713 |
+
self.n_generators = len(self.generators) # actions = generators
|
714 |
+
self.n_actions = self.vocab_size + self.n_generators # 1 reset symbol per state, 1 apply symbol per generator
|
715 |
|
716 |
+
self.init_state = np.arange(self.n) # identity permutation
|
717 |
+
|
718 |
+
# lookup tables
|
719 |
+
self.int2perm = list(map(np.array, itertools.permutations(range(self.n))))
|
720 |
+
self.perm2int = {tuple(p):i for i,p in enumerate(self.int2perm)}
|
721 |
+
|
722 |
+
# interval lengths
|
723 |
+
T = self.sample_length()
|
724 |
+
self.lags = [1]
|
725 |
+
while self.lags[-1]*2 < T:
|
726 |
+
self.lags.append(self.lags[-1]*2)
|
727 |
+
|
728 |
+
def f(self, x):
|
729 |
+
curr_state = self.init_state
|
730 |
+
states = []
|
731 |
+
for action_id in x:
|
732 |
+
if action_id >= self.vocab_size:
|
733 |
+
curr_state = self.generators[action_id - self.vocab_size][curr_state]
|
734 |
+
else:
|
735 |
+
curr_state = self.int2perm[action_id]
|
736 |
+
states.append(self.perm2int[tuple(curr_state)])
|
737 |
+
return np.array(states, dtype=np.int64)
|
738 |
+
|
739 |
+
def sample(self):
|
740 |
+
T = self.sample_length()
|
741 |
+
x = self.np_rng.choice(range(self.n_generators), p=self.perm_probs, size=T) + self.vocab_size
|
742 |
+
|
743 |
+
i = 0
|
744 |
+
reset_in = 0 # always start with reset
|
745 |
+
while i < T:
|
746 |
+
if reset_in == 0:
|
747 |
+
x[i] = self.np_rng.choice(range(self.vocab_size))
|
748 |
+
i += self.np_rng.choice(self.lags)
|
749 |
+
|
750 |
+
return x, self.f(x)
|
751 |
|
752 |
|
753 |
|
|
|
762 |
'parity': ParitySampler,
|
763 |
'quaternion': QuaternionSampler,
|
764 |
'symmetric': SymmetricSampler,
|
765 |
+
'permutation_reset': PermutationResetSampler
|
766 |
# TODO: add Dyck
|
767 |
}
|
768 |
|