5 import torch, torchvision
8 from torch.nn import functional as F
10 ######################################################################
14 def generate_sequences(self, nb):
17 def seq2str(self, seq):
18 return "[NOT IMPLEMENTED]"
20 def compute_nb_correct(self, input, ar_mask, result):
21 nb_total = ar_mask.sum().item()
22 nb_correct = ((result == input).long() * ar_mask).sum().item()
23 return nb_total, nb_correct
28 class ProblemDegradation(Problem):
29 def __init__(self, nb_state_tokens=7, nb_time_steps=10, value_max=100, hard=False):
30 self.nb_state_tokens = nb_state_tokens
31 self.nb_time_steps = nb_time_steps
32 self.value_max = value_max
35 def generate_sequences(self,nb):
37 x = (torch.rand(nb,self.nb_state_tokens).sort(dim=-1).indices == 0).long() * self.value_max
40 for t in range(self.nb_time_steps-1):
41 v = torch.rand(x.size()) * (x > 0).float()
42 u = (v.max(dim=-1,keepdim=True).values == v).long()
43 n = (u*x*torch.rand(x.size())).long().sum(dim=-1,keepdim=True) // 2
44 x = x + n * (u.roll(shifts=-1,dims=-1) - 2 * u + u.roll(shifts=1,dims=-1))
47 if self.hard: seq.reverse()
49 seq = torch.cat(seq,dim=1)
50 return seq,seq.new_full(seq.size(), 1, dtype=torch.int64)
52 def compute_nb_correct(self, input, ar_mask, result):
53 nb_total = result.size(0)
55 e=result.new_zeros(self.nb_state_tokens)
58 states = list(seq.split(self.nb_state_tokens))
63 j=d.sort(descending=True).indices[0]
66 if (d-e).abs().sum() == 0:
68 for k in range(len(states)-1):
69 d=states[k]-states[k+1]
70 j=d.sort(descending=True).indices[0]
73 e[(j+1)%e.size(0)]=-d[j]//2
74 e[(j-1)%e.size(0)]=-d[j]//2
75 if (d-e).abs().sum() > 0:
80 return nb_total, nb_correct
82 def seq2str(self, seq):
83 return " | ".join( [ " ".join([f"{x:02d}" for x in s ]) for s in seq.split(self.nb_state_tokens) ] )
88 class ProblemTwoTargets(Problem):
89 def __init__(self, len_total=10, len_targets=3):
90 assert len_targets >= 3
91 assert len_total >= 3 * len_targets - 1
92 self.len_total = len_total
93 self.len_targets = len_targets
95 def generate_sequences(self, nb):
96 k = torch.arange(self.len_total)[None, :]
97 s = torch.randint(10, (nb, self.len_total))
98 l = torch.rand(nb, self.len_total)
99 l = l * (k <= self.len_total - self.len_targets).long()
100 k1 = l.argmax(dim=1, keepdim=True)
101 m = (k != k1).long() * (k != k1 + self.len_targets - 1).long()
102 s = s * m + 10 * (1 - m)
105 - (k + self.len_targets - 1 >= k1).long()
106 * (k < k1 + self.len_targets).long()
108 k2 = l.argmax(dim=1, keepdim=True)
109 m = (k != k2).long() * (k != k2 + self.len_targets - 1).long()
110 s = s * m + 11 * (1 - m)
111 a1 = s.gather(dim=1, index=k1 + 1 + torch.arange(self.len_targets - 2)[None, :])
112 a2 = s.gather(dim=1, index=k2 + 1 + torch.arange(self.len_targets - 2)[None, :])
113 sequences = torch.cat(
116 torch.full((nb, 1), 12),
118 torch.full((nb, 1), 12),
120 torch.full((nb, 1), 12),
124 ar_mask = (sequences == 12).long()
125 ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
126 return sequences, ar_mask
128 def seq2str(self, seq):
129 return "".join("0123456789-+|"[x.item()] for x in seq)
135 class ProblemByHeart(Problem):
136 def __init__(self, nb_sentences=100, len_prompt=8, len_result=8):
137 self.seq = torch.randint(10, (nb_sentences, len_prompt + 1 + len_result))
138 self.seq[:, len_prompt] = 10
140 def generate_sequences(self, nb):
141 sequences = self.seq[torch.randint(self.seq.size(0), (nb,))]
142 ar_mask = (sequences == 10).long()
143 ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
144 return sequences, ar_mask
146 def seq2str(self, seq):
147 return "".join("0123456789|"[x.item()] for x in seq)
153 class ProblemLearnOperator(Problem):
154 def __init__(self, nb_operators=100, len_source=6, len_result=9):
155 self.len_source = len_source
156 self.len_result = len_result
157 self.len_nb_operator = int(math.log(nb_operators) / math.log(10)) + 1
158 self.operators = F.one_hot(
159 torch.rand(nb_operators, len_result, len_source).argmax(-1),
160 num_classes=len_source,
163 def generate_sequences(self, nb):
164 nb_operators = torch.randint(self.operators.size(0), (nb,))
165 operators = self.operators[nb_operators]
167 nb_operators[:, None]
168 // 10 ** torch.arange(self.len_nb_operator - 1, -1, -1)
170 marker1 = torch.full((nb, 1), 10)
171 source = torch.rand(nb, 10).sort(dim=1).indices[:, : self.len_source]
172 marker2 = torch.full((nb, 1), 11)
173 result = operators.bmm(source[:, :, None]).squeeze(-1)
174 sequences = torch.cat((nb_operators, marker1, source, marker2, result), 1)
175 ar_mask = (sequences == 11).long()
176 ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
177 return sequences, ar_mask
179 def seq2str(self, seq):
180 return "".join("0123456789|>"[x.item()] for x in seq)
186 class ProblemGuessOperator(Problem):
187 def __init__(self, len_source=5, len_result=8):
188 self.len_source = len_source
189 self.len_result = len_result
191 def generate_sequences(self, nb):
192 operators = F.one_hot(
193 torch.rand(nb, self.len_result, self.len_source).argmax(-1),
194 num_classes=self.len_source,
196 source1 = torch.rand(nb, 10).sort(dim=1).indices[:, : self.len_source]
197 marker1 = torch.full((nb, 1), 10)
198 result1 = operators.bmm(source1[:, :, None]).squeeze(-1)
199 marker2 = torch.full((nb, 1), 11)
200 source2 = torch.randint(10, (nb, self.len_source))
201 marker3 = torch.full((nb, 1), 12)
202 result2 = operators.bmm(source2[:, :, None]).squeeze(-1)
204 sequences = torch.cat(
205 (source1, marker1, result1, marker2, source2, marker3, result2), 1
207 ar_mask = (sequences == 12).long()
208 ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
209 return sequences, ar_mask
211 def seq2str(self, seq):
212 return "".join("0123456789>|~"[x.item()] for x in seq)
218 class ProblemAddition(Problem):
219 def __init__(self, nb_digits=10, zero_padded=False, inverted_result=False):
220 self.nb_digits = nb_digits
221 self.zero_padded = zero_padded
222 self.inverted_result = inverted_result
223 self.char2id = dict([(c, n) for n, c in enumerate("0123456789+=$")])
224 self.id2char = dict([(n, c) for c, n in self.char2id.items()])
226 def tensorize(self, strings):
227 len_max = max([len(x) for x in strings])
232 [self.char2id[c] for c in s + "$" * (len_max - len(s))]
240 def generate_sequences(self, nb):
243 a, b = torch.randint(10**self.nb_digits, (2,))
245 a, b, c = str(a.item()), str(b.item()), str(c.item())
247 a = "0" * (self.nb_digits - len(a)) + a
248 b = "0" * (self.nb_digits - len(b)) + b
249 c = "0" * (self.nb_digits + 1 - len(c)) + c
250 if self.inverted_result:
252 sequences.append(f"{a}+{b}={c}$")
254 sequences = self.tensorize(sequences)
255 ar_mask = (sequences == self.char2id["="]).long()
256 ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
257 return sequences, ar_mask
259 def seq2str(self, seq):
260 return "".join(self.id2char[x.item()] for x in seq)
263 if __name__ == "__main__":
264 p = ProblemDegradation(hard=False)
265 s, m = p.generate_sequences(10000)
268 print(p.compute_nb_correct(None, None, s))