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]"
24 class ProblemTwoTargets(Problem):
25 def __init__(self, len_total=10, len_targets=3):
26 assert len_targets >= 3
27 assert len_total >= 3 * len_targets - 1
28 self.len_total = len_total
29 self.len_targets = len_targets
31 def generate_sequences(self, nb):
32 k = torch.arange(self.len_total)[None, :]
33 s = torch.randint(10, (nb, self.len_total))
34 l = torch.rand(nb, self.len_total)
35 l = l * (k <= self.len_total - self.len_targets).long()
36 k1 = l.argmax(dim=1, keepdim=True)
37 m = (k != k1).long() * (k != k1 + self.len_targets - 1).long()
38 s = s * m + 10 * (1 - m)
41 - (k + self.len_targets - 1 >= k1).long()
42 * (k < k1 + self.len_targets).long()
44 k2 = l.argmax(dim=1, keepdim=True)
45 m = (k != k2).long() * (k != k2 + self.len_targets - 1).long()
46 s = s * m + 11 * (1 - m)
47 a1 = s.gather(dim=1, index=k1 + 1 + torch.arange(self.len_targets - 2)[None, :])
48 a2 = s.gather(dim=1, index=k2 + 1 + torch.arange(self.len_targets - 2)[None, :])
49 sequences = torch.cat(
50 (s, torch.full((nb, 1), 12), a1, torch.full((nb, 1), 12), a2), 1
52 ar_mask = (sequences == 12).long()
53 ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
54 return sequences, ar_mask
56 def seq2str(self, seq):
57 return "".join("0123456789+-|"[x.item()] for x in seq)
63 class ProblemLenId(Problem):
64 def __init__(self, len_max=10):
65 self.len_max = len_max
67 def generate_sequences(self, nb):
68 k = torch.arange(self.len_max * 3 + 3)[None, :]
69 l = torch.randint(self.len_max, (2, nb))[:, :, None] + 1
70 i = torch.randint(10, (2, nb))[:, :, None]
73 c = l[0] + 1 + l[1] + 1 + l[0]
77 + (k > a) * (k < b) * i[1]
79 + (k > b) * (k < c) * i[1]
82 ar_mask = (sequences == 11).long()
83 ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
84 return sequences, ar_mask
86 def seq2str(self, seq):
87 return "".join("0123456789|>_"[x.item()] for x in seq)
93 class ProblemLevel0(Problem):
94 def __init__(self, nb_sentences=100, len_prompt=5, len_result=5):
95 self.seq = torch.randint(10, (nb_sentences, len_prompt + 1 + len_result))
96 self.seq[:, len_prompt] = 10
98 def generate_sequences(self, nb):
99 sequences = self.seq[torch.randint(self.seq.size(0), (nb,))]
100 ar_mask = (sequences == 10).long()
101 ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
102 return sequences, ar_mask
104 def seq2str(self, seq):
105 return "".join("0123456789|"[x.item()] for x in seq)
111 class ProblemLevel1(Problem):
112 def __init__(self, nb_operators=100, len_source=5, len_result=8):
113 self.len_source = len_source
114 self.len_result = len_result
115 self.len_nb_operator = int(math.log(nb_operators) / math.log(10)) + 1
116 self.operators = F.one_hot(
117 torch.rand(nb_operators, len_result, len_source).argmax(-1),
118 num_classes=len_source,
121 def generate_sequences(self, nb):
122 nb_operators = torch.randint(self.operators.size(0), (nb,))
123 operators = self.operators[nb_operators]
125 nb_operators[:, None]
126 // 10 ** torch.arange(self.len_nb_operator - 1, -1, -1)
128 marker1 = torch.full((nb, 1), 10)
129 # source = torch.randint(10, (nb, self.len_source))
130 source = torch.rand(nb, 10).sort(dim=1).indices[:, : self.len_source]
131 marker2 = torch.full((nb, 1), 11)
132 result = operators.bmm(source[:, :, None]).squeeze(-1)
133 sequences = torch.cat((nb_operators, marker1, source, marker2, result), 1)
134 ar_mask = (sequences == 11).long()
135 ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
136 return sequences, ar_mask
138 def seq2str(self, seq):
139 return "".join("0123456789|>"[x.item()] for x in seq)
145 class ProblemLevel2(Problem):
146 def __init__(self, len_source=5, len_result=8):
147 self.len_source = len_source
148 self.len_result = len_result
150 def generate_sequences(self, nb):
151 operators = F.one_hot(
152 torch.rand(nb, self.len_result, self.len_source).argmax(-1),
153 num_classes=self.len_source,
155 source1 = torch.rand(nb, 10).sort(dim=1).indices[:, : self.len_source]
156 marker1 = torch.full((nb, 1), 10)
157 result1 = operators.bmm(source1[:, :, None]).squeeze(-1)
158 marker2 = torch.full((nb, 1), 11)
159 source2 = torch.randint(10, (nb, self.len_source))
160 marker3 = torch.full((nb, 1), 12)
161 result2 = operators.bmm(source2[:, :, None]).squeeze(-1)
163 sequences = torch.cat(
164 (source1, marker1, result1, marker2, source2, marker3, result2), 1
166 ar_mask = (sequences == 12).long()
167 ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
168 return sequences, ar_mask
170 def seq2str(self, seq):
171 return "".join("0123456789>|~"[x.item()] for x in seq)
177 class ProblemAddition(Problem):
178 def __init__(self, nb_digits=10, zero_padded=False, inverted_result=False):
179 self.nb_digits = nb_digits
180 self.zero_padded = zero_padded
181 self.inverted_result = inverted_result
182 self.char2id = dict([(c, n) for n, c in enumerate("0123456789+=$")])
183 self.id2char = dict([(n, c) for c, n in self.char2id.items()])
185 def tensorize(self, strings):
186 len_max = max([len(x) for x in strings])
191 [self.char2id[c] for c in s + "$" * (len_max - len(s))]
199 def generate_sequences(self, nb):
202 a, b = torch.randint(10**self.nb_digits, (2,))
204 a, b, c = str(a.item()), str(b.item()), str(c.item())
206 a = "0" * (self.nb_digits - len(a)) + a
207 b = "0" * (self.nb_digits - len(b)) + b
208 c = "0" * (self.nb_digits + 1 - len(c)) + c
209 if self.inverted_result:
211 sequences.append(f"{a}+{b}={c}$")
213 sequences = self.tensorize(sequences)
214 ar_mask = (sequences == self.char2id["="]).long()
215 ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
216 return sequences, ar_mask
218 def seq2str(self, seq):
219 return "".join(self.id2char[x.item()] for x in seq)
222 if __name__ == "__main__":
223 p = ProblemTwoTargets(12, 4)
224 s, m = p.generate_sequences(10)