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 ProblemLevel0(Problem):
25 def __init__(self, nb_sentences=100, len_prompt=5, len_result=5):
26 self.seq = torch.randint(10, (nb_sentences, len_prompt + 1 + len_result))
27 self.seq[:, len_prompt] = 10
29 def generate_sequences(self, nb):
30 sequences = self.seq[torch.randint(self.seq.size(0), (nb,))]
31 ar_mask = (sequences == 10).long()
32 ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
33 return sequences, ar_mask
36 class ProblemLevel1(Problem):
37 def __init__(self, nb_operators=100, len_source=5, len_result=8):
38 self.len_source = len_source
39 self.len_result = len_result
40 self.len_nb_operator = int(math.log(nb_operators) / math.log(10)) + 1
41 self.operators = F.one_hot(
42 torch.rand(nb_operators, len_result, len_source).argmax(-1),
43 num_classes=len_source,
46 def generate_sequences(self, nb):
47 nb_operators = torch.randint(self.operators.size(0), (nb,))
48 operators = self.operators[nb_operators]
51 // 10 ** torch.arange(self.len_nb_operator - 1, -1, -1)
53 marker1 = torch.full((nb, 1), 10)
54 # source = torch.randint(10, (nb, self.len_source))
55 source = torch.rand(nb, 10).sort(dim=1).indices[:, : self.len_source]
56 marker2 = torch.full((nb, 1), 11)
57 result = operators.bmm(source[:, :, None]).squeeze(-1)
58 sequences = torch.cat((nb_operators, marker1, source, marker2, result), 1)
59 ar_mask = (sequences == 11).long()
60 ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
61 return sequences, ar_mask
63 def seq2str(self, seq):
64 return "".join("0123456789|>"[x.item()] for x in seq)
67 class ProblemLevel2(Problem):
68 def __init__(self, len_source=5, len_result=8):
69 self.len_source = len_source
70 self.len_result = len_result
72 def generate_sequences(self, nb):
73 operators = F.one_hot(
74 torch.rand(nb, self.len_result, self.len_source).argmax(-1),
75 num_classes=self.len_source,
77 source1 = torch.rand(nb, 10).sort(dim=1).indices[:, : self.len_source]
78 marker1 = torch.full((nb, 1), 10)
79 result1 = operators.bmm(source1[:, :, None]).squeeze(-1)
80 marker2 = torch.full((nb, 1), 11)
81 source2 = torch.randint(10, (nb, self.len_source))
82 marker3 = torch.full((nb, 1), 12)
83 result2 = operators.bmm(source2[:, :, None]).squeeze(-1)
85 sequences = torch.cat(
86 (source1, marker1, result1, marker2, source2, marker3, result2), 1
88 ar_mask = (sequences == 12).long()
89 ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
90 return sequences, ar_mask
92 def seq2str(self, seq):
93 return "".join("0123456789>|~"[x.item()] for x in seq)
99 class ProblemAddition(Problem):
100 def __init__(self, nb_digits=10, zero_padded=False, inverted_result=False):
101 self.nb_digits = nb_digits
102 self.zero_padded = zero_padded
103 self.inverted_result = inverted_result
104 self.char2id = dict([(c, n) for n, c in enumerate("0123456789+=$")])
105 self.id2char = dict([(n, c) for c, n in self.char2id.items()])
107 def tensorize(self, strings):
108 len_max = max([len(x) for x in strings])
113 [self.char2id[c] for c in s + "$" * (len_max - len(s))]
121 def generate_sequences(self, nb):
124 a, b = torch.randint(10**self.nb_digits, (2,))
126 a, b, c = str(a.item()), str(b.item()), str(c.item())
128 a = "0" * (self.nb_digits - len(a)) + a
129 b = "0" * (self.nb_digits - len(b)) + b
130 c = "0" * (self.nb_digits + 1 - len(c)) + c
131 if self.inverted_result:
133 sequences.append(f"{a}+{b}={c}$")
135 sequences = self.tensorize(sequences)
136 ar_mask = (sequences == self.char2id["="]).long()
137 ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
138 return sequences, ar_mask
140 def seq2str(self, seq):
141 return "".join(self.id2char[x.item()] for x in seq)
144 # class ProblemUnion(Problem):
145 # problems = [ProblemByheart()]
146 # nb_common_codes = 100
148 # def generate_sequences(nb_samples):
149 # problem_indexes = torch.randint(len(problems), (nb_samples,))
150 # nb_samples_per_problem = torch.one_hot(problem_indexes).sum(0)
151 # print(f"{nb_samples_per_problem}")
153 # for nb, p in zip(nb_samples_per_problem, problems):
154 # all_seq.append(p.generate_sequences(nb_samples_per_problem[nb]))
157 # for strain, stest in zip(train_seq, test_seq):
158 # s = torch.cat((strain, stest), 0)