+class ProblemDegradation(Problem):
+ def __init__(self, nb_state_tokens=5, nb_time_steps=5, value_max=25, hard=False):
+ self.nb_state_tokens = nb_state_tokens
+ self.nb_time_steps = nb_time_steps
+ self.value_max = value_max
+ self.hard = hard
+
+ def generate_sequences(self,nb):
+
+ x = (torch.rand(nb,self.nb_state_tokens).sort(dim=-1).indices == 0).long() * self.value_max
+ seq = [x]
+
+ for t in range(self.nb_time_steps-1):
+ v = torch.rand(x.size()) * (x > 0).float()
+ u = (v.max(dim=-1,keepdim=True).values == v).long()
+ n = (u*x*torch.rand(x.size())).long().sum(dim=-1,keepdim=True) // 2
+ x = x + n * (u.roll(shifts=-1,dims=-1) - 2 * u + u.roll(shifts=1,dims=-1))
+ seq.append(x)
+
+ if self.hard: seq.reverse()
+
+ seq = torch.cat(seq,dim=1)
+ return seq,seq.new_full(seq.size(), 1, dtype=torch.int64)
+
+ def compute_nb_correct(self, input, ar_mask, result):
+ nb_total = result.size(0)
+ nb_correct = 0
+ e=result.new_zeros(self.nb_state_tokens)
+
+ for seq in result:
+ states = list(seq.split(self.nb_state_tokens))
+ if self.hard:
+ states.reverse()
+
+ d = states[0]
+ j=d.sort(descending=True).indices[0]
+ e.zero_()
+ e[j]=self.value_max
+ if (d-e).abs().sum() == 0:
+ nb_errors = 0
+ for k in range(len(states)-1):
+ d=states[k]-states[k+1]
+ j=d.sort(descending=True).indices[0]
+ e.zero_()
+ e[j]=d[j]
+ e[(j+1)%e.size(0)]=-d[j]//2
+ e[(j-1)%e.size(0)]=-d[j]//2
+ if (d-e).abs().sum() > 0:
+ nb_errors += 1
+ if nb_errors == 0:
+ nb_correct += 1
+
+ return nb_total, nb_correct
+
+ def seq2str(self, seq):
+ return " | ".join( [ " ".join([f"{x:02d}" for x in s ]) for s in seq.split(self.nb_state_tokens) ] )
+
+####################
+
+
+class ProblemTwoTargets(Problem):
+ def __init__(self, len_total=10, len_targets=3):
+ assert len_targets >= 3
+ assert len_total >= 3 * len_targets - 1
+ self.len_total = len_total
+ self.len_targets = len_targets
+
+ def generate_sequences(self, nb):
+ k = torch.arange(self.len_total)[None, :]
+ s = torch.randint(10, (nb, self.len_total))
+ l = torch.rand(nb, self.len_total)
+ l = l * (k <= self.len_total - self.len_targets).long()
+ k1 = l.argmax(dim=1, keepdim=True)
+ m = (k != k1).long() * (k != k1 + self.len_targets - 1).long()
+ s = s * m + 10 * (1 - m)
+ l = l * (
+ 1
+ - (k + self.len_targets - 1 >= k1).long()
+ * (k < k1 + self.len_targets).long()
+ )
+ k2 = l.argmax(dim=1, keepdim=True)
+ m = (k != k2).long() * (k != k2 + self.len_targets - 1).long()
+ s = s * m + 11 * (1 - m)
+ a1 = s.gather(dim=1, index=k1 + 1 + torch.arange(self.len_targets - 2)[None, :])
+ a2 = s.gather(dim=1, index=k2 + 1 + torch.arange(self.len_targets - 2)[None, :])
+ sequences = torch.cat(
+ (
+ s,
+ torch.full((nb, 1), 12),
+ a1,
+ torch.full((nb, 1), 12),
+ a2,
+ torch.full((nb, 1), 12),
+ ),
+ 1,
+ )
+ ar_mask = (sequences == 12).long()
+ ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
+ return sequences, ar_mask
+
+ def seq2str(self, seq):
+ return "".join("0123456789-+|"[x.item()] for x in seq)
+
+
+####################
+
+
+class ProblemByHeart(Problem):
+ def __init__(self, nb_sentences=100, len_prompt=8, len_result=8):