-class ProblemTwoTargets(Problem):
- def __init__(self, len_total=10, len_target=2):
- assert len_total >= 3 * (2 + len_target) - 1
- self.len_total = len_total
- self.len_target = len_target
-
- def generate_sequences(self, nb):
- k = torch.arange(self.len_total)[None, :]
- l = torch.randint(self.len_total, (2, nb))[:, :, None] + 1
- i = torch.randint(10, (2, nb))[:, :, None]
- a = l[0]
- b = l[0] + 1 + l[1]
- c = l[0] + 1 + l[1] + 1 + l[0]
- sequences = (
- (k < a) * i[0]
- + (k == a) * 10
- + (k > a) * (k < b) * i[1]
- + (k == b) * 11
- + (k > b) * (k < c) * i[1]
- + (k >= c) * 12
- )
- ar_mask = (sequences == 11).long()
- ar_mask = (ar_mask.cumsum(1) - ar_mask).clamp(max=1)
- return sequences, ar_mask
+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