nb_correct = ((result == input).long() * ar_mask).sum().item()
return nb_total, nb_correct
-####################
+####################
+class ProblemDegradation(Problem):
+ def __init__(self, nb_state_tokens=5, nb_time_steps=12, value_max=25, hard=False):
+ assert value_max // nb_state_tokens >= 2
+ self.nb_state_tokens = nb_state_tokens
+ self.nb_time_steps = nb_time_steps
+ self.value_max = value_max
+ self.hard = hard
-class ProblemTwoCuts(Problem):
- def __init__(self, len_total=50, nb_values=100, global_constraint=True):
- self.len_total = len_total
- self.nb_values = nb_values
- self.global_constraint = global_constraint
-
- def generate_sequences_internal(self, nb):
- return u,v,a,b,c
-
- def generate_sequences(self,nb):
-
- u = torch.randint(self.len_total, (nb,))
- v = torch.randint(self.len_total, (nb,))
-
- a = torch.randint(self.nb_values, (nb,))
- b = torch.randint(self.nb_values, (nb,))
- c = torch.randint(self.nb_values, (nb,))
-
- while True:
- to_compute = torch.logical_or(u>=v-self.len_total//10,u<v-self.len_total//5)
- to_compute =torch.logical_or(to_compute, u == 0)
- to_compute =torch.logical_or(to_compute, v == self.len_total)
- n = to_compute.long().sum()
- if n == 0:
- break
- else:
- u[to_compute] = torch.randint(self.len_total, (n,))
- v[to_compute] = torch.randint(self.len_total, (n,))
-
- while True:
- to_compute = a==b
- to_compute = torch.logical_or(to_compute,b==c)
- to_compute = torch.logical_or(to_compute,a==c)
-
- if self.global_constraint:
- to_compute = torch.logical_or(to_compute,(a*u+b*(v-u)+c*(self.len_total-v)) // self.len_total != self.nb_values//2)
-
- n = to_compute.long().sum()
- if n == 0:
- break
- else:
- a[to_compute] = torch.randint(self.nb_values, (n,))
- b[to_compute] = torch.randint(self.nb_values, (n,))
- c[to_compute] = torch.randint(self.nb_values, (n,))
-
- assert (u>=v).long().sum() == 0
- assert (a==b).long().sum() == 0
- assert (a==c).long().sum() == 0
- assert (c==b).long().sum() == 0
-
- t = torch.arange(self.len_total)
- seq = (t[None,:] < u[:,None]).long() * a[:,None] + \
- (t[None,:] >= u[:,None]).long() * (t[None,:] < v[:,None]).long() * b[:,None] + \
- (t[None,:] >= v[:,None]).long() * c[:,None]
-
- return seq,seq.new_full(seq.size(), 1, dtype=torch.int64)
+ 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()).sort(dim=-1).indices + 1) * (x >= 2).long()
+ u = (v.max(dim=-1, keepdim=True).values == v).long()
+ n = (
+ (u * x)
+ .minimum(2 + torch.randint(self.value_max // 4 - 2, x.size()))
+ .sum(dim=-1, keepdim=True)
+ )
+ m = 1 + ((n - 1) * torch.rand(n.size())).long()
+ x = (
+ x
+ + m * u.roll(shifts=-1, dims=-1)
+ - n * u
+ + (n - m) * 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
- i = torch.arange(result.size(1), device=result.device)
-
- for k in range(nb_total):
- s = result[k]
- a = s[0]
- uu = (s != a).nonzero()
- if uu.size(0) > 0:
- u = uu.min()
- b = s[u]
- vv = torch.logical_and(s != b, i >= u).nonzero()
- if vv.size(0) > 0:
- v = vv.min()
- c = s[v]
- ww = torch.logical_and(s != c, i >= v).nonzero()
- if ww.size(0) == 0:
- if not self.global_constraint or (a*u+b*(v-u)+c*(self.len_total-v)) // self.len_total == self.nb_values//2:
- nb_correct += 1
+ 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 + 1] - states[k]
+ j = d.sort(descending=False).indices[0]
+ if (
+ d[j] == 0
+ or d[j] > self.value_max // 4
+ or d[(j + 1) % e.size(0)] <= 0
+ or d[(j + 1) % e.size(0)] >= -d[j]
+ ):
+ nb_errors += 1
+ else:
+ e.zero_()
+ e[j] = d[j]
+ e[(j + 1) % e.size(0)] = d[(j + 1) % e.size(0)]
+ e[(j - 1) % e.size(0)] = -d[(j + 1) % e.size(0)] - d[j]
+ 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( [ f"{x:02d}" for x in seq ] )
+ return " | ".join(
+ [" ".join([f"{x:02d}" for x in s]) for s in seq.split(self.nb_state_tokens)]
+ )
+
####################
return "".join(self.id2char[x.item()] for x in seq)
+####################
+
+
+class ProblemMixing(Problem):
+ def __init__(self, height=3, width=3, nb_time_steps=12, hard=False):
+ self.height = height
+ self.width = width
+ self.nb_time_steps = nb_time_steps
+ self.hard = hard
+
+ def start_random(self, nb):
+ y = torch.arange(self.height * self.width).reshape(1, -1).expand(nb, -1)
+
+ m = (torch.rand(y.size()).sort(dim=-1).indices < y.size(1) // 2).long()
+
+ y = (y * m + self.height * self.width * (1 - m)).reshape(
+ nb, self.height, self.width
+ )
+
+ return y
+
+ def start_error(self, x):
+ x = x.flatten(1)
+ u = torch.arange(self.height * self.width).reshape(1, -1)
+ m = ((x - u).abs() == 0).long()
+ d = (x - (m * u + (1-m) * self.height * self.width)).abs().sum(-1) + (
+ m.sum(dim=-1) != self.height * self.width // 2
+ ).long()
+ return d
+
+ def moves(self, x):
+ y = (
+ x[:, None, :, :]
+ .expand(-1, self.height * 2 + self.width * 2, -1, -1)
+ .clone()
+ )
+ k = 0
+
+ for i in range(self.height):
+ y[:, k, i, :] = y[:, k, i, :].roll(dims=-1, shifts=-1)
+ k += 1
+ y[:, k, i, :] = y[:, k, i, :].roll(dims=-1, shifts=1)
+ k += 1
+
+ for j in range(self.width):
+ y[:, k, :, j] = y[:, k, :, j].roll(dims=-1, shifts=-1)
+ k += 1
+ y[:, k, :, j] = y[:, k, :, j].roll(dims=-1, shifts=1)
+ k += 1
+
+ return y
+
+ def generate_sequences(self, nb):
+ x = self.start_random(nb)
+
+ seq = [x.flatten(1)]
+
+ for t in range(self.nb_time_steps - 1):
+ y = self.moves(x)
+ x = y[torch.arange(nb), torch.randint(y.size(1), (nb,))]
+ seq.append(x.flatten(1))
+
+ 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):
+ a = [
+ x.reshape(result.size(0), self.height, self.width)
+ for x in result.split(self.height * self.width, dim=1)
+ ]
+ if self.hard:
+ a.reverse()
+
+ x = a[0]
+
+ d = self.start_error(x)
+
+ for t in range(self.nb_time_steps - 1):
+ x0, x = a[t], a[t + 1]
+ y = self.moves(x0)
+ d = d + (x[:, None] - y).abs().sum((-1, -2)).min(dim=-1).values
+
+ nb_total, nb_correct = result.size(0), (d == 0).long().sum().item()
+
+ return nb_total, nb_correct
+
+ def seq2str(self, seq):
+ return " | ".join(
+ [
+ " ".join(
+ ["-".join([f"{x:02d}" for x in s]) for s in r.split(self.width)]
+ )
+ for r in seq.split(self.height * self.width)
+ ]
+ )
+
+
+####################
+
if __name__ == "__main__":
- p = ProblemTwoCuts(12)
+ p = ProblemMixing(width=4, hard=True)
s, m = p.generate_sequences(10000)
+ for x in s[:5]:
+ print(p.seq2str(x))
print(p.compute_nb_correct(None, None, s))
projects / picoclvr.git / blobdiff