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=5, value_max=25, hard=False):
+ 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
- def generate_sequences(self,nb):
-
- x = (torch.rand(nb,self.nb_state_tokens).sort(dim=-1).indices == 0).long() * self.value_max
+ 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))
+ 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()
+ if self.hard:
+ seq.reverse()
- seq = torch.cat(seq,dim=1)
- return seq,seq.new_full(seq.size(), 1, dtype=torch.int64)
+ 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)
+ e = result.new_zeros(self.nb_state_tokens)
for seq in result:
states = list(seq.split(self.nb_state_tokens))
states.reverse()
d = states[0]
- j=d.sort(descending=True).indices[0]
+ j = d.sort(descending=True).indices[0]
e.zero_()
- e[j]=self.value_max
- if (d-e).abs().sum() == 0:
+ 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:
+ 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( [ " ".join([f"{x:02d}" for x in s ]) for s in seq.split(self.nb_state_tokens) ] )
+ return " | ".join(
+ [" ".join([f"{x:02d}" for x in s]) for s in seq.split(self.nb_state_tokens)]
+ )
+
####################
+class ProblemMemory(Problem):
+ def __init__(self, len_total=25):
+ self.len_total = len_total
+ self.max_len_pattern = 5
+ self.nb_noise_tokens = 10
+ self.start_pattern_token = 0
+ self.end_pattern_token = 1
+ self.start_result_token = 2
+ self.end_result_token = 3
+ self.token_string = "[]<>" + "".join(
+ [chr(ord("a") + k) for k in range(self.nb_noise_tokens)]
+ )
+
+ def generate_sequences(self, nb):
+ sequences = (
+ torch.randint(self.nb_noise_tokens, (nb, self.len_total))
+ + self.end_result_token
+ + 1
+ )
+ len_patterns = torch.randint(self.max_len_pattern, (nb,)) + 1
+ pattern_positions = torch.randint(
+ self.len_total - (5 + 2 * self.max_len_pattern), (nb,)
+ )
+ k = self.len_total - (3 + self.max_len_pattern)
+ for i in range(nb):
+ l = len_patterns[i]
+ j = pattern_positions[i]
+ sequences[i, j] = self.start_pattern_token
+ sequences[i, j + l + 2] = self.end_pattern_token
+ sequences[i, k] = self.start_result_token
+ sequences[i, k + l + 2] = self.end_result_token
+ sequences[i, k + 1 : k + 2 + l] = sequences[i, j + 1 : j + 2 + l]
+
+ j = torch.arange(self.len_total)[None, :]
+ ar_mask = (j > k).long() * (j <= k + 1 + len_patterns[:, None]).long()
+
+ return sequences, ar_mask
+
+ def seq2str(self, seq):
+ return "".join(self.token_string[x.item()] for x in seq)
+
+
class ProblemTwoTargets(Problem):
def __init__(self, len_total=10, len_targets=3):
assert len_targets >= 3
return "".join(self.id2char[x.item()] for x in seq)
+####################
+
+
+class ProblemMixing(Problem):
+ def __init__(
+ self, height=4, width=4, nb_time_steps=9, hard=False, random_start=True
+ ):
+ self.height = height
+ self.width = width
+ self.nb_time_steps = nb_time_steps
+ self.hard = hard
+ self.random_start = random_start
+
+ def start_random(self, nb):
+ y = torch.arange(self.height * self.width).reshape(1, -1).expand(nb, -1)
+
+ if self.random_start:
+ i = (
+ torch.arange(self.height)
+ .reshape(1, -1, 1)
+ .expand(nb, self.height, self.width)
+ )
+ j = (
+ torch.arange(self.width)
+ .reshape(1, 1, -1)
+ .expand(nb, self.height, self.width)
+ )
+
+ ri = torch.randint(self.height, (nb,)).reshape(nb, 1, 1)
+ rj = torch.randint(self.width, (nb,)).reshape(nb, 1, 1)
+
+ m = 1 - torch.logical_or(i == ri, j == rj).long().flatten(1)
+
+ y = y * m + self.height * self.width * (1 - m)
+
+ y = y.reshape(nb, self.height, self.width)
+
+ return y
+
+ def start_error(self, x):
+ if self.random_start:
+ i = (
+ torch.arange(self.height, device=x.device)
+ .reshape(1, -1, 1)
+ .expand_as(x)
+ )
+ j = torch.arange(self.width, device=x.device).reshape(1, 1, -1).expand_as(x)
+
+ ri = (
+ (x == self.height * self.width)
+ .long()
+ .sum(dim=-1)
+ .argmax(-1)
+ .view(-1, 1, 1)
+ )
+ rj = (
+ (x == self.height * self.width)
+ .long()
+ .sum(dim=-2)
+ .argmax(-1)
+ .view(-1, 1, 1)
+ )
+
+ m = 1 - torch.logical_or(i == ri, j == rj).long().flatten(1)
+ else:
+ m = 1
+
+ x = x.flatten(1)
+ u = torch.arange(self.height * self.width, device=x.device).reshape(1, -1)
+
+ d = (x - (m * u + (1 - m) * self.height * self.width)).abs().sum(-1)
+
+ 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}" if x < self.height * self.width else "**"
+ for x in s
+ ]
+ )
+ for s in r.split(self.width)
+ ]
+ )
+ for r in seq.split(self.height * self.width)
+ ]
+ )
+
+
+####################
+
if __name__ == "__main__":
- p = ProblemDegradation(hard=False)
+ p = ProblemMixing(height=3, width=3, random_start=False)
+
s, m = p.generate_sequences(10000)
- print(p.seq2str(s[0]))
+ for x in s[:5]:
+ print(p.seq2str(x))
print(p.compute_nb_correct(None, None, s))
projects / picoclvr.git / blobdiff