X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=problems.py;h=d7dbc542aa14b0c77999b7b7a55ece9394e57c90;hb=b1d28a1ed672be21947509dac2f90666b65b5034;hp=22b651795b0501331c69fd24c3ea499779576e48;hpb=0d86d8ca945722438d3c85cd01b3740269ed3546;p=picoclvr.git diff --git a/problems.py b/problems.py index 22b6517..d7dbc54 100755 --- a/problems.py +++ b/problems.py @@ -110,6 +110,48 @@ class ProblemDegradation(Problem): #################### +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 @@ -285,9 +327,164 @@ class ProblemAddition(Problem): 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) - for x in s[:100]: + for x in s[:5]: print(p.seq2str(x)) print(p.compute_nb_correct(None, None, s))