wireworld.py

   1 #!/usr/bin/env python
   2
   3 # Any copyright is dedicated to the Public Domain.
   4 # https://creativecommons.org/publicdomain/zero/1.0/
   5
   6 # Written by Francois Fleuret <francois@fleuret.org>
   7
   8 import math, sys, tqdm, os
   9
  10 import torch, torchvision
  11
  12 from torch import nn
  13 from torch.nn import functional as F
  14
  15 ######################################################################
  16
  17 import problem
  18
  19
  20 class Wireworld(problem.Problem):
  21     colors = torch.tensor(
  22         [
  23             [128, 128, 128],
  24             [128, 128, 255],
  25             [255, 0, 0],
  26             [255, 255, 0],
  27         ]
  28     )
  29
  30     token_empty = 0
  31     token_head = 1
  32     token_tail = 2
  33     token_conductor = 3
  34     token_forward = 4
  35     token_backward = 5
  36
  37     token2char = (
  38         "_" + "".join([chr(ord("A") + n) for n in range(len(colors) - 1)]) + "><"
  39     )
  40
  41     def __init__(
  42         self, height=6, width=8, nb_objects=2, nb_walls=2, speed=1, nb_iterations=4
  43     ):
  44         self.height = height
  45         self.width = width
  46         self.nb_objects = nb_objects
  47         self.nb_walls = nb_walls
  48         self.speed = speed
  49         self.nb_iterations = nb_iterations
  50
  51     def direction_tokens(self):
  52         return self.token_forward, self.token_backward
  53
  54     def generate_frame_sequences(self, nb):
  55         frame_sequences = []
  56
  57         result = torch.full(
  58             (nb * 4, self.nb_iterations * self.speed, self.height, self.width),
  59             self.token_empty,
  60         )
  61
  62         for n in range(result.size(0)):
  63             while True:
  64                 i = torch.randint(self.height, (1,))
  65                 j = torch.randint(self.width, (1,))
  66                 v = torch.randint(2, (2,))
  67                 vi = v[0] * (v[1] * 2 - 1)
  68                 vj = (1 - v[0]) * (v[1] * 2 - 1)
  69                 while True:
  70                     if i < 0 or i >= self.height or j < 0 or j >= self.width:
  71                         break
  72                     o = 0
  73                     if i > 0:
  74                         o += (result[n, 0, i - 1, j] == self.token_conductor).long()
  75                     if i < self.height - 1:
  76                         o += (result[n, 0, i + 1, j] == self.token_conductor).long()
  77                     if j > 0:
  78                         o += (result[n, 0, i, j - 1] == self.token_conductor).long()
  79                     if j < self.width - 1:
  80                         o += (result[n, 0, i, j + 1] == self.token_conductor).long()
  81                     if o > 1:
  82                         break
  83                     result[n, 0, i, j] = self.token_conductor
  84                     i += vi
  85                     j += vj
  86                 if (
  87                     result[n, 0] == self.token_conductor
  88                 ).long().sum() > self.width and torch.rand(1) < 0.5:
  89                     break
  90
  91             while True:
  92                 for _ in range(self.height * self.width):
  93                     i = torch.randint(self.height, (1,))
  94                     j = torch.randint(self.width, (1,))
  95                     v = torch.randint(2, (2,))
  96                     vi = v[0] * (v[1] * 2 - 1)
  97                     vj = (1 - v[0]) * (v[1] * 2 - 1)
  98                     if (
  99                         i + vi >= 0
 100                         and i + vi < self.height
 101                         and j + vj >= 0
 102                         and j + vj < self.width
 103                         and result[n, 0, i, j] == self.token_conductor
 104                         and result[n, 0, i + vi, j + vj] == self.token_conductor
 105                     ):
 106                         result[n, 0, i, j] = self.token_head
 107                         result[n, 0, i + vi, j + vj] = self.token_tail
 108                         break
 109
 110                 if torch.rand(1) < 0.75:
 111                     break
 112
 113         weight = torch.full((1, 1, 3, 3), 1.0)
 114
 115         # mask = (torch.rand(result[:, 0].size()) < 0.01).long()
 116         # rand = torch.randint(4, mask.size())
 117         # result[:, 0] = mask * rand + (1 - mask) * result[:, 0]
 118
 119         # empty->empty
 120         # head->tail
 121         # tail->conductor
 122         # conductor->head if 1 or 2 head in the neighborhood, or remains conductor
 123
 124         for l in range(self.nb_iterations * self.speed - 1):
 125             nb_head_neighbors = (
 126                 F.conv2d(
 127                     input=(result[:, l] == self.token_head).float()[:, None, :, :],
 128                     weight=weight,
 129                     padding=1,
 130                 )
 131                 .long()
 132                 .squeeze(1)
 133             )
 134             mask_1_or_2_heads = (nb_head_neighbors == 1).long() + (
 135                 nb_head_neighbors == 2
 136             ).long()
 137             result[:, l + 1] = (
 138                 (result[:, l] == self.token_empty).long() * self.token_empty
 139                 + (result[:, l] == self.token_head).long() * self.token_tail
 140                 + (result[:, l] == self.token_tail).long() * self.token_conductor
 141                 + (result[:, l] == self.token_conductor).long()
 142                 * (
 143                     mask_1_or_2_heads * self.token_head
 144                     + (1 - mask_1_or_2_heads) * self.token_conductor
 145                 )
 146             )
 147
 148         result = result[
 149             :, torch.arange(self.nb_iterations, device=result.device) * self.speed
 150         ]
 151
 152         i = (result[:, -1] == self.token_head).flatten(1).max(dim=1).values > 0
 153         result = result[i]
 154
 155         print(f"{result.size(0)=} {nb=}")
 156
 157         if result.size(0) < nb:
 158             # print(result.size(0))
 159             result = torch.cat(
 160                 [result, self.generate_frame_sequences(nb - result.size(0))], dim=0
 161             )
 162
 163         return result[:nb]
 164
 165     def generate_token_sequences(self, nb):
 166         frame_sequences = self.generate_frame_sequences(nb)
 167
 168         result = []
 169
 170         for frame_sequence in frame_sequences:
 171             a = []
 172             if torch.rand(1) < 0.5:
 173                 for frame in frame_sequence:
 174                     if len(a) > 0:
 175                         a.append(torch.tensor([self.token_forward]))
 176                     a.append(frame.flatten())
 177             else:
 178                 for frame in reversed(frame_sequence):
 179                     if len(a) > 0:
 180                         a.append(torch.tensor([self.token_backward]))
 181                     a.append(frame.flatten())
 182
 183             result.append(torch.cat(a, dim=0)[None, :])
 184
 185         return torch.cat(result, dim=0)
 186
 187     ######################################################################
 188
 189     def frame2img(self, x, scale=15):
 190         x = x.reshape(-1, self.height, self.width)
 191         m = torch.logical_and(x >= 0, x < 4).long()
 192
 193         x = self.colors[x * m].permute(0, 3, 1, 2)
 194         s = x.shape
 195         x = x[:, :, :, None, :, None].expand(-1, -1, -1, scale, -1, scale)
 196         x = x.reshape(s[0], s[1], s[2] * scale, s[3] * scale)
 197
 198         x[:, :, :, torch.arange(0, x.size(3), scale)] = 0
 199         x[:, :, torch.arange(0, x.size(2), scale), :] = 0
 200         x = x[:, :, 1:, 1:]
 201
 202         for n in range(m.size(0)):
 203             for i in range(m.size(1)):
 204                 for j in range(m.size(2)):
 205                     if m[n, i, j] == 0:
 206                         for k in range(2, scale - 2):
 207                             for l in [0, 1]:
 208                                 x[n, :, i * scale + k, j * scale + k - l] = 0
 209                                 x[
 210                                     n, :, i * scale + scale - 1 - k, j * scale + k - l
 211                                 ] = 0
 212
 213         return x
 214
 215     def seq2img(self, seq, scale=15):
 216         all = [
 217             self.frame2img(
 218                 seq[:, : self.height * self.width].reshape(-1, self.height, self.width),
 219                 scale,
 220             )
 221         ]
 222
 223         separator = torch.full((seq.size(0), 3, self.height * scale - 1, 1), 0)
 224
 225         t = self.height * self.width
 226
 227         while t < seq.size(1):
 228             direction_tokens = seq[:, t]
 229             t += 1
 230
 231             direction_images = self.colors[
 232                 torch.full(
 233                     (direction_tokens.size(0), self.height * scale - 1, scale), 0
 234                 )
 235             ].permute(0, 3, 1, 2)
 236
 237             for n in range(direction_tokens.size(0)):
 238                 if direction_tokens[n] == self.token_forward:
 239                     for k in range(scale):
 240                         for l in [0, 1]:
 241                             direction_images[
 242                                 n,
 243                                 :,
 244                                 (self.height * scale) // 2 - scale // 2 + k - l,
 245                                 3 + scale // 2 - abs(k - scale // 2),
 246                             ] = 0
 247                 elif direction_tokens[n] == self.token_backward:
 248                     for k in range(scale):
 249                         for l in [0, 1]:
 250                             direction_images[
 251                                 n,
 252                                 :,
 253                                 (self.height * scale) // 2 - scale // 2 + k - l,
 254                                 3 + abs(k - scale // 2),
 255                             ] = 0
 256                 else:
 257                     for k in range(2, scale - 2):
 258                         for l in [0, 1]:
 259                             direction_images[
 260                                 n,
 261                                 :,
 262                                 (self.height * scale) // 2 - scale // 2 + k - l,
 263                                 k,
 264                             ] = 0
 265                             direction_images[
 266                                 n,
 267                                 :,
 268                                 (self.height * scale) // 2 - scale // 2 + k - l,
 269                                 scale - 1 - k,
 270                             ] = 0
 271
 272             all += [
 273                 separator,
 274                 direction_images,
 275                 separator,
 276                 self.frame2img(
 277                     seq[:, t : t + self.height * self.width].reshape(
 278                         -1, self.height, self.width
 279                     ),
 280                     scale,
 281                 ),
 282             ]
 283
 284             t += self.height * self.width
 285
 286         return torch.cat(all, dim=3)
 287
 288     def seq2str(self, seq):
 289         result = []
 290         for s in seq:
 291             result.append("".join([self.token2char[v] for v in s]))
 292         return result
 293
 294     def save_image(self, input, result_dir, filename):
 295         img = self.seq2img(input.to("cpu"))
 296         image_name = os.path.join(result_dir, filename)
 297         torchvision.utils.save_image(img.float() / 255.0, image_name, nrow=6, padding=4)
 298
 299     def save_quizzes(self, input, result_dir, filename_prefix):
 300         self.save_image(input, result_dir, filename_prefix + ".png")
 301
 302
 303 ######################################################################
 304
 305 if __name__ == "__main__":
 306     import time
 307
 308     wireworld = Wireworld(height=10, width=15, nb_iterations=2, speed=5)
 309
 310     start_time = time.perf_counter()
 311     frame_sequences = wireworld.generate_frame_sequences(nb=96)
 312     delay = time.perf_counter() - start_time
 313     print(f"{frame_sequences.size(0)/delay:02f} seq/s")
 314
 315     # print(wireworld.seq2str(seq[:4]))
 316
 317     # for t in range(frame_sequences.size(1)):
 318     # img = wireworld.seq2img(frame_sequences[:, t])
 319     # torchvision.utils.save_image(
 320     # img.float() / 255.0,
 321     # f"/tmp/frame_{t:03d}.png",
 322     # nrow=8,
 323     # padding=6,
 324     # pad_value=0,
 325     # )
 326
 327     # m = (torch.rand(seq.size()) < 0.05).long()
 328     # seq = (1 - m) * seq + m * 23
 329
 330     token_sequences = wireworld.generate_token_sequences(32)
 331     wireworld.save_quizzes(token_sequences, "/tmp", "seq")
 332     # img = wireworld.seq2img(frame_sequences[:60])
 333
 334     # torchvision.utils.save_image(
 335     # img.float() / 255.0, "/tmp/world.png", nrow=6, padding=10, pad_value=0.1
 336     # )