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         result = []
  56         N = 100
  57         for _ in tqdm.tqdm(
  58             range(0, nb + N, N), dynamic_ncols=True, desc="world generation"
  59         ):
  60             result.append(self.generate_frame_sequences_hard(100))
  61         return torch.cat(result, dim=0)[:nb]
  62
  63     def generate_frame_sequences_hard(self, nb):
  64         frame_sequences = []
  65
  66         result = torch.full(
  67             (nb * 4, self.nb_iterations * self.speed, self.height, self.width),
  68             self.token_empty,
  69         )
  70
  71         for n in range(result.size(0)):
  72             while True:
  73                 i = torch.randint(self.height, (1,))
  74                 j = torch.randint(self.width, (1,))
  75                 v = torch.randint(2, (2,))
  76                 vi = v[0] * (v[1] * 2 - 1)
  77                 vj = (1 - v[0]) * (v[1] * 2 - 1)
  78                 while True:
  79                     if i < 0 or i >= self.height or j < 0 or j >= self.width:
  80                         break
  81                     o = 0
  82                     if i > 0:
  83                         o += (result[n, 0, i - 1, j] == self.token_conductor).long()
  84                     if i < self.height - 1:
  85                         o += (result[n, 0, i + 1, j] == self.token_conductor).long()
  86                     if j > 0:
  87                         o += (result[n, 0, i, j - 1] == self.token_conductor).long()
  88                     if j < self.width - 1:
  89                         o += (result[n, 0, i, j + 1] == self.token_conductor).long()
  90                     if o > 1:
  91                         break
  92                     result[n, 0, i, j] = self.token_conductor
  93                     i += vi
  94                     j += vj
  95                 if (
  96                     result[n, 0] == self.token_conductor
  97                 ).long().sum() > self.width and torch.rand(1) < 0.5:
  98                     break
  99
 100             while True:
 101                 for _ in range(self.height * self.width):
 102                     i = torch.randint(self.height, (1,))
 103                     j = torch.randint(self.width, (1,))
 104                     v = torch.randint(2, (2,))
 105                     vi = v[0] * (v[1] * 2 - 1)
 106                     vj = (1 - v[0]) * (v[1] * 2 - 1)
 107                     if (
 108                         i + vi >= 0
 109                         and i + vi < self.height
 110                         and j + vj >= 0
 111                         and j + vj < self.width
 112                         and result[n, 0, i, j] == self.token_conductor
 113                         and result[n, 0, i + vi, j + vj] == self.token_conductor
 114                     ):
 115                         result[n, 0, i, j] = self.token_head
 116                         result[n, 0, i + vi, j + vj] = self.token_tail
 117                         break
 118
 119                 if torch.rand(1) < 0.75:
 120                     break
 121
 122         weight = torch.full((1, 1, 3, 3), 1.0)
 123
 124         # mask = (torch.rand(result[:, 0].size()) < 0.01).long()
 125         # rand = torch.randint(4, mask.size())
 126         # result[:, 0] = mask * rand + (1 - mask) * result[:, 0]
 127
 128         # empty->empty
 129         # head->tail
 130         # tail->conductor
 131         # conductor->head if 1 or 2 head in the neighborhood, or remains conductor
 132
 133         for l in range(self.nb_iterations * self.speed - 1):
 134             nb_head_neighbors = (
 135                 F.conv2d(
 136                     input=(result[:, l] == self.token_head).float()[:, None, :, :],
 137                     weight=weight,
 138                     padding=1,
 139                 )
 140                 .long()
 141                 .squeeze(1)
 142             )
 143             mask_1_or_2_heads = (nb_head_neighbors == 1).long() + (
 144                 nb_head_neighbors == 2
 145             ).long()
 146             result[:, l + 1] = (
 147                 (result[:, l] == self.token_empty).long() * self.token_empty
 148                 + (result[:, l] == self.token_head).long() * self.token_tail
 149                 + (result[:, l] == self.token_tail).long() * self.token_conductor
 150                 + (result[:, l] == self.token_conductor).long()
 151                 * (
 152                     mask_1_or_2_heads * self.token_head
 153                     + (1 - mask_1_or_2_heads) * self.token_conductor
 154                 )
 155             )
 156
 157         result = result[
 158             :, torch.arange(self.nb_iterations, device=result.device) * self.speed
 159         ]
 160
 161         i = (result[:, -1] == self.token_head).flatten(1).max(dim=1).values > 0
 162         result = result[i]
 163
 164         # print(f"{result.size(0)=} {nb=}")
 165
 166         if result.size(0) < nb:
 167             # print(result.size(0))
 168             result = torch.cat(
 169                 [result, self.generate_frame_sequences(nb - result.size(0))], dim=0
 170             )
 171
 172         return result[:nb]
 173
 174     def generate_token_sequences(self, nb):
 175         frame_sequences = self.generate_frame_sequences(nb)
 176
 177         result = []
 178
 179         for frame_sequence in frame_sequences:
 180             a = []
 181             if torch.rand(1) < 0.5:
 182                 for frame in frame_sequence:
 183                     if len(a) > 0:
 184                         a.append(torch.tensor([self.token_forward]))
 185                     a.append(frame.flatten())
 186             else:
 187                 for frame in reversed(frame_sequence):
 188                     if len(a) > 0:
 189                         a.append(torch.tensor([self.token_backward]))
 190                     a.append(frame.flatten())
 191
 192             result.append(torch.cat(a, dim=0)[None, :])
 193
 194         return torch.cat(result, dim=0)
 195
 196     ######################################################################
 197
 198     def frame2img(self, x, scale=15):
 199         x = x.reshape(-1, self.height, self.width)
 200         m = torch.logical_and(x >= 0, x < 4).long()
 201
 202         x = self.colors[x * m].permute(0, 3, 1, 2)
 203         s = x.shape
 204         x = x[:, :, :, None, :, None].expand(-1, -1, -1, scale, -1, scale)
 205         x = x.reshape(s[0], s[1], s[2] * scale, s[3] * scale)
 206
 207         x[:, :, :, torch.arange(0, x.size(3), scale)] = 0
 208         x[:, :, torch.arange(0, x.size(2), scale), :] = 0
 209         x = x[:, :, 1:, 1:]
 210
 211         for n in range(m.size(0)):
 212             for i in range(m.size(1)):
 213                 for j in range(m.size(2)):
 214                     if m[n, i, j] == 0:
 215                         for k in range(2, scale - 2):
 216                             for l in [0, 1]:
 217                                 x[n, :, i * scale + k, j * scale + k - l] = 0
 218                                 x[
 219                                     n, :, i * scale + scale - 1 - k, j * scale + k - l
 220                                 ] = 0
 221
 222         return x
 223
 224     def seq2img(self, seq, scale=15):
 225         all = [
 226             self.frame2img(
 227                 seq[:, : self.height * self.width].reshape(-1, self.height, self.width),
 228                 scale,
 229             )
 230         ]
 231
 232         separator = torch.full((seq.size(0), 3, self.height * scale - 1, 1), 0)
 233
 234         t = self.height * self.width
 235
 236         while t < seq.size(1):
 237             direction_tokens = seq[:, t]
 238             t += 1
 239
 240             direction_images = self.colors[
 241                 torch.full(
 242                     (direction_tokens.size(0), self.height * scale - 1, scale), 0
 243                 )
 244             ].permute(0, 3, 1, 2)
 245
 246             for n in range(direction_tokens.size(0)):
 247                 if direction_tokens[n] == self.token_forward:
 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 + scale // 2 - abs(k - scale // 2),
 255                             ] = 0
 256                 elif direction_tokens[n] == self.token_backward:
 257                     for k in range(scale):
 258                         for l in [0, 1]:
 259                             direction_images[
 260                                 n,
 261                                 :,
 262                                 (self.height * scale) // 2 - scale // 2 + k - l,
 263                                 3 + abs(k - scale // 2),
 264                             ] = 0
 265                 else:
 266                     for k in range(2, scale - 2):
 267                         for l in [0, 1]:
 268                             direction_images[
 269                                 n,
 270                                 :,
 271                                 (self.height * scale) // 2 - scale // 2 + k - l,
 272                                 k,
 273                             ] = 0
 274                             direction_images[
 275                                 n,
 276                                 :,
 277                                 (self.height * scale) // 2 - scale // 2 + k - l,
 278                                 scale - 1 - k,
 279                             ] = 0
 280
 281             all += [
 282                 separator,
 283                 direction_images,
 284                 separator,
 285                 self.frame2img(
 286                     seq[:, t : t + self.height * self.width].reshape(
 287                         -1, self.height, self.width
 288                     ),
 289                     scale,
 290                 ),
 291             ]
 292
 293             t += self.height * self.width
 294
 295         return torch.cat(all, dim=3)
 296
 297     def seq2str(self, seq):
 298         result = []
 299         for s in seq:
 300             result.append("".join([self.token2char[v] for v in s]))
 301         return result
 302
 303     def save_image(self, input, result_dir, filename):
 304         img = self.seq2img(input.to("cpu"))
 305         image_name = os.path.join(result_dir, filename)
 306         torchvision.utils.save_image(img.float() / 255.0, image_name, nrow=6, padding=4)
 307
 308     def save_quizzes(self, input, result_dir, filename_prefix):
 309         self.save_image(input, result_dir, filename_prefix + ".png")
 310
 311
 312 ######################################################################
 313
 314 if __name__ == "__main__":
 315     import time
 316
 317     wireworld = Wireworld(height=8, width=10, nb_iterations=2, speed=5)
 318
 319     start_time = time.perf_counter()
 320     frame_sequences = wireworld.generate_frame_sequences(nb=96)
 321     delay = time.perf_counter() - start_time
 322     print(f"{frame_sequences.size(0)/delay:02f} seq/s")
 323
 324     # print(wireworld.seq2str(seq[:4]))
 325
 326     # for t in range(frame_sequences.size(1)):
 327     # img = wireworld.seq2img(frame_sequences[:, t])
 328     # torchvision.utils.save_image(
 329     # img.float() / 255.0,
 330     # f"/tmp/frame_{t:03d}.png",
 331     # nrow=8,
 332     # padding=6,
 333     # pad_value=0,
 334     # )
 335
 336     # m = (torch.rand(seq.size()) < 0.05).long()
 337     # seq = (1 - m) * seq + m * 23
 338
 339     token_sequences = wireworld.generate_token_sequences(32)
 340     wireworld.save_quizzes(token_sequences, "/tmp", "seq")
 341     # img = wireworld.seq2img(frame_sequences[:60])
 342
 343     # torchvision.utils.save_image(
 344     # img.float() / 255.0, "/tmp/world.png", nrow=6, padding=10, pad_value=0.1
 345     # )