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 Physics(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 * 100, self.nb_iterations, self.height, self.width), self.token_empty
  59         )
  60
  61         for n in range(result.size(0)):
  62             while True:
  63                 i = torch.randint(self.height, (1,))
  64                 j = torch.randint(self.width, (1,))
  65                 v = torch.randint(2, (2,))
  66                 vi = v[0] * (v[1] * 2 - 1)
  67                 vj = (1 - v[0]) * (v[1] * 2 - 1)
  68                 while True:
  69                     if i < 0 or i >= self.height or j < 0 or j >= self.width:
  70                         break
  71                     result[n, 0, i, j] = self.token_conductor
  72                     i += vi
  73                     j += vj
  74                 if torch.rand(1) < 0.5:
  75                     break
  76
  77         weight = torch.full((1, 1, 3, 3), 1.0)
  78
  79         mask = (torch.rand(result[:, 0].size()) < 0.01).long()
  80         rand = torch.randint(4, mask.size())
  81         result[:, 0] = mask * rand + (1 - mask) * result[:, 0]
  82
  83         # empty->empty
  84         # head->tail
  85         # tail->conductor
  86         # conductor->head if 1 or 2 head in the neighborhood, or remains conductor
  87
  88         for l in range(self.nb_iterations - 1):
  89             nb_head_neighbors = (
  90                 F.conv2d(
  91                     input=(result[:, l] == self.token_head).float()[:, None, :, :],
  92                     weight=weight,
  93                     padding=1,
  94                 )
  95                 .long()
  96                 .squeeze(1)
  97             )
  98             mask_1_or_2_heads = (nb_head_neighbors == 1).long() + (
  99                 nb_head_neighbors == 2
 100             ).long()
 101             result[:, l + 1] = (
 102                 (result[:, l] == self.token_empty).long() * self.token_empty
 103                 + (result[:, l] == self.token_head).long() * self.token_tail
 104                 + (result[:, l] == self.token_tail).long() * self.token_conductor
 105                 + (result[:, l] == self.token_conductor).long()
 106                 * (
 107                     mask_1_or_2_heads * self.token_head
 108                     + (1 - mask_1_or_2_heads) * self.token_conductor
 109                 )
 110             )
 111
 112         i = (result[:, -1] == self.token_head).flatten(1).max(dim=1).values > 0
 113
 114         result = result[i]
 115
 116         if result.size(0) < nb:
 117             print(result.size(0))
 118             result = torch.cat(
 119                 [result, self.generate_frame_sequences(nb - result.size(0))], dim=0
 120             )
 121
 122         return result
 123
 124     def generate_token_sequences(self, nb):
 125         frame_sequences = self.generate_frame_sequences(nb)
 126
 127         result = []
 128
 129         for frame_sequence in frame_sequences:
 130             a = []
 131             if torch.rand(1) < 0.5:
 132                 for frame in frame_sequence:
 133                     if len(a) > 0:
 134                         a.append(torch.tensor([self.token_forward]))
 135                     a.append(frame.flatten())
 136             else:
 137                 for frame in reversed(frame_sequence):
 138                     if len(a) > 0:
 139                         a.append(torch.tensor([self.token_backward]))
 140                     a.append(frame.flatten())
 141
 142             result.append(torch.cat(a, dim=0)[None, :])
 143
 144         return torch.cat(result, dim=0)
 145
 146     ######################################################################
 147
 148     def frame2img(self, x, scale=15):
 149         x = x.reshape(-1, self.height, self.width)
 150         m = torch.logical_and(x >= 0, x < 4).long()
 151
 152         x = self.colors[x * m].permute(0, 3, 1, 2)
 153         s = x.shape
 154         x = x[:, :, :, None, :, None].expand(-1, -1, -1, scale, -1, scale)
 155         x = x.reshape(s[0], s[1], s[2] * scale, s[3] * scale)
 156
 157         x[:, :, :, torch.arange(0, x.size(3), scale)] = 0
 158         x[:, :, torch.arange(0, x.size(2), scale), :] = 0
 159         x = x[:, :, 1:, 1:]
 160
 161         for n in range(m.size(0)):
 162             for i in range(m.size(1)):
 163                 for j in range(m.size(2)):
 164                     if m[n, i, j] == 0:
 165                         for k in range(2, scale - 2):
 166                             for l in [0, 1]:
 167                                 x[n, :, i * scale + k, j * scale + k - l] = 0
 168                                 x[
 169                                     n, :, i * scale + scale - 1 - k, j * scale + k - l
 170                                 ] = 0
 171
 172         return x
 173
 174     def seq2img(self, seq, scale=15):
 175         all = [
 176             self.frame2img(
 177                 seq[:, : self.height * self.width].reshape(-1, self.height, self.width),
 178                 scale,
 179             )
 180         ]
 181
 182         separator = torch.full((seq.size(0), 3, self.height * scale - 1, 1), 0)
 183
 184         t = self.height * self.width
 185
 186         while t < seq.size(1):
 187             direction_tokens = seq[:, t]
 188             t += 1
 189
 190             direction_images = self.colors[
 191                 torch.full(
 192                     (direction_tokens.size(0), self.height * scale - 1, scale), 0
 193                 )
 194             ].permute(0, 3, 1, 2)
 195
 196             for n in range(direction_tokens.size(0)):
 197                 if direction_tokens[n] == self.token_forward:
 198                     for k in range(scale):
 199                         for l in [0, 1]:
 200                             direction_images[
 201                                 n,
 202                                 :,
 203                                 (self.height * scale) // 2 - scale // 2 + k - l,
 204                                 3 + scale // 2 - abs(k - scale // 2),
 205                             ] = 0
 206                 elif direction_tokens[n] == self.token_backward:
 207                     for k in range(scale):
 208                         for l in [0, 1]:
 209                             direction_images[
 210                                 n,
 211                                 :,
 212                                 (self.height * scale) // 2 - scale // 2 + k - l,
 213                                 3 + abs(k - scale // 2),
 214                             ] = 0
 215                 else:
 216                     for k in range(2, scale - 2):
 217                         for l in [0, 1]:
 218                             direction_images[
 219                                 n,
 220                                 :,
 221                                 (self.height * scale) // 2 - scale // 2 + k - l,
 222                                 k,
 223                             ] = 0
 224                             direction_images[
 225                                 n,
 226                                 :,
 227                                 (self.height * scale) // 2 - scale // 2 + k - l,
 228                                 scale - 1 - k,
 229                             ] = 0
 230
 231             all += [
 232                 separator,
 233                 direction_images,
 234                 separator,
 235                 self.frame2img(
 236                     seq[:, t : t + self.height * self.width].reshape(
 237                         -1, self.height, self.width
 238                     ),
 239                     scale,
 240                 ),
 241             ]
 242
 243             t += self.height * self.width
 244
 245         return torch.cat(all, dim=3)
 246
 247     def seq2str(self, seq):
 248         result = []
 249         for s in seq:
 250             result.append("".join([self.token2char[v] for v in s]))
 251         return result
 252
 253     def save_image(self, input, result_dir, filename):
 254         img = self.seq2img(input.to("cpu"))
 255         image_name = os.path.join(result_dir, filename)
 256         torchvision.utils.save_image(img.float() / 255.0, image_name, nrow=6, padding=4)
 257
 258     def save_quizzes(self, input, result_dir, filename_prefix):
 259         self.save_image(input, result_dir, filename_prefix + ".png")
 260
 261
 262 ######################################################################
 263
 264 if __name__ == "__main__":
 265     import time
 266
 267     sky = Physics(height=10, width=15, speed=1, nb_iterations=100)
 268
 269     start_time = time.perf_counter()
 270     frame_sequences = sky.generate_frame_sequences(nb=96)
 271     delay = time.perf_counter() - start_time
 272     print(f"{frame_sequences.size(0)/delay:02f} seq/s")
 273
 274     # print(sky.seq2str(seq[:4]))
 275
 276     for t in range(frame_sequences.size(1)):
 277         img = sky.seq2img(frame_sequences[:, t])
 278         torchvision.utils.save_image(
 279             img.float() / 255.0,
 280             f"/tmp/frame_{t:03d}.png",
 281             nrow=8,
 282             padding=6,
 283             pad_value=0,
 284         )
 285
 286     # m = (torch.rand(seq.size()) < 0.05).long()
 287     # seq = (1 - m) * seq + m * 23
 288
 289     # img = sky.seq2img(frame_sequences[:60])
 290
 291     # torchvision.utils.save_image(
 292     # img.float() / 255.0, "/tmp/world.png", nrow=6, padding=10, pad_value=0.1
 293     # )