sky.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 Sky(problem.Problem):
  21     colors = torch.tensor(
  22         [
  23             [255, 255, 255],
  24             [255, 0, 0],
  25             [0, 192, 0],
  26             [0, 0, 255],
  27             [255, 192, 0],
  28             [0, 255, 255],
  29             [255, 0, 255],
  30             [192, 255, 192],
  31             [255, 192, 192],
  32             [192, 192, 255],
  33             [192, 192, 192],
  34         ]
  35     )
  36
  37     token_background = 0
  38     first_bird_token = 1
  39     nb_bird_tokens = colors.size(0) - 1
  40     token_forward = first_bird_token + nb_bird_tokens
  41     token_backward = token_forward + 1
  42
  43     token2char = (
  44         "_" + "".join([chr(ord("A") + n) for n in range(len(colors) - 1)]) + "><"
  45     )
  46
  47     def __init__(
  48         self,
  49         height=6,
  50         width=8,
  51         nb_birds=3,
  52         speed=2,
  53         nb_iterations=2,
  54         avoid_collision=True,
  55     ):
  56         self.height = height
  57         self.width = width
  58         self.nb_birds = nb_birds
  59         self.speed = speed
  60         self.nb_iterations = nb_iterations
  61         self.avoid_collision = avoid_collision
  62
  63     def direction_tokens(self):
  64         return self.token_forward, self.token_backward
  65
  66     def generate_frame_sequences(self, nb):
  67         frame_sequences = []
  68
  69         for _ in tqdm.tqdm(range(nb), dynamic_ncols=True, desc="world generation"):
  70             i, j, vi, vj = (
  71                 torch.empty(self.nb_birds, dtype=torch.int64),
  72                 torch.empty(self.nb_birds, dtype=torch.int64),
  73                 torch.empty(self.nb_birds, dtype=torch.int64),
  74                 torch.empty(self.nb_birds, dtype=torch.int64),
  75             )
  76
  77             def collision_okay():
  78                 if not self.avoid_collision:
  79                     return True
  80
  81                 count = torch.zeros(self.height, self.width, dtype=torch.int64)
  82
  83                 for n in range(self.nb_birds):
  84                     count[i[n], j[n]] += 1
  85                     count[i[n] - vi[n], j[n]] += 1
  86                     count[i[n], j[n] - vj[n]] += 1
  87
  88                 return count.max() <= 1
  89
  90             col = (
  91                 torch.randperm(self.colors.size(0) - 1)[: self.nb_birds].sort().values
  92                 + 1
  93             )
  94
  95             while True:
  96                 while True:
  97                     for n in range(self.nb_birds):
  98                         while True:
  99                             i[n] = torch.randint(self.height, (1,))
 100                             j[n] = torch.randint(self.width, (1,))
 101                             vm = torch.randint(4, (1,))
 102                             vi[n], vj[n] = (vm % 2) * 2 - 1, (vm // 2) * 2 - 1
 103                             if (
 104                                 i[n] - vi[n] >= 0
 105                                 and i[n] - vi[n] < self.height
 106                                 and j[n] - vj[n] >= 0
 107                                 and j[n] - vj[n] < self.width
 108                             ):
 109                                 break
 110
 111                     if collision_okay():
 112                         break
 113
 114                 result = torch.zeros(
 115                     self.nb_iterations * self.speed,
 116                     self.height,
 117                     self.width,
 118                     dtype=torch.int64,
 119                 )
 120
 121                 fine = torch.empty(self.nb_iterations * self.speed)
 122
 123                 t_to_keep = (
 124                     torch.arange(self.nb_iterations, device=result.device) * self.speed
 125                 )
 126
 127                 for l in range(self.nb_iterations * self.speed):
 128                     fine[l] = collision_okay()
 129                     for n in range(self.nb_birds):
 130                         c = col[n]
 131                         result[l, i[n], j[n]] = c
 132                         result[l, i[n] - vi[n], j[n]] = c
 133                         result[l, i[n], j[n] - vj[n]] = c
 134
 135                         if (i[n] == 0 and vi[n] == -1) or (
 136                             i[n] == self.height - 1 and vi[n] == 1
 137                         ):
 138                             vi[n] = -vi[n]
 139
 140                         if (j[n] == 0 and vj[n] == -1) or (
 141                             j[n] == self.width - 1 and vj[n] == 1
 142                         ):
 143                             vj[n] = -vj[n]
 144
 145                         i[n] += vi[n]
 146                         j[n] += vj[n]
 147
 148                 result = result[t_to_keep]
 149                 fine = fine[t_to_keep]
 150
 151                 if fine[-1]:
 152                     break
 153
 154             frame_sequences.append(result)
 155
 156         return frame_sequences
 157
 158     ######################################################################
 159
 160     def generate_token_sequences(self, nb):
 161         frame_sequences = self.generate_frame_sequences(nb)
 162
 163         result = []
 164
 165         for frame_sequence in frame_sequences:
 166             a = []
 167             if torch.rand(1) < 0.5:
 168                 for frame in frame_sequence:
 169                     if len(a) > 0:
 170                         a.append(torch.tensor([self.token_forward]))
 171                     a.append(frame.flatten())
 172             else:
 173                 for frame in reversed(frame_sequence):
 174                     if len(a) > 0:
 175                         a.append(torch.tensor([self.token_backward]))
 176                     a.append(frame.flatten())
 177
 178             result.append(torch.cat(a, dim=0)[None, :])
 179
 180         return torch.cat(result, dim=0)
 181
 182     ######################################################################
 183
 184     def frame2img(self, x, scale=15):
 185         x = x.reshape(-1, self.height, self.width)
 186         m = torch.logical_and(
 187             x >= 0, x < self.first_bird_token + self.nb_bird_tokens
 188         ).long()
 189         x = self.colors[x * m].permute(0, 3, 1, 2)
 190         s = x.shape
 191         x = x[:, :, :, None, :, None].expand(-1, -1, -1, scale, -1, scale)
 192         x = x.reshape(s[0], s[1], s[2] * scale, s[3] * scale)
 193
 194         x[:, :, :, torch.arange(0, x.size(3), scale)] = 0
 195         x[:, :, torch.arange(0, x.size(2), scale), :] = 0
 196         x = x[:, :, 1:, 1:]
 197
 198         for n in range(m.size(0)):
 199             for i in range(m.size(1)):
 200                 for j in range(m.size(2)):
 201                     if m[n, i, j] == 0:
 202                         for k in range(2, scale - 2):
 203                             for l in [0, 1]:
 204                                 x[n, :, i * scale + k, j * scale + k - l] = 0
 205                                 x[
 206                                     n, :, i * scale + scale - 1 - k, j * scale + k - l
 207                                 ] = 0
 208
 209         return x
 210
 211     def seq2img(self, seq, scale=15):
 212         all = [
 213             self.frame2img(
 214                 seq[:, : self.height * self.width].reshape(-1, self.height, self.width),
 215                 scale,
 216             )
 217         ]
 218
 219         separator = torch.full((seq.size(0), 3, self.height * scale - 1, 1), 0)
 220
 221         t = self.height * self.width
 222
 223         while t < seq.size(1):
 224             direction_tokens = seq[:, t]
 225             t += 1
 226
 227             direction_images = self.colors[
 228                 torch.full(
 229                     (direction_tokens.size(0), self.height * scale - 1, scale), 0
 230                 )
 231             ].permute(0, 3, 1, 2)
 232
 233             for n in range(direction_tokens.size(0)):
 234                 if direction_tokens[n] == self.token_forward:
 235                     for k in range(scale):
 236                         for l in [0, 1]:
 237                             direction_images[
 238                                 n,
 239                                 :,
 240                                 (self.height * scale) // 2 - scale // 2 + k - l,
 241                                 3 + scale // 2 - abs(k - scale // 2),
 242                             ] = 0
 243                 elif direction_tokens[n] == self.token_backward:
 244                     for k in range(scale):
 245                         for l in [0, 1]:
 246                             direction_images[
 247                                 n,
 248                                 :,
 249                                 (self.height * scale) // 2 - scale // 2 + k - l,
 250                                 3 + abs(k - scale // 2),
 251                             ] = 0
 252                 else:
 253                     for k in range(2, scale - 2):
 254                         for l in [0, 1]:
 255                             direction_images[
 256                                 n,
 257                                 :,
 258                                 (self.height * scale) // 2 - scale // 2 + k - l,
 259                                 k,
 260                             ] = 0
 261                             direction_images[
 262                                 n,
 263                                 :,
 264                                 (self.height * scale) // 2 - scale // 2 + k - l,
 265                                 scale - 1 - k,
 266                             ] = 0
 267
 268             all += [
 269                 separator,
 270                 direction_images,
 271                 separator,
 272                 self.frame2img(
 273                     seq[:, t : t + self.height * self.width].reshape(
 274                         -1, self.height, self.width
 275                     ),
 276                     scale,
 277                 ),
 278             ]
 279
 280             t += self.height * self.width
 281
 282         return torch.cat(all, dim=3)
 283
 284     def seq2str(self, seq):
 285         result = []
 286         for s in seq:
 287             result.append("".join([self.token2char[v] for v in s]))
 288         return result
 289
 290     def save_image(self, input, result_dir, filename):
 291         img = self.seq2img(input.to("cpu"))
 292         image_name = os.path.join(result_dir, filename)
 293         torchvision.utils.save_image(img.float() / 255.0, image_name, nrow=6, padding=4)
 294
 295     def save_quizzes(self, input, result_dir, filename_prefix):
 296         self.save_image(input, result_dir, filename_prefix + ".png")
 297
 298
 299 ######################################################################
 300
 301 if __name__ == "__main__":
 302     import time
 303
 304     sky = Sky(height=6, width=8, speed=4, nb_iterations=2)
 305
 306     start_time = time.perf_counter()
 307     token_sequences = sky.generate_token_sequences(nb=64)
 308     delay = time.perf_counter() - start_time
 309     print(f"{token_sequences.size(0)/delay:02f} seq/s")
 310
 311     # print(sky.seq2str(seq[:4]))
 312
 313     # for t in range(len(it[0])):
 314     # img = torch.cat([sky.frame2img(f[t]) for f in it], dim=0)
 315     # torchvision.utils.save_image(
 316     # img.float() / 255.0,
 317     # f"/tmp/frame_{t:03d}.png",
 318     # nrow=8,
 319     # padding=6,
 320     # pad_value=0,
 321     # )
 322
 323     # m = (torch.rand(seq.size()) < 0.05).long()
 324     # seq = (1 - m) * seq + m * 23
 325
 326     # print(seq.size())
 327     img = sky.seq2img(token_sequences)
 328     # print(img.size())
 329
 330     torchvision.utils.save_image(
 331         img.float() / 255.0, "/tmp/world.png", nrow=6, padding=6, pad_value=0
 332     )