3 # Any copyright is dedicated to the Public Domain.
4 # https://creativecommons.org/publicdomain/zero/1.0/
6 # Written by Francois Fleuret <francois@fleuret.org>
8 import math, sys, tqdm, os
10 import torch, torchvision
13 from torch.nn import functional as F
15 ######################################################################
20 class Physics(problem.Problem):
21 colors = torch.tensor(
38 "_" + "".join([chr(ord("A") + n) for n in range(len(colors) - 1)]) + "><"
42 self, height=6, width=8, nb_objects=2, nb_walls=2, speed=1, nb_iterations=4
46 self.nb_objects = nb_objects
47 self.nb_walls = nb_walls
49 self.nb_iterations = nb_iterations
51 def direction_tokens(self):
52 return self.token_forward, self.token_backward
54 def generate_frame_sequences(self, nb):
58 (nb * 100, self.nb_iterations, self.height, self.width), self.token_empty
61 for n in range(result.size(0)):
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)
69 if i < 0 or i >= self.height or j < 0 or j >= self.width:
71 result[n, 0, i, j] = self.token_conductor
74 if torch.rand(1) < 0.5:
77 weight = torch.full((1, 1, 3, 3), 1.0)
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]
86 # conductor->head if 1 or 2 head in the neighborhood, or remains conductor
88 for l in range(self.nb_iterations - 1):
91 input=(result[:, l] == self.token_head).float()[:, None, :, :],
98 mask_1_or_2_heads = (nb_head_neighbors == 1).long() + (
99 nb_head_neighbors == 2
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()
107 mask_1_or_2_heads * self.token_head
108 + (1 - mask_1_or_2_heads) * self.token_conductor
112 i = (result[:, -1] == self.token_head).flatten(1).max(dim=1).values > 0
116 if result.size(0) < nb:
117 print(result.size(0))
119 [result, self.generate_frame_sequences(nb - result.size(0))], dim=0
124 def generate_token_sequences(self, nb):
125 frame_sequences = self.generate_frame_sequences(nb)
129 for frame_sequence in frame_sequences:
131 if torch.rand(1) < 0.5:
132 for frame in frame_sequence:
134 a.append(torch.tensor([self.token_forward]))
135 a.append(frame.flatten())
137 for frame in reversed(frame_sequence):
139 a.append(torch.tensor([self.token_backward]))
140 a.append(frame.flatten())
142 result.append(torch.cat(a, dim=0)[None, :])
144 return torch.cat(result, dim=0)
146 ######################################################################
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()
152 x = self.colors[x * m].permute(0, 3, 1, 2)
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)
157 x[:, :, :, torch.arange(0, x.size(3), scale)] = 0
158 x[:, :, torch.arange(0, x.size(2), scale), :] = 0
161 for n in range(m.size(0)):
162 for i in range(m.size(1)):
163 for j in range(m.size(2)):
165 for k in range(2, scale - 2):
167 x[n, :, i * scale + k, j * scale + k - l] = 0
169 n, :, i * scale + scale - 1 - k, j * scale + k - l
174 def seq2img(self, seq, scale=15):
177 seq[:, : self.height * self.width].reshape(-1, self.height, self.width),
182 separator = torch.full((seq.size(0), 3, self.height * scale - 1, 1), 0)
184 t = self.height * self.width
186 while t < seq.size(1):
187 direction_tokens = seq[:, t]
190 direction_images = self.colors[
192 (direction_tokens.size(0), self.height * scale - 1, scale), 0
194 ].permute(0, 3, 1, 2)
196 for n in range(direction_tokens.size(0)):
197 if direction_tokens[n] == self.token_forward:
198 for k in range(scale):
203 (self.height * scale) // 2 - scale // 2 + k - l,
204 3 + scale // 2 - abs(k - scale // 2),
206 elif direction_tokens[n] == self.token_backward:
207 for k in range(scale):
212 (self.height * scale) // 2 - scale // 2 + k - l,
213 3 + abs(k - scale // 2),
216 for k in range(2, scale - 2):
221 (self.height * scale) // 2 - scale // 2 + k - l,
227 (self.height * scale) // 2 - scale // 2 + k - l,
236 seq[:, t : t + self.height * self.width].reshape(
237 -1, self.height, self.width
243 t += self.height * self.width
245 return torch.cat(all, dim=3)
247 def seq2str(self, seq):
250 result.append("".join([self.token2char[v] for v in s]))
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)
258 def save_quizzes(self, input, result_dir, filename_prefix):
259 self.save_image(input, result_dir, filename_prefix + ".png")
262 ######################################################################
264 if __name__ == "__main__":
267 sky = Physics(height=10, width=15, speed=1, nb_iterations=100)
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")
274 # print(sky.seq2str(seq[:4]))
276 for t in range(frame_sequences.size(1)):
277 img = sky.seq2img(frame_sequences[:, t])
278 torchvision.utils.save_image(
280 f"/tmp/frame_{t:03d}.png",
286 # m = (torch.rand(seq.size()) < 0.05).long()
287 # seq = (1 - m) * seq + m * 23
289 # img = sky.seq2img(frame_sequences[:60])
291 # torchvision.utils.save_image(
292 # img.float() / 255.0, "/tmp/world.png", nrow=6, padding=10, pad_value=0.1