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 torch, torchvision
10 ######################################################################
12 v_empty, v_wall, v_start, v_goal, v_path = 0, 1, 2, 3, 4
15 def create_maze(h=11, w=17, nb_walls=8):
16 assert h % 2 == 1 and w % 2 == 1
23 m = torch.zeros(h, w, dtype=torch.int64)
33 int((r[1] * h).item()),
34 int((r[2] * h).item()),
35 int((r[3] * w).item()),
37 i1, i2, j = i1 - i1 % 2, i2 - i2 % 2, j - j % 2
38 i1, i2 = min(i1, i2), max(i1, i2)
39 if i2 - i1 > 1 and i2 - i1 <= h / 2 and m[i1 : i2 + 1, j].sum() <= 1:
44 int((r[1] * h).item()),
45 int((r[2] * w).item()),
46 int((r[3] * w).item()),
48 i, j1, j2 = i - i % 2, j1 - j1 % 2, j2 - j2 % 2
49 j1, j2 = min(j1, j2), max(j1, j2)
50 if j2 - j1 > 1 and j2 - j1 <= w / 2 and m[i, j1 : j2 + 1].sum() <= 1:
63 ######################################################################
66 def compute_distance(walls, goal_i, goal_j):
67 max_length = walls.numel()
68 dist = torch.full_like(walls, max_length)
70 dist[goal_i, goal_j] = 0
71 pred_dist = torch.empty_like(dist)
78 dist[None, 1:-1, 0:-2],
81 dist[None, 0:-2, 1:-1],
88 dist[1:-1, 1:-1] = torch.min(dist[1:-1, 1:-1], d)
89 dist = walls * max_length + (1 - walls) * dist
91 if dist.equal(pred_dist):
92 return dist * (1 - walls)
95 ######################################################################
98 def compute_policy(walls, goal_i, goal_j):
99 distance = compute_distance(walls, goal_i, goal_j)
100 distance = distance + walls.numel() * walls
102 value = distance.new_full((4,) + distance.size(), walls.numel())
103 value[0, :, 1:] = distance[:, :-1] # <
104 value[1, :, :-1] = distance[:, 1:] # >
105 value[2, 1:, :] = distance[:-1, :] # ^
106 value[3, :-1, :] = distance[1:, :] # v
108 proba = (value.min(dim=0)[0][None] == value).float()
109 proba = proba / proba.sum(dim=0)[None]
110 proba = proba * (1 - walls) + walls.float() / 4
115 def stationary_densities(mazes, policies):
116 policies = policies * (mazes != v_goal)[:, None]
117 start = (mazes == v_start).nonzero(as_tuple=True)
118 probas = mazes.new_zeros(mazes.size(), dtype=torch.float32)
119 pred_probas = probas.clone()
122 while not pred_probas.equal(probas):
123 pred_probas.copy_(probas)
125 probas[:, 1:, :] += pred_probas[:, :-1, :] * policies[:, 3, :-1, :]
126 probas[:, :-1, :] += pred_probas[:, 1:, :] * policies[:, 2, 1:, :]
127 probas[:, :, 1:] += pred_probas[:, :, :-1] * policies[:, 1, :, :-1]
128 probas[:, :, :-1] += pred_probas[:, :, 1:] * policies[:, 0, :, 1:]
134 ######################################################################
137 def mark_path(walls, i, j, goal_i, goal_j, policy):
138 action = torch.distributions.categorical.Categorical(
139 policy.permute(1, 2, 0)
141 n, nmax = 0, walls.numel()
142 while i != goal_i or j != goal_j:
143 di, dj = [(0, -1), (0, 1), (-1, 0), (1, 0)][action[i, j]]
144 i, j = i + di, j + dj
145 assert walls[i, j] == 0
151 def path_optimality(ref_paths, paths):
152 return (ref_paths == v_path).long().flatten(1).sum(1) == (
154 ).long().flatten(1).sum(1)
157 def path_correctness(mazes, paths):
158 still_ok = (mazes - (paths * (paths != v_path))).view(mazes.size(0), -1).abs().sum(
161 reached = still_ok.new_zeros(still_ok.size())
162 current, pred_current = paths.clone(), paths.new_zeros(paths.size())
163 goal = (mazes == v_goal).long()
164 while not pred_current.equal(current):
165 pred_current.copy_(current)
166 u = (current == v_start).long()
168 u[:, 2:, 1:-1] + u[:, 0:-2, 1:-1] + u[:, 1:-1, 2:] + u[:, 1:-1, 0:-2] > 0
171 reached += ((goal[:, 1:-1, 1:-1] * possible_next).sum((1, 2)) == 1) * (
172 (current == v_path).sum((1, 2)) == 0
174 current[:, 1:-1, 1:-1] = (1 - u) * current[:, 1:-1, 1:-1] + (
176 ) * (possible_next * (current[:, 1:-1, 1:-1] == v_path))
177 still_ok *= (current == v_start).sum((1, 2)) <= 1
179 return still_ok * reached
182 ######################################################################
185 def create_maze_data(
186 nb, height=11, width=17, nb_walls=8, dist_min=10, progress_bar=lambda x: x
188 mazes = torch.empty(nb, height, width, dtype=torch.int64)
189 paths = torch.empty(nb, height, width, dtype=torch.int64)
190 policies = torch.empty(nb, 4, height, width)
192 for n in progress_bar(range(nb)):
193 maze = create_maze(height, width, nb_walls)
194 i = (maze == v_empty).nonzero()
196 start, goal = i[torch.randperm(i.size(0))[:2]]
197 if (start - goal).abs().sum() >= dist_min:
199 start_i, start_j, goal_i, goal_j = start[0], start[1], goal[0], goal[1]
201 policy = compute_policy(maze, goal_i, goal_j)
203 mark_path(path, start_i, start_j, goal_i, goal_j, policy)
204 maze[start_i, start_j] = v_start
205 maze[goal_i, goal_j] = v_goal
206 path[start_i, start_j] = v_start
207 path[goal_i, goal_j] = v_goal
213 return mazes, paths, policies
216 ######################################################################
223 predicted_paths=None,
227 colors = torch.tensor(
229 [255, 255, 255], # empty
232 [127, 127, 255], # goal
240 colors[mazes.reshape(-1)].reshape(mazes.size() + (-1,)).permute(0, 3, 1, 2)
243 imgs = c_mazes.unsqueeze(1)
245 if target_paths is not None:
246 target_paths = target_paths.cpu()
249 colors[target_paths.reshape(-1)]
250 .reshape(target_paths.size() + (-1,))
254 imgs = torch.cat((imgs, c_target_paths.unsqueeze(1)), 1)
256 if predicted_paths is not None:
257 predicted_paths = predicted_paths.cpu()
258 c_predicted_paths = (
259 colors[predicted_paths.reshape(-1)]
260 .reshape(predicted_paths.size() + (-1,))
263 imgs = torch.cat((imgs, c_predicted_paths.unsqueeze(1)), 1)
265 img = torch.tensor([255, 255, 0]).view(1, -1, 1, 1)
268 if path_optimal is not None:
269 path_optimal = path_optimal.cpu().long().view(-1, 1, 1, 1)
271 img * (1 - path_optimal)
272 + torch.tensor([0, 255, 0]).view(1, -1, 1, 1) * path_optimal
275 if path_correct is not None:
276 path_correct = path_correct.cpu().long().view(-1, 1, 1, 1)
277 img = img * path_correct + torch.tensor([255, 0, 0]).view(1, -1, 1, 1) * (
282 -1, -1, imgs.size(3) + 2, 1 + imgs.size(1) * (1 + imgs.size(4))
285 print(f"{img.size()=} {imgs.size()=}")
287 for k in range(imgs.size(1)):
291 1 : 1 + imgs.size(3),
292 1 + k * (1 + imgs.size(4)) : 1 + k * (1 + imgs.size(4)) + imgs.size(4),
295 img = img.float() / 255.0
297 torchvision.utils.save_image(img, name, nrow=4, padding=1, pad_value=224.0 / 256)
300 ######################################################################
302 if __name__ == "__main__":
303 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
304 mazes, paths, policies = create_maze_data(8)
305 mazes, paths = mazes.to(device), paths.to(device)
306 save_image("test.png", mazes=mazes, target_paths=paths, predicted_paths=paths)
307 print(path_correctness(mazes, paths))
309 ######################################################################