X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=maze.py;h=6c3fe94fe189159a4a450fb90c33f02d73cb6e92;hb=7143cc544d2b0af03150d9ee05f3cf21319c693b;hp=f4a4840c352d26406e9bd80077ef4cf407f9bc95;hpb=d44d0605fed828b8cea08c8e1c5bda7e4528ea97;p=beaver.git

diff --git a/maze.py b/maze.py
index f4a4840..6c3fe94 100755
--- a/maze.py
+++ b/maze.py
@@ -113,18 +113,16 @@ def compute_policy(walls, i, j):
 ######################################################################
 
 
-def mark_path(walls, i, j, goal_i, goal_j):
-    policy = compute_policy(walls, goal_i, goal_j)
+def mark_path(walls, i, j, goal_i, goal_j, policy):
     action = torch.distributions.categorical.Categorical(
         policy.permute(1, 2, 0)
     ).sample()
-    walls[i, j] = 4
     n, nmax = 0, walls.numel()
     while i != goal_i or j != goal_j:
         di, dj = [(0, -1), (0, 1), (-1, 0), (1, 0)][action[i, j]]
         i, j = i + di, j + dj
         assert walls[i, j] == 0
-        walls[i, j] = 4
+        walls[i, j] = v_path
         n += 1
         assert n < nmax
 
@@ -160,32 +158,36 @@ def create_maze_data(
 ):
     mazes = torch.empty(nb, height, width, dtype=torch.int64)
     paths = torch.empty(nb, height, width, dtype=torch.int64)
+    policies = torch.empty(nb, 4, height, width)
 
     for n in progress_bar(range(nb)):
         maze = create_maze(height, width, nb_walls)
-        i = (1 - maze).nonzero()
+        i = (maze == v_empty).nonzero()
         while True:
             start, goal = i[torch.randperm(i.size(0))[:2]]
             if (start - goal).abs().sum() >= dist_min:
                 break
+        start_i, start_j, goal_i, goal_j = start[0], start[1], goal[0], goal[1]
 
+        policy = compute_policy(maze, goal_i, goal_j)
         path = maze.clone()
-        mark_path(path, start[0], start[1], goal[0], goal[1])
-        maze[start[0], start[1]] = v_start
-        maze[goal[0], goal[1]] = v_goal
-        path[start[0], start[1]] = v_start
-        path[goal[0], goal[1]] = v_goal
+        mark_path(path, start_i, start_j, goal_i, goal_j, policy)
+        maze[start_i, start_j] = v_start
+        maze[goal_i, goal_j] = v_goal
+        path[start_i, start_j] = v_start
+        path[goal_i, goal_j] = v_goal
 
         mazes[n] = maze
         paths[n] = path
+        policies[n] = policy
 
-    return mazes, paths
+    return mazes, paths, policies
 
 
 ######################################################################
 
 
-def save_image(name, mazes, target_paths, predicted_paths=None):
+def save_image(name, mazes, target_paths, predicted_paths=None, path_correct=None):
     mazes, target_paths = mazes.cpu(), target_paths.cpu()
 
     colors = torch.tensor(
@@ -204,7 +206,7 @@ def save_image(name, mazes, target_paths, predicted_paths=None):
         .reshape(target_paths.size() + (-1,))
         .permute(0, 3, 1, 2)
     )
-    img = torch.cat((mazes.unsqueeze(1), target_paths.unsqueeze(1)), 1)
+    imgs = torch.cat((mazes.unsqueeze(1), target_paths.unsqueeze(1)), 1)
 
     if predicted_paths is not None:
         predicted_paths = predicted_paths.cpu()
@@ -213,17 +215,34 @@ def save_image(name, mazes, target_paths, predicted_paths=None):
             .reshape(predicted_paths.size() + (-1,))
             .permute(0, 3, 1, 2)
         )
-        img = torch.cat((img, predicted_paths.unsqueeze(1)), 1)
-
-    img = img.reshape((-1,) + img.size()[2:]).float() / 255.0
-
-    torchvision.utils.save_image(img, name, padding=1, pad_value=0.85, nrow=6)
+        imgs = torch.cat((imgs, predicted_paths.unsqueeze(1)), 1)
+
+    # NxKxCxHxW
+    if path_correct is None:
+        path_correct = torch.zeros(imgs.size(0)) <= 1
+    path_correct = path_correct.cpu().long().view(-1, 1, 1, 1)
+    img = torch.tensor([224, 224, 224]).view(1, -1, 1, 1) * path_correct + torch.tensor(
+        [255, 0, 0]
+    ).view(1, -1, 1, 1) * (1 - path_correct)
+    img = img.expand(
+        -1, -1, imgs.size(3) + 2, 1 + imgs.size(1) * (1 + imgs.size(4))
+    ).clone()
+    for k in range(imgs.size(1)):
+        img[
+            :,
+            :,
+            1 : 1 + imgs.size(3),
+            1 + k * (1 + imgs.size(4)) : 1 + k * (1 + imgs.size(4)) + imgs.size(4),
+        ] = imgs[:, k]
+
+    img = img.float() / 255.0
+
+    torchvision.utils.save_image(img, name, nrow=4, padding=1, pad_value=224.0 / 256)
 
 
 ######################################################################
 
 if __name__ == "__main__":
-
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
     mazes, paths = create_maze_data(8)
     mazes, paths = mazes.to(device), paths.to(device)