X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=maze.py;h=d5662f0500e24dad299210a69efed402e3db2339;hb=HEAD;hp=81afcd94b7e12eedb0721887b6861de4bc7982bf;hpb=c921b95d0ea5b94a893447fbd4792e5047ba6e99;p=picoclvr.git

diff --git a/maze.py b/maze.py
index 81afcd9..d5662f0 100755
--- a/maze.py
+++ b/maze.py
@@ -13,11 +13,13 @@ v_empty, v_wall, v_start, v_goal, v_path = 0, 1, 2, 3, 4
 
 
 def create_maze(h=11, w=17, nb_walls=8):
-    a, k = 0, 0
+    assert h % 2 == 1 and w % 2 == 1
 
-    while k < nb_walls:
+    nb_attempts, nb_added_walls = 0, 0
+
+    while nb_added_walls < nb_walls:
         while True:
-            if a == 0:
+            if nb_attempts == 0:
                 m = torch.zeros(h, w, dtype=torch.int64)
                 m[0, :] = 1
                 m[-1, :] = 1
@@ -27,6 +29,7 @@ def create_maze(h=11, w=17, nb_walls=8):
             r = torch.rand(4)
 
             if r[0] <= 0.5:
+                # Add a vertical wall
                 i1, i2, j = (
                     int((r[1] * h).item()),
                     int((r[2] * h).item()),
@@ -34,10 +37,14 @@ def create_maze(h=11, w=17, nb_walls=8):
                 )
                 i1, i2, j = i1 - i1 % 2, i2 - i2 % 2, j - j % 2
                 i1, i2 = min(i1, i2), max(i1, i2)
+
+                # If this wall does not hit another one, add it
                 if i2 - i1 > 1 and i2 - i1 <= h / 2 and m[i1 : i2 + 1, j].sum() <= 1:
                     m[i1 : i2 + 1, j] = 1
                     break
+
             else:
+                # Add an horizontal wall
                 i, j1, j2 = (
                     int((r[1] * h).item()),
                     int((r[2] * w).item()),
@@ -45,15 +52,18 @@ def create_maze(h=11, w=17, nb_walls=8):
                 )
                 i, j1, j2 = i - i % 2, j1 - j1 % 2, j2 - j2 % 2
                 j1, j2 = min(j1, j2), max(j1, j2)
+
+                # If this wall does not hit another one, add it
                 if j2 - j1 > 1 and j2 - j1 <= w / 2 and m[i, j1 : j2 + 1].sum() <= 1:
                     m[i, j1 : j2 + 1] = 1
                     break
-            a += 1
 
-            if a > 10 * nb_walls:
-                a, k = 0, 0
+            nb_attempts += 1
+
+            if nb_attempts > 10 * nb_walls:
+                nb_attempts, nb_added_walls = 0, 0
 
-        k += 1
+        nb_added_walls += 1
 
     return m
 
@@ -146,8 +156,16 @@ def mark_path(walls, i, j, goal_i, goal_j, policy):
         assert n < nmax
 
 
+def path_optimality(ref_paths, paths):
+    return (ref_paths == v_path).long().flatten(1).sum(1) == (
+        paths == v_path
+    ).long().flatten(1).sum(1)
+
+
 def path_correctness(mazes, paths):
-    still_ok = (mazes - (paths * (paths < 4))).view(mazes.size(0), -1).abs().sum(1) == 0
+    still_ok = (mazes - (paths * (paths != v_path))).view(mazes.size(0), -1).abs().sum(
+        1
+    ) == 0
     reached = still_ok.new_zeros(still_ok.size())
     current, pred_current = paths.clone(), paths.new_zeros(paths.size())
     goal = (mazes == v_goal).long()
@@ -211,9 +229,8 @@ def save_image(
     mazes,
     target_paths=None,
     predicted_paths=None,
-    score_paths=None,
-    score_truth=None,
     path_correct=None,
+    path_optimal=None,
 ):
     colors = torch.tensor(
         [
@@ -231,17 +248,6 @@ def save_image(
         colors[mazes.reshape(-1)].reshape(mazes.size() + (-1,)).permute(0, 3, 1, 2)
     )
 
-    if score_truth is not None:
-        score_truth = score_truth.cpu()
-        c_score_truth = score_truth.unsqueeze(1).expand(-1, 3, -1, -1)
-        c_score_truth = (
-            c_score_truth * colors[4].reshape(1, 3, 1, 1)
-            + (1 - c_score_truth) * colors[0].reshape(1, 3, 1, 1)
-        ).long()
-        c_mazes = (mazes.unsqueeze(1) != v_empty) * c_mazes + (
-            mazes.unsqueeze(1) == v_empty
-        ) * c_score_truth
-
     imgs = c_mazes.unsqueeze(1)
 
     if target_paths is not None:
@@ -264,28 +270,28 @@ def save_image(
         )
         imgs = torch.cat((imgs, c_predicted_paths.unsqueeze(1)), 1)
 
-    if score_paths is not None:
-        score_paths = score_paths.cpu()
-        c_score_paths = score_paths.unsqueeze(1).expand(-1, 3, -1, -1)
-        c_score_paths = (
-            c_score_paths * colors[4].reshape(1, 3, 1, 1)
-            + (1 - c_score_paths) * colors[0].reshape(1, 3, 1, 1)
-        ).long()
-        c_score_paths = c_score_paths * (mazes.unsqueeze(1) == v_empty) + c_mazes * (
-            mazes.unsqueeze(1) != v_empty
-        )
-        imgs = torch.cat((imgs, c_score_paths.unsqueeze(1)), 1)
+    img = torch.tensor([255, 255, 0]).view(1, -1, 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)
+    if path_optimal is not None:
+        path_optimal = path_optimal.cpu().long().view(-1, 1, 1, 1)
+        img = (
+            img * (1 - path_optimal)
+            + torch.tensor([0, 255, 0]).view(1, -1, 1, 1) * path_optimal
+        )
+
+    if path_correct is not None:
+        path_correct = path_correct.cpu().long().view(-1, 1, 1, 1)
+        img = img * 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()
+
+    print(f"{img.size()=} {imgs.size()=}")
+
     for k in range(imgs.size(1)):
         img[
             :,
@@ -303,9 +309,9 @@ def save_image(
 
 if __name__ == "__main__":
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-    mazes, paths = create_maze_data(8)
+    mazes, paths, policies = create_maze_data(8)
     mazes, paths = mazes.to(device), paths.to(device)
-    save_image("test.png", mazes, paths, paths)
+    save_image("test.png", mazes=mazes, target_paths=paths, predicted_paths=paths)
     print(path_correctness(mazes, paths))
 
 ######################################################################