Update.

[picoclvr.git] / maze.py

diff --git a/maze.py b/maze.py
index 81afcd9..8ac9fce 100755 (executable)
--- a/maze.py
+++ b/maze.py
@@ -13,6 +13,8 @@ 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):
+    assert h % 2 == 1 and w % 2 == 1
+
     a, k = 0, 0
 
     while k < nb_walls:
@@ -146,8 +148,16 @@ 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 +221,8 @@ 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 +240,6 @@ 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 +262,28 @@ 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 +301,9 @@ 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))
 
 ######################################################################