X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=maze.py;h=36eef25540b5b52a9b5688c12573b1bcd03bf6f2;hb=29cd6ffe24dfbc5720efe8b123ec1973d868881a;hp=d11ab6ef177fbef75dcd354e38da45e4df4f717f;hpb=39e24a2f9076db2d512791e723e7f2dc0275d99c;p=beaver.git

diff --git a/maze.py b/maze.py
index d11ab6e..36eef25 100755
--- a/maze.py
+++ b/maze.py
@@ -61,11 +61,11 @@ def create_maze(h=11, w=17, nb_walls=8):
 ######################################################################
 
 
-def compute_distance(walls, i, j):
+def compute_distance(walls, goal_i, goal_j):
     max_length = walls.numel()
     dist = torch.full_like(walls, max_length)
 
-    dist[i, j] = 0
+    dist[goal_i, goal_j] = 0
     pred_dist = torch.empty_like(dist)
 
     while True:
@@ -93,8 +93,8 @@ def compute_distance(walls, i, j):
 ######################################################################
 
 
-def compute_policy(walls, i, j):
-    distance = compute_distance(walls, i, j)
+def compute_policy(walls, goal_i, goal_j):
+    distance = compute_distance(walls, goal_i, goal_j)
     distance = distance + walls.numel() * walls
 
     value = distance.new_full((4,) + distance.size(), walls.numel())
@@ -110,6 +110,21 @@ def compute_policy(walls, i, j):
     return proba
 
 
+def stationary_density(policy, start_i, start_j):
+    probas = policy.new_zeros(policy.size()[:-1])
+    pred_probas = probas.clone()
+    probas[start_i, start_j] = 1.0
+
+    while not pred_probas.equal(probas):
+        pred_probas.copy_(probas)
+        probas.zero_()
+        probas[1:, :] = pred_probas[:-1, :] * policy[0, :-1, :]
+        probas[:-1, :] = pred_probas[1:, :] * policy[1, 1:, :]
+        probas[:, 1:] = pred_probas[:, :-1] * policy[2, :, :-1]
+        probas[:, :-1] = pred_probas[:, 1:] * policy[3, :, 1:]
+        probas[start_i, start_j] = 1.0
+
+
 ######################################################################
 
 
@@ -158,11 +173,11 @@ 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, 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:
@@ -187,35 +202,63 @@ def create_maze_data(
 ######################################################################
 
 
-def save_image(name, mazes, target_paths, predicted_paths=None, path_correct=None):
-    mazes, target_paths = mazes.cpu(), target_paths.cpu()
-
+def save_image(
+    name,
+    mazes,
+    target_paths=None,
+    predicted_paths=None,
+    score_paths=None,
+    path_correct=None,
+):
     colors = torch.tensor(
         [
             [255, 255, 255],  # empty
             [0, 0, 0],  # wall
             [0, 255, 0],  # start
-            [0, 0, 255],  # goal
+            [127, 127, 255],  # goal
             [255, 0, 0],  # path
         ]
     )
 
-    mazes = colors[mazes.reshape(-1)].reshape(mazes.size() + (-1,)).permute(0, 3, 1, 2)
-    target_paths = (
-        colors[target_paths.reshape(-1)]
-        .reshape(target_paths.size() + (-1,))
-        .permute(0, 3, 1, 2)
+    mazes = mazes.cpu()
+
+    c_mazes = (
+        colors[mazes.reshape(-1)].reshape(mazes.size() + (-1,)).permute(0, 3, 1, 2)
     )
-    imgs = torch.cat((mazes.unsqueeze(1), target_paths.unsqueeze(1)), 1)
+
+    imgs = c_mazes.unsqueeze(1)
+
+    if target_paths is not None:
+        target_paths = target_paths.cpu()
+
+        c_target_paths = (
+            colors[target_paths.reshape(-1)]
+            .reshape(target_paths.size() + (-1,))
+            .permute(0, 3, 1, 2)
+        )
+
+        imgs = torch.cat((imgs, c_target_paths.unsqueeze(1)), 1)
 
     if predicted_paths is not None:
         predicted_paths = predicted_paths.cpu()
-        predicted_paths = (
+        c_predicted_paths = (
             colors[predicted_paths.reshape(-1)]
             .reshape(predicted_paths.size() + (-1,))
             .permute(0, 3, 1, 2)
         )
-        imgs = torch.cat((imgs, predicted_paths.unsqueeze(1)), 1)
+        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)
 
     # NxKxCxHxW
     if path_correct is None: