Update.

diff --git a/evasion.py b/evasion.py
new file mode 100755 (executable)
index 0000000..4efa4b3
--- /dev/null
+++ b/evasion.py
@@ -0,0 +1,117 @@
+#!/usr/bin/env python
+
+import torch
+
+from torch.nn import functional as F
+
+######################################################################
+
+
+def generate_sequence(nb, height=6, width=6, T=10):
+    rnd = torch.rand(nb, height, width)
+    rnd[:, 0, :] = 0
+    rnd[:, -1, :] = 0
+    rnd[:, :, 0] = 0
+    rnd[:, :, -1] = 0
+    wall = 0
+
+    for k in range(3):
+        wall = wall + (
+            rnd.flatten(1).argmax(dim=1)[:, None]
+            == torch.arange(rnd.flatten(1).size(1))[None, :]
+        ).long().reshape(rnd.size())
+        rnd = rnd * (1 - wall.clamp(max=1))
+
+    seq = wall[:, None, :, :].expand(-1, T, -1, -1).clone()
+
+    agent = torch.zeros(seq.size(), dtype=torch.int64)
+    agent[:, 0, 0, 0] = 1
+    agent_actions = torch.randint(5, (nb, T))
+    monster = torch.zeros(seq.size(), dtype=torch.int64)
+    monster[:, 0, -1, -1] = 1
+    monster_actions = torch.randint(5, (nb, T))
+
+    all_moves = agent.new(nb, 5, height, width)
+    for t in range(T - 1):
+        all_moves.zero_()
+        all_moves[:, 0] = agent[:, t]
+        all_moves[:, 1, 1:, :] = agent[:, t, :-1, :]
+        all_moves[:, 2, :-1, :] = agent[:, t, 1:, :]
+        all_moves[:, 3, :, 1:] = agent[:, t, :, :-1]
+        all_moves[:, 4, :, :-1] = agent[:, t, :, 1:]
+        a = F.one_hot(agent_actions[:, t], num_classes=5)[:, :, None, None]
+        after_move = (all_moves * a).sum(dim=1)
+        collision = (
+            (after_move * (1 - wall) * (1 - monster[:, t]))
+            .flatten(1)
+            .sum(dim=1)[:, None, None]
+            == 0
+        ).long()
+        agent[:, t + 1] = collision * agent[:, t] + (1 - collision) * after_move
+
+        all_moves.zero_()
+        all_moves[:, 0] = monster[:, t]
+        all_moves[:, 1, 1:, :] = monster[:, t, :-1, :]
+        all_moves[:, 2, :-1, :] = monster[:, t, 1:, :]
+        all_moves[:, 3, :, 1:] = monster[:, t, :, :-1]
+        all_moves[:, 4, :, :-1] = monster[:, t, :, 1:]
+        a = F.one_hot(monster_actions[:, t], num_classes=5)[:, :, None, None]
+        after_move = (all_moves * a).sum(dim=1)
+        collision = (
+            (after_move * (1 - wall) * (1 - agent[:, t + 1]))
+            .flatten(1)
+            .sum(dim=1)[:, None, None]
+            == 0
+        ).long()
+        monster[:, t + 1] = collision * monster[:, t] + (1 - collision) * after_move
+
+    seq += 2 * agent + 3 * monster
+
+    return seq, agent_actions
+
+
+######################################################################
+
+
+def seq2str(seq, actions=None):
+    # symbols=" #@$"
+    symbols = " █@$"
+
+    hline = ("+" + "-" * seq.size(-1)) * seq.size(1) + "+" + "\n"
+
+    result = hline
+
+    for n in range(seq.size(0)):
+        for i in range(seq.size(2)):
+            result += (
+                "|"
+                + "|".join(
+                    ["".join([symbols[v.item()] for v in row]) for row in seq[n, :, i]]
+                )
+                + "|"
+                + "\n"
+            )
+
+        result += hline
+
+        if actions is not None:
+            result += (
+                "|"
+                + "|".join(
+                    ["INESW"[a.item()] + " " * (seq.size(-1) - 1) for a in actions[n]]
+                )
+                + "|"
+                + "\n"
+            )
+
+        result += hline
+
+    return result
+
+
+######################################################################
+
+if __name__ == "__main__":
+    seq, actions = generate_sequence(40, 4, 6, T=20)
+
+    print(seq2str(seq, actions))

diff --git a/maze.py b/maze.py
index 8ac9fce..d5662f0 100755 (executable)
--- a/maze.py
+++ b/maze.py
@@ -15,11 +15,11 @@ 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
+    nb_attempts, nb_added_walls = 0, 0
 
-    while k < nb_walls:
+    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
@@ -29,6 +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()),
@@ -36,10 +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()),
@@ -47,15 +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
 

diff --git a/tasks.py b/tasks.py
index d21e264..d680951 100755 (executable)
--- a/tasks.py
+++ b/tasks.py
@@ -99,7 +99,6 @@ class TaskFromFile(Task):
         ).to("cpu")
 
         if shuffle:
-            print("SHUFFLING!")
             i = torch.randperm(input.size(0))
             input = input[i].contiguous()
             pred_mask = pred_mask[i].contiguous()

diff --git a/world.py b/world.py
index aad0bfb..d95bddb 100755 (executable)
--- a/world.py
+++ b/world.py
@@ -464,7 +464,7 @@ if __name__ == "__main__":
         frame2seq,
         seq2frame,
     ) = create_data_and_processors(
-        25000,
+        250,
         1000,
         nb_epochs=5,
         mode="first_last",