Update.

diff --git a/escape.py b/escape.py
index 1c1bc20..43843f0 100755 (executable)
--- a/escape.py
+++ b/escape.py
@@ -25,7 +25,7 @@ nb_codes = first_lookahead_rewards_code + nb_lookahead_rewards_codes
 ######################################################################
 
 
-def generate_episodes(nb, height=6, width=6, T=10):
+def generate_episodes(nb, height=6, width=6, T=10, nb_walls=3):
     rnd = torch.rand(nb, height, width)
     rnd[:, 0, :] = 0
     rnd[:, -1, :] = 0
@@ -33,11 +33,12 @@ def generate_episodes(nb, height=6, width=6, T=10):
     rnd[:, :, -1] = 0
     wall = 0
 
-    for k in range(3):
+    for k in range(nb_walls):
         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))
 
     states = wall[:, None, :, :].expand(-1, T, -1, -1).clone()
@@ -280,8 +281,8 @@ def episodes2str(
 ######################################################################
 
 if __name__ == "__main__":
-    nb, height, width, T = 25, 5, 7, 25
-    states, actions, rewards = generate_episodes(nb, height, width, T)
+    nb, height, width, T, nb_walls = 25, 5, 7, 25, 5
+    states, actions, rewards = generate_episodes(nb, height, width, T, nb_walls)
     seq = episodes2seq(states, actions, rewards, lookahead_delta=T)
     s, a, r, lr = seq2episodes(seq, height, width, lookahead=True)
     print(episodes2str(s, a, r, lookahead_rewards=lr, unicode=True, ansi_colors=True))

diff --git a/main.py b/main.py
index 2edfa14..e855b06 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -178,12 +178,14 @@ parser.add_argument("--mixing_deterministic_start", action="store_true", default
 ##############################
 # escape options
 
-parser.add_argument("--escape_height", type=int, default=4)
+parser.add_argument("--escape_height", type=int, default=5)
 
-parser.add_argument("--escape_width", type=int, default=6)
+parser.add_argument("--escape_width", type=int, default=7)
 
 parser.add_argument("--escape_T", type=int, default=25)
 
+parser.add_argument("--escape_nb_walls", type=int, default=5)
+
 ######################################################################
 
 args = parser.parse_args()
@@ -622,6 +624,7 @@ elif args.task == "escape":
         height=args.escape_height,
         width=args.escape_width,
         T=args.escape_T,
+        nb_walls=args.escape_nb_walls,
         logger=log_string,
         device=device,
     )

diff --git a/tasks.py b/tasks.py
index 829eb24..56c2b0f 100755 (executable)
--- a/tasks.py
+++ b/tasks.py
@@ -1874,6 +1874,7 @@ class Escape(Task):
         height,
         width,
         T,
+        nb_walls,
         logger=None,
         device=torch.device("cpu"),
     ):
@@ -1885,7 +1886,7 @@ class Escape(Task):
         self.width = width
 
         states, actions, rewards = escape.generate_episodes(
-            nb_train_samples + nb_test_samples, height, width, T
+            nb_train_samples + nb_test_samples, height, width, T, nb_walls
         )
         seq = escape.episodes2seq(states, actions, rewards, lookahead_delta=T)
         # seq = seq[:, seq.size(1) // 3 : 2 * seq.size(1) // 3]
@@ -1912,7 +1913,7 @@ class Escape(Task):
     def thinking_autoregression(
         self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
     ):
-        result = self.test_input[:100].clone()
+        result = self.test_input[:250].clone()
         t = torch.arange(result.size(1), device=result.device)[None, :]
 
         state_len = self.height * self.width
@@ -1936,9 +1937,9 @@ class Escape(Task):
         for u in tqdm.tqdm(
             range(it_len, result.size(1) - it_len + 1, it_len), desc="thinking"
         ):
-            # Put the lookahead reward to -1 for the current iteration,
-            # sample the next state
-            s = -1
+            # Put the lookahead reward to either 0 or -1 for the
+            # current iteration, sample the next state
+            s = -1  # (torch.rand(result.size(0), device = result.device) < 0.2).long()
             result[:, u - 1] = s + 1 + escape.first_lookahead_rewards_code
             ar_mask = (t >= u).long() * (t < u + state_len).long()
             ar(result, ar_mask)
@@ -1954,7 +1955,7 @@ class Escape(Task):
             for v in range(0, u, it_len):
                 # Extract the rewards
                 r = result[:, range(v + state_len + 1 + it_len, u + it_len - 1, it_len)]
-                r = r - escape.first_lookahead_rewards_code - 1
+                r = r - escape.first_rewards_code - 1
                 a = r.min(dim=1).values
                 b = r.max(dim=1).values
                 s = (a < 0).long() * a + (a >= 0).long() * b