Update.

diff --git a/escape.py b/escape.py
index a3d8c85..7596bea 100755 (executable)
--- a/escape.py
+++ b/escape.py
@@ -14,7 +14,7 @@ from torch.nn import functional as F
 nb_states_codes = 5
 nb_actions_codes = 5
 nb_rewards_codes = 3
-nb_lookahead_rewards_codes = 3
+nb_lookahead_rewards_codes = 4  # stands for -1, 0, +1, and UNKNOWN
 
 first_states_code = 0
 first_actions_code = first_states_code + nb_states_codes
@@ -50,6 +50,7 @@ def code2reward(r):
 
 
 def lookahead_reward2code(r):
+    # -1, 0, +1 or 2 for UNKNOWN
     return r + 1 + first_lookahead_rewards_code
 
 
@@ -60,7 +61,7 @@ def code2lookahead_reward(r):
 ######################################################################
 
 
-def generate_episodes(nb, height=6, width=6, T=10, nb_walls=3, nb_coins=3):
+def generate_episodes(nb, height=6, width=6, T=10, nb_walls=3, nb_coins=2):
     rnd = torch.rand(nb, height, width)
     rnd[:, 0, :] = 0
     rnd[:, -1, :] = 0
@@ -195,7 +196,7 @@ def seq2str(seq):
             t >= first_lookahead_rewards_code
             and t < first_lookahead_rewards_code + nb_lookahead_rewards_codes
         ):
-            return "n.p"[t - first_lookahead_rewards_code]
+            return "n.pU"[t - first_lookahead_rewards_code]
         else:
             return "?"
 

diff --git a/tasks.py b/tasks.py
index 11879fd..57a4c39 100755 (executable)
--- a/tasks.py
+++ b/tasks.py
@@ -1898,6 +1898,13 @@ class Escape(Task):
         self.train_input = seq[:nb_train_samples].to(self.device)
         self.test_input = seq[nb_train_samples:].to(self.device)
 
+        self.state_len = self.height * self.width
+        self.index_lookahead_reward = 0
+        self.index_states = 1
+        self.index_action = self.state_len + 1
+        self.index_reward = self.state_len + 2
+        self.it_len = self.state_len + 3  # lookahead_reward / state / action / reward
+
     def batches(self, split="train", nb_to_use=-1, desc=None):
         assert split in {"train", "test"}
         input = self.train_input if split == "train" else self.test_input
@@ -1908,6 +1915,13 @@ class Escape(Task):
         for batch in tqdm.tqdm(
             input.split(self.batch_size), dynamic_ncols=True, desc=desc
         ):
+            t = torch.arange(input.size(1), device=input.device)[None, :]
+            u = torch.randint(input.size(1), (input.size(0), 1), device=input.device)
+            lr_mask = (t <= u).long() * (
+                t % self.it_len == self.index_lookahead_reward
+            ).long()
+
+            input = lr_mask * escape.lookahead_reward2code(2) + (1 - lr_mask) * input
             yield batch
 
     def vocabulary_size(self):
@@ -1919,14 +1933,7 @@ class Escape(Task):
         result = self.test_input[:250].clone()
         t = torch.arange(result.size(1), device=result.device)[None, :]
 
-        state_len = self.height * self.width
-        index_lookahead_reward = 0
-        index_states = 1
-        index_action = state_len + 1
-        index_reward = state_len + 2
-        it_len = state_len + 3  # lookahead_reward / state / action / reward
-
-        result[:, it_len:] = -1
+        result[:, self.it_len :] = -1
 
         snapshots = []
 
@@ -1953,30 +1960,31 @@ class Escape(Task):
         optimistic_bias[escape.lookahead_reward2code(1)] = math.log(1e1)
 
         for u in tqdm.tqdm(
-            range(it_len, result.size(1) - it_len + 1, it_len), desc="thinking"
+            range(self.it_len, result.size(1) - self.it_len + 1, self.it_len),
+            desc="thinking",
         ):
-            lr, _, _, _ = escape.seq2episodes(result[:, :u], self.height, self.width)
-
             # Generate the lookahead_reward and state
-            ar_mask = (t % it_len == index_lookahead_reward).long() * (
-                t <= u + index_lookahead_reward
+            ar_mask = (t % self.it_len == self.index_lookahead_reward).long() * (
+                t <= u + self.index_lookahead_reward
             ).long()
             ar(result, ar_mask)
 
             # Generate the lookahead_reward and state
-            ar_mask = (t >= u + index_states).long() * (
-                t < u + index_states + state_len
+            ar_mask = (t >= u + self.index_states).long() * (
+                t < u + self.index_states + self.state_len
             ).long()
             ar(result, ar_mask)
 
             # Re-generate the lookahead_reward
-            ar_mask = (t % it_len == index_lookahead_reward).long() * (
-                t <= u + index_lookahead_reward
+            ar_mask = (t % self.it_len == self.index_lookahead_reward).long() * (
+                t <= u + self.index_lookahead_reward
             ).long()
             ar(result, ar_mask, logit_biases=optimistic_bias)
 
             # Generate the action and reward
-            ar_mask = (t >= u + index_action).long() * (t <= u + index_reward).long()
+            ar_mask = (t >= u + self.index_action).long() * (
+                t <= u + self.index_reward
+            ).long()
             ar(result, ar_mask)
 
         filename = os.path.join(result_dir, f"test_thinking_compute_{n_epoch:04d}.txt")