Update.

diff --git a/escape.py b/escape.py
index f2a1662..8066479 100755 (executable)
--- a/escape.py
+++ b/escape.py
@@ -266,11 +266,11 @@ def episodes2str(
 ######################################################################
 
 if __name__ == "__main__":
-    nb, height, width, T, nb_walls = 5, 5, 7, 20, 5
+    nb, height, width, T, nb_walls = 5, 5, 7, 4, 5
     states, actions, rewards = generate_episodes(nb, height, width, T, nb_walls)
     seq = episodes2seq(states, actions, rewards)
     lr, s, a, r = seq2episodes(seq, height, width)
     print(episodes2str(lr, s, a, r, unicode=True, ansi_colors=True))
-    # print()
-    # for s in seq2str(seq):
-    # print(s)
+    print()
+    for s in seq2str(seq):
+        print(s)

diff --git a/tasks.py b/tasks.py
index 12c6125..f2b7709 100755 (executable)
--- a/tasks.py
+++ b/tasks.py
@@ -1893,7 +1893,7 @@ class Escape(Task):
         states, actions, rewards = escape.generate_episodes(
             nb_train_samples + nb_test_samples, height, width, T, nb_walls
         )
-        seq = escape.episodes2seq(states, actions, rewards, lookahead_delta=T)
+        seq = escape.episodes2seq(states, actions, rewards)
         # seq = seq[:, seq.size(1) // 3 : 2 * seq.size(1) // 3]
         self.train_input = seq[:nb_train_samples].to(self.device)
         self.test_input = seq[nb_train_samples:].to(self.device)
@@ -1920,10 +1920,11 @@ class Escape(Task):
         t = torch.arange(result.size(1), device=result.device)[None, :]
 
         state_len = self.height * self.width
-        index_action = state_len
-        index_reward = state_len + 1
-        index_lookahead_reward = state_len + 2
-        it_len = state_len + 3  # state / action / reward / lookahead_reward
+        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
 
@@ -1952,20 +1953,16 @@ class Escape(Task):
         for u in tqdm.tqdm(
             range(it_len, result.size(1) - it_len + 1, it_len), desc="thinking"
         ):
-            # Re-generate the lookahead_reward pessimistically in the
-            # previous iterations
-            ar_mask = (t < u).long() * (t % it_len == index_lookahead_reward).long()
-            ar(result, ar_mask, logit_biases=-optimistic_bias)
-            snapshots.append(result[:10].detach().clone())
-
-            # Generate the state
-            ar_mask = (t >= u).long() * (t < u + state_len).long()
+            # Generate the lookahead_reward and state
+            ar_mask = (t >= u + index_lookahead_reward).long() * (
+                t < u + index_states + state_len
+            ).long()
             ar(result, ar_mask)
             snapshots.append(result[:10].detach().clone())
+            backup_lookahead_reward = result[:, u + index_lookahead_reward]
 
-            # Re-generate the lookahead_reward optimistically in the
-            # previous iterations
-            ar_mask = (t < u).long() * (t % it_len == index_lookahead_reward).long()
+            # Re-generate the lookahead_reward
+            ar_mask = (t == u + index_lookahead_reward).long()
             ar(result, ar_mask, logit_biases=optimistic_bias)
             snapshots.append(result[:10].detach().clone())
 
@@ -1974,12 +1971,14 @@ class Escape(Task):
             ar(result, ar_mask)
             snapshots.append(result[:10].detach().clone())
 
+            result[:, u + index_lookahead_reward] = backup_lookahead_reward
+
         filename = os.path.join(result_dir, f"test_thinking_compute_{n_epoch:04d}.txt")
         with open(filename, "w") as f:
             for n in range(10):
                 for s in snapshots:
                     lr, s, a, r = escape.seq2episodes(
-                        s[n : n + 1], self.height, self.width, lookahead=True
+                        s[n : n + 1], self.height, self.width
                     )
                     str = escape.episodes2str(
                         lr, s, a, r, unicode=True, ansi_colors=True
@@ -1989,7 +1988,7 @@ class Escape(Task):
 
         # Saving the generated sequences
 
-        s, a, r, lr = escape.seq2episodes(result, self.height, self.width)
+        lr, s, a, r = escape.seq2episodes(result, self.height, self.width)
         str = escape.episodes2str(lr, s, a, r, unicode=True, ansi_colors=True)
 
         filename = os.path.join(result_dir, f"test_thinking_seq_{n_epoch:04d}.txt")
@@ -2004,7 +2003,7 @@ class Escape(Task):
 
         # Saving the ground truth
 
-        s, a, r, lr = escape.seq2episodes(
+        lr, s, a, r = escape.seq2episodes(
             result,
             self.height,
             self.width,
@@ -2036,7 +2035,7 @@ class Escape(Task):
 
         # Saving the generated sequences
 
-        s, a, r, lr = escape.seq2episodes(
+        lr, s, a, r = escape.seq2episodes(
             result,
             self.height,
             self.width,