Update.

diff --git a/escape.py b/escape.py
index a2e378d..da3e495 100755 (executable)
--- a/escape.py
+++ b/escape.py
@@ -110,13 +110,22 @@ def episodes2seq(states, actions, rewards, lookahead_delta=None):
     actions = actions[:, :, None] + first_actions_code
 
     if lookahead_delta is not None:
-        r = rewards
-        u = F.pad(r, (0, lookahead_delta - 1)).as_strided(
-            (r.size(0), r.size(1), lookahead_delta),
-            (r.size(1) + lookahead_delta - 1, 1, 1),
-        )
-        a = u[:, :, 1:].min(dim=-1).values
-        b = u[:, :, 1:].max(dim=-1).values
+        # r = rewards
+        # u = F.pad(r, (0, lookahead_delta - 1)).as_strided(
+        # (r.size(0), r.size(1), lookahead_delta),
+        # (r.size(1) + lookahead_delta - 1, 1, 1),
+        # )
+        # a = u[:, :, 1:].min(dim=-1).values
+        # b = u[:, :, 1:].max(dim=-1).values
+        # s = (a < 0).long() * a + (a >= 0).long() * b
+        # lookahead_rewards = (1 + s[:, :, None]) + first_lookahead_rewards_code
+
+        # a[n,t]=min_s>t r[n,s]
+        a = rewards.new_zeros(rewards.size())
+        b = rewards.new_zeros(rewards.size())
+        for t in range(a.size(1) - 1):
+            a[:, t] = rewards[:, t + 1 :].min(dim=-1).values
+            b[:, t] = rewards[:, t + 1 :].max(dim=-1).values
         s = (a < 0).long() * a + (a >= 0).long() * b
         lookahead_rewards = (1 + s[:, :, None]) + first_lookahead_rewards_code
 
@@ -271,11 +280,11 @@ def episodes2str(
 ######################################################################
 
 if __name__ == "__main__":
-    nb, height, width, T = 10, 4, 6, 20
+    nb, height, width, T = 1000, 4, 6, 20
     states, actions, rewards = generate_episodes(nb, height, width, T)
     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))
-    print()
-    for s in seq2str(seq):
-        print(s)
+    # print()
+    # for s in seq2str(seq):
+    # print(s)

diff --git a/main.py b/main.py
index b4f831f..2edfa14 100755 (executable)
--- a/main.py
+++ b/main.py
@@ -182,7 +182,7 @@ parser.add_argument("--escape_height", type=int, default=4)
 
 parser.add_argument("--escape_width", type=int, default=6)
 
-parser.add_argument("--escape_T", type=int, default=20)
+parser.add_argument("--escape_T", type=int, default=25)
 
 ######################################################################
 

diff --git a/tasks.py b/tasks.py
index a4ef557..38c85ed 100755 (executable)
--- a/tasks.py
+++ b/tasks.py
@@ -1885,10 +1885,10 @@ class Escape(Task):
         self.width = width
 
         states, actions, rewards = escape.generate_episodes(
-            nb_train_samples + nb_test_samples, height, width, 3 * T
+            nb_train_samples + nb_test_samples, height, width, T
         )
         seq = escape.episodes2seq(states, actions, rewards, lookahead_delta=T)
-        seq = seq[:, seq.size(1) // 3 : 2 * seq.size(1) // 3]
+        # 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)