Update.
authorFrançois Fleuret <francois@fleuret.org>
Mon, 25 Mar 2024 06:22:23 +0000 (07:22 +0100)
committerFrançois Fleuret <francois@fleuret.org>
Mon, 25 Mar 2024 06:22:23 +0000 (07:22 +0100)
escape.py
tasks.py

index 43843f0..f51863b 100755 (executable)
--- a/escape.py
+++ b/escape.py
@@ -94,7 +94,7 @@ def generate_episodes(nb, height=6, width=6, T=10, nb_walls=3):
         )
         hit = (hit > 0).long()
 
-        assert hit.min() == 0 and hit.max() <= 1
+        assert hit.min() == 0 and hit.max() <= 1
 
         rewards[:, t + 1] = -hit + (1 - hit) * agent[:, t + 1, -1, -1]
 
@@ -133,27 +133,27 @@ def episodes2seq(states, actions, rewards, lookahead_delta=None):
     r = rewards[:, :, None]
     rewards = (r + 1) + first_rewards_code
 
-    assert (
-        states.min() >= first_state_code
-        and states.max() < first_state_code + nb_state_codes
-    )
-    assert (
-        actions.min() >= first_actions_code
-        and actions.max() < first_actions_code + nb_actions_codes
-    )
-    assert (
-        rewards.min() >= first_rewards_code
-        and rewards.max() < first_rewards_code + nb_rewards_codes
-    )
+    assert (
+    # states.min() >= first_state_code
+    # and states.max() < first_state_code + nb_state_codes
+    )
+    assert (
+    # actions.min() >= first_actions_code
+    # and actions.max() < first_actions_code + nb_actions_codes
+    )
+    assert (
+    # rewards.min() >= first_rewards_code
+    # and rewards.max() < first_rewards_code + nb_rewards_codes
+    )
 
     if lookahead_delta is None:
         return torch.cat([states, actions, rewards], dim=2).flatten(1)
     else:
-        assert (
-            lookahead_rewards.min() >= first_lookahead_rewards_code
-            and lookahead_rewards.max()
-            < first_lookahead_rewards_code + nb_lookahead_rewards_codes
-        )
+        assert (
+        # lookahead_rewards.min() >= first_lookahead_rewards_code
+        # and lookahead_rewards.max()
+        # < first_lookahead_rewards_code + nb_lookahead_rewards_codes
+        )
         return torch.cat([states, actions, rewards, lookahead_rewards], dim=2).flatten(
             1
         )
index fddcaff..5153836 100755 (executable)
--- a/tasks.py
+++ b/tasks.py
@@ -1938,8 +1938,13 @@ class Escape(Task):
             range(it_len, result.size(1) - it_len + 1, it_len), desc="thinking"
         ):
             # Put the lookahead reward to either 0 or -1 for the
-            # current iteration, sample the next state
-            s = -(torch.rand(result.size(0), device=result.device) < 0.2).long()
+            # current iteration, with a proba that depends with the
+            # sequence index, so that we have diverse examples, sample
+            # the next state
+            s = -(
+                torch.rand(result.size(0), device=result.device)
+                <= torch.linspace(0, 1, result.size(0), device=result.device)
+            ).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)
@@ -1956,6 +1961,7 @@ class Escape(Task):
                 # Extract the rewards
                 r = result[:, range(v + state_len + 1 + it_len, u + it_len - 1, it_len)]
                 r = r - escape.first_rewards_code - 1
+                r = r.clamp(min=-1, max=1)  # the reward is predicted hence can be weird
                 a = r.min(dim=1).values
                 b = r.max(dim=1).values
                 s = (a < 0).long() * a + (a >= 0).long() * b