Update.

[picoclvr.git] / escape.py

diff --git a/escape.py b/escape.py
index 43843f0..6f4af35 100755 (executable)
--- a/escape.py
+++ b/escape.py
@@ -25,6 +25,33 @@ nb_codes = first_lookahead_rewards_code + nb_lookahead_rewards_codes
 ######################################################################
 
 
+def action2code(r):
+    return first_actions_code + r
+
+
+def code2action(r):
+    return r - first_actions_code
+
+
+def reward2code(r):
+    return first_rewards_code + r + 1
+
+
+def code2reward(r):
+    return r - first_rewards_code - 1
+
+
+def lookahead_reward2code(r):
+    return first_lookahead_rewards_code + r + 1
+
+
+def code2lookahead_reward(r):
+    return r - first_lookahead_rewards_code - 1
+
+
+######################################################################
+
+
 def generate_episodes(nb, height=6, width=6, T=10, nb_walls=3):
     rnd = torch.rand(nb, height, width)
     rnd[:, 0, :] = 0
@@ -94,7 +121,7 @@ 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]
 
@@ -111,17 +138,6 @@ 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
-        # 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):
@@ -133,27 +149,27 @@ 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
         )