X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=escape.py;fp=escape.py;h=a3d8c85962296639f5d2fc53ff84d6df4c2f9d5e;hb=0b8185b90014369f0d39892e128ad04a7d9ae872;hp=8066479e3aafd104001773df583dcf2082f41804;hpb=1eeba5d817d6e440a93895d42f6e580e9ba273fd;p=picoclvr.git

diff --git a/escape.py b/escape.py
index 8066479..a3d8c85 100755
--- a/escape.py
+++ b/escape.py
@@ -11,7 +11,7 @@ from torch.nn import functional as F
 
 ######################################################################
 
-nb_states_codes = 4
+nb_states_codes = 5
 nb_actions_codes = 5
 nb_rewards_codes = 3
 nb_lookahead_rewards_codes = 3
@@ -60,14 +60,13 @@ def code2lookahead_reward(r):
 ######################################################################
 
 
-def generate_episodes(nb, height=6, width=6, T=10, nb_walls=3):
+def generate_episodes(nb, height=6, width=6, T=10, nb_walls=3, nb_coins=3):
     rnd = torch.rand(nb, height, width)
     rnd[:, 0, :] = 0
     rnd[:, -1, :] = 0
     rnd[:, :, 0] = 0
     rnd[:, :, -1] = 0
     wall = 0
-
     for k in range(nb_walls):
         wall = wall + (
             rnd.flatten(1).argmax(dim=1)[:, None]
@@ -76,6 +75,17 @@ def generate_episodes(nb, height=6, width=6, T=10, nb_walls=3):
 
         rnd = rnd * (1 - wall.clamp(max=1))
 
+    rnd = torch.rand(nb, height, width)
+    coins = torch.zeros(nb, T, height, width, dtype=torch.int64)
+    rnd = rnd * (1 - wall.clamp(max=1))
+    for k in range(nb_coins):
+        coins[:, 0] = coins[:, 0] + (
+            rnd.flatten(1).argmax(dim=1)[:, None]
+            == torch.arange(rnd.flatten(1).size(1))[None, :]
+        ).long().reshape(rnd.size())
+
+        rnd = rnd * (1 - coins[:, 0].clamp(max=1))
+
     states = wall[:, None, :, :].expand(-1, T, -1, -1).clone()
 
     agent = torch.zeros(states.size(), dtype=torch.int64)
@@ -131,9 +141,12 @@ def generate_episodes(nb, height=6, width=6, T=10, nb_walls=3):
 
         # assert hit.min() == 0 and hit.max() <= 1
 
-        rewards[:, t + 1] = -hit + (1 - hit) * agent[:, t + 1, -1, -1]
+        got_coin = (agent[:, t + 1] * coins[:, t]).flatten(1).sum(dim=1)
+        coins[:, t + 1] = coins[:, t] * (1 - agent[:, t + 1])
+
+        rewards[:, t + 1] = -hit + (1 - hit) * got_coin
 
-    states += 2 * agent + 3 * monster
+    states = states + 2 * agent + 3 * monster + 4 * coins
 
     return states, agent_actions, rewards
 
@@ -173,7 +186,7 @@ def seq2episodes(seq, height, width):
 def seq2str(seq):
     def token2str(t):
         if t >= first_states_code and t < first_states_code + nb_states_codes:
-            return " #@$"[t - first_states_code]
+            return " #@T$"[t - first_states_code]
         elif t >= first_actions_code and t < first_actions_code + nb_actions_codes:
             return "ISNEW"[t - first_actions_code]
         elif t >= first_rewards_code and t < first_rewards_code + nb_rewards_codes:
@@ -196,11 +209,11 @@ def episodes2str(
     lookahead_rewards, states, actions, rewards, unicode=False, ansi_colors=False
 ):
     if unicode:
-        symbols = "·█@$"
+        symbols = "·█@T$"
         # vert, hori, cross, thin_hori = "║", "═", "╬", "─"
         vert, hori, cross, thin_vert, thin_hori = "┃", "━", "╋", "│", "─"
     else:
-        symbols = " #@$"
+        symbols = " #@T$"
         vert, hori, cross, thin_vert, thin_hori = "|", "-", "+", "|", "-"
 
     hline = (cross + hori * states.size(-1)) * states.size(1) + cross + "\n"
@@ -266,11 +279,11 @@ def episodes2str(
 ######################################################################
 
 if __name__ == "__main__":
-    nb, height, width, T, nb_walls = 5, 5, 7, 4, 5
+    nb, height, width, T, nb_walls = 5, 5, 7, 10, 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)