Update.
authorFrançois Fleuret <francois@fleuret.org>
Wed, 27 Mar 2024 06:22:09 +0000 (07:22 +0100)
committerFrançois Fleuret <francois@fleuret.org>
Wed, 27 Mar 2024 06:22:09 +0000 (07:22 +0100)
greed.py
tasks.py

index 3cbe886..20cef79 100755 (executable)
--- a/greed.py
+++ b/greed.py
@@ -300,7 +300,8 @@ def save_seq_as_anim_script(seq, filename):
             )
             f.write(episodes2str(lr, s, a, r, unicode=True, ansi_colors=True))
             f.write("EOF\n")
-            f.write("sleep 0.5\n")
+            f.write("sleep 0.25\n")
+        print(f"Saved {filename}")
 
 
 if __name__ == "__main__":
index 845b5b3..aa5df72 100755 (executable)
--- a/tasks.py
+++ b/tasks.py
@@ -1905,6 +1905,15 @@ class Greed(Task):
         self.index_reward = self.state_len + 2
         self.it_len = self.state_len + 3  # lookahead_reward / state / action / reward
 
+    def wipe_lookahead_rewards(self, batch):
+        t = torch.arange(batch.size(1), device=batch.device)[None, :]
+        u = torch.randint(batch.size(1), (batch.size(0), 1), device=batch.device)
+        lr_mask = (t <= u).long() * (
+            t % self.it_len == self.index_lookahead_reward
+        ).long()
+
+        return lr_mask * greed.lookahead_reward2code(2) + (1 - lr_mask) * batch
+
     def batches(self, split="train", nb_to_use=-1, desc=None):
         assert split in {"train", "test"}
         input = self.train_input if split == "train" else self.test_input
@@ -1915,14 +1924,7 @@ class Greed(Task):
         for batch in tqdm.tqdm(
             input.split(self.batch_size), dynamic_ncols=True, desc=desc
         ):
-            t = torch.arange(batch.size(1), device=batch.device)[None, :]
-            u = torch.randint(batch.size(1), (batch.size(0), 1), device=batch.device)
-            lr_mask = (t <= u).long() * (
-                t % self.it_len == self.index_lookahead_reward
-            ).long()
-
-            batch = lr_mask * greed.lookahead_reward2code(2) + (1 - lr_mask) * batch
-            yield batch
+            yield self.wipe_lookahead_rewards(batch)
 
     def vocabulary_size(self):
         return greed.nb_codes
@@ -2010,7 +2012,7 @@ class Greed(Task):
     def produce_results(
         self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
     ):
-        result = self.test_input[:250].clone()
+        result = self.wipe_lookahead_rewards(self.test_input[:250].clone())
 
         # Saving the ground truth