Update.

[picoclvr.git] / tasks.py

diff --git a/tasks.py b/tasks.py
index 11879fd..870ab95 100755 (executable)
--- a/tasks.py
+++ b/tasks.py
@@ -1898,6 +1898,13 @@ class Escape(Task):
         self.train_input = seq[:nb_train_samples].to(self.device)
         self.test_input = seq[nb_train_samples:].to(self.device)
 
+        self.state_len = self.height * self.width
+        self.index_lookahead_reward = 0
+        self.index_states = 1
+        self.index_action = self.state_len + 1
+        self.index_reward = self.state_len + 2
+        self.it_len = self.state_len + 3  # lookahead_reward / state / action / reward
+
     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
@@ -1908,6 +1915,13 @@ class Escape(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 * escape.lookahead_reward2code(2) + (1 - lr_mask) * batch
             yield batch
 
     def vocabulary_size(self):
@@ -1916,18 +1930,6 @@ class Escape(Task):
     def thinking_autoregression(
         self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
     ):
-        result = self.test_input[:250].clone()
-        t = torch.arange(result.size(1), device=result.device)[None, :]
-
-        state_len = self.height * self.width
-        index_lookahead_reward = 0
-        index_states = 1
-        index_action = state_len + 1
-        index_reward = state_len + 2
-        it_len = state_len + 3  # lookahead_reward / state / action / reward
-
-        result[:, it_len:] = -1
-
         snapshots = []
 
         def ar(result, ar_mask, logit_biases=None):
@@ -1948,36 +1950,36 @@ class Escape(Task):
 
         # Generate iteration after iteration
 
-        optimistic_bias = result.new_zeros(escape.nb_codes, device=result.device)
-        optimistic_bias[escape.lookahead_reward2code(-1)] = -math.log(1e1)
-        optimistic_bias[escape.lookahead_reward2code(1)] = math.log(1e1)
+        result = self.test_input[:250].clone()
+        result[:, self.it_len :] = -1
+        result[:, self.index_lookahead_reward] = escape.lookahead_reward2code(2)
+        t = torch.arange(result.size(1), device=result.device)[None, :]
 
         for u in tqdm.tqdm(
-            range(it_len, result.size(1) - it_len + 1, it_len), desc="thinking"
+            range(0, result.size(1), self.it_len),
+            desc="thinking",
         ):
-            lr, _, _, _ = escape.seq2episodes(result[:, :u], self.height, self.width)
-
-            # Generate the lookahead_reward and state
-            ar_mask = (t % it_len == index_lookahead_reward).long() * (
-                t <= u + index_lookahead_reward
-            ).long()
-            ar(result, ar_mask)
-
-            # Generate the lookahead_reward and state
-            ar_mask = (t >= u + index_states).long() * (
-                t < u + index_states + state_len
+            # Generate the next state but keep the initial one, the
+            # lookahead_reward of previous iterations are set to
+            # UNKNOWN
+            if u > 0:
+                result[
+                    :, u + self.index_lookahead_reward
+                ] = escape.lookahead_reward2code(2)
+                ar_mask = (t >= u + self.index_states).long() * (
+                    t < u + self.index_states + self.state_len
+                ).long()
+                ar(result, ar_mask)
+
+            # Generate the action and reward with lookahead_reward to +1
+            result[:, u + self.index_lookahead_reward] = escape.lookahead_reward2code(1)
+            ar_mask = (t >= u + self.index_action).long() * (
+                t <= u + self.index_reward
             ).long()
             ar(result, ar_mask)
 
-            # Re-generate the lookahead_reward
-            ar_mask = (t % it_len == index_lookahead_reward).long() * (
-                t <= u + index_lookahead_reward
-            ).long()
-            ar(result, ar_mask, logit_biases=optimistic_bias)
-
-            # Generate the action and reward
-            ar_mask = (t >= u + index_action).long() * (t <= u + index_reward).long()
-            ar(result, ar_mask)
+            # Set the lookahead_reward to UNKNOWN for the next iterations
+            result[:, u + self.index_lookahead_reward] = escape.lookahead_reward2code(2)
 
         filename = os.path.join(result_dir, f"test_thinking_compute_{n_epoch:04d}.txt")
         with open(filename, "w") as f: