Update.

[picoclvr.git] / tasks.py

diff --git a/tasks.py b/tasks.py
index 29f1e5a..1d967f9 100755 (executable)
--- a/tasks.py
+++ b/tasks.py
@@ -1898,8 +1898,6 @@ class Escape(Task):
         self.train_input = seq[:nb_train_samples].to(self.device)
         self.test_input = seq[nb_train_samples:].to(self.device)
 
-        self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
-
     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
@@ -1913,7 +1911,7 @@ class Escape(Task):
             yield batch
 
     def vocabulary_size(self):
-        return self.nb_codes
+        return escape.nb_codes
 
     def thinking_autoregression(
         self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
@@ -1927,6 +1925,8 @@ class Escape(Task):
         index_lookahead_reward = state_len + 2
         it_len = state_len + 3  # state / action / reward / lookahead_reward
 
+        result[:, it_len:] = -1
+
         def ar(result, ar_mask, logit_biases=None):
             ar_mask = ar_mask.expand_as(result)
             result *= 1 - ar_mask
@@ -1943,10 +1943,12 @@ class Escape(Task):
 
         # Generate iteration after iteration
 
-        optimistic_bias = result.new_zeros(self.nb_codes, device=result.device)
+        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)
 
+        snapshots = []
+
         for u in tqdm.tqdm(
             range(it_len, result.size(1) - it_len + 1, it_len), desc="thinking"
         ):
@@ -1954,19 +1956,36 @@ class Escape(Task):
             # previous iterations
             ar_mask = (t < u).long() * (t % it_len == index_lookahead_reward).long()
             ar(result, ar_mask, logit_biases=-optimistic_bias)
+            snapshots.append(result[:10].detach().clone())
 
             # Generate the state
             ar_mask = (t >= u).long() * (t < u + state_len).long()
             ar(result, ar_mask)
+            snapshots.append(result[:10].detach().clone())
 
             # Re-generate the lookahead_reward optimistically in the
             # previous iterations
             ar_mask = (t < u).long() * (t % it_len == index_lookahead_reward).long()
             ar(result, ar_mask, logit_biases=optimistic_bias)
+            snapshots.append(result[:10].detach().clone())
 
             # Generate the action and reward
             ar_mask = (t >= u + index_action).long() * (t <= u + index_reward).long()
             ar(result, ar_mask)
+            snapshots.append(result[:10].detach().clone())
+
+        filename = os.path.join(result_dir, f"test_thinking_compute_{n_epoch:04d}.txt")
+        with open(filename, "w") as f:
+            for n in range(10):
+                for s in snapshots:
+                    s, a, r, lr = escape.seq2episodes(
+                        s[n : n + 1], self.height, self.width, lookahead=True
+                    )
+                    str = escape.episodes2str(
+                        s, a, r, lookahead_rewards=lr, unicode=True, ansi_colors=True
+                    )
+                    f.write(str)
+                f.write("\n\n")
 
         # Saving the generated sequences