Update.

[picoclvr.git] / tasks.py

diff --git a/tasks.py b/tasks.py
index 02c44bb..6b6b8f2 100755 (executable)
--- a/tasks.py
+++ b/tasks.py
@@ -27,6 +27,7 @@ def masked_inplace_autoregression(
     ar_mask,
     deterministic_synthesis,
     forbidden_tokens=None,
+    logit_biases=None,
     progress_bar_desc="autoregression",
     device=torch.device("cpu"),
 ):
@@ -48,7 +49,11 @@ def masked_inplace_autoregression(
 
         for input, ar_mask in batches:
             model.masked_inplace_autoregression(
-                input, ar_mask, forbidden_tokens, deterministic_synthesis
+                input,
+                ar_mask,
+                deterministic_synthesis,
+                forbidden_tokens,
+                logit_biases,
             )
 
         model.train(t)
@@ -1874,6 +1879,7 @@ class Escape(Task):
         height,
         width,
         T,
+        nb_walls,
         logger=None,
         device=torch.device("cpu"),
     ):
@@ -1885,10 +1891,10 @@ class Escape(Task):
         self.width = width
 
         states, actions, rewards = escape.generate_episodes(
-            nb_train_samples + nb_test_samples, height, width, 3 * T
+            nb_train_samples + nb_test_samples, height, width, T, nb_walls
         )
         seq = escape.episodes2seq(states, actions, rewards, lookahead_delta=T)
-        seq = seq[:, seq.size(1) // 3 : 2 * seq.size(1) // 3]
+        # seq = seq[:, seq.size(1) // 3 : 2 * seq.size(1) // 3]
         self.train_input = seq[:nb_train_samples].to(self.device)
         self.test_input = seq[nb_train_samples:].to(self.device)
 
@@ -1912,36 +1918,53 @@ class Escape(Task):
     def thinking_autoregression(
         self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
     ):
-        result = self.test_input[:100].clone()
-        t = torch.arange(result.size(1), device=result.device)
-        itl = self.height * self.width + 3
+        result = self.test_input[:250].clone()
+        t = torch.arange(result.size(1), device=result.device)[None, :]
 
-        def ar():
+        state_len = self.height * self.width
+        index_action = state_len
+        index_reward = state_len + 1
+        index_lookahead_reward = state_len + 2
+        it_len = state_len + 3  # state / action / reward / lookahead_reward
+
+        def ar(result, ar_mask, logit_biases=None):
+            ar_mask = ar_mask.expand_as(result)
+            result *= 1 - ar_mask
             masked_inplace_autoregression(
                 model,
                 self.batch_size,
                 result,
                 ar_mask,
-                deterministic_synthesis,
+                deterministic_synthesis=deterministic_synthesis,
+                logit_biases=logit_biases,
                 device=self.device,
+                progress_bar_desc=None,
             )
 
-        for u in range(itl, result.size(1) - itl + 1, itl):
-            print(f"{itl=} {u=} {result.size(1)=}")
-            result[:, u - 1] = (-1) + 1 + escape.first_lookahead_rewards_code
-            ar_mask = (t >= u).long() * (t < u + self.height * self.width).long()
-            ar_mask = ar_mask[None, :]
-            ar_mask = ar_mask.expand_as(result)
-            result *= 1 - ar_mask
-            ar()
-            result[:, u - 1] = (1) + 1 + escape.first_lookahead_rewards_code
-            ar_mask = (t >= self.height * self.width).long() * (
-                t < self.height * self.width + 2
-            ).long()
-            ar_mask = ar_mask[None, :]
-            ar_mask = ar_mask.expand_as(result)
-            result *= 1 - ar_mask
-            ar()
+        # Generate iteration after iteration
+
+        optimistic_bias = result.new_zeros(self.nb_codes, device=result.device)
+        optimistic_bias[(-1) + escape.first_lookahead_rewards_code + 1] = math.log(1e-1)
+        optimistic_bias[(1) + escape.first_lookahead_rewards_code + 1] = math.log(1e1)
+
+        for u in tqdm.tqdm(
+            range(it_len, result.size(1) - it_len + 1, it_len), desc="thinking"
+        ):
+            # Generate the lookahead_reward pessimistically
+            ar_mask = (t < u).long() * (t % it_len == index_lookahead_reward).long()
+            ar(result, ar_mask, logit_biases=-optimistic_bias)
+
+            # Generate the state
+            ar_mask = (t >= u).long() * (t < u + state_len).long()
+            ar(result, ar_mask)
+
+            # Generate the lookahead_reward optimistically
+            ar_mask = (t < u).long() * (t % it_len == 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)
 
         # Saving the generated sequences
 
@@ -1960,7 +1983,7 @@ class Escape(Task):
     def produce_results(
         self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
     ):
-        result = self.test_input[:100].clone()
+        result = self.test_input[:250].clone()
 
         # Saving the ground truth