X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=tasks.py;h=29f1e5a783024a3948867d2ffc0dae8c71b09954;hb=2be22c9825d8aebe8d184e9501355a31318abf2b;hp=38c85ed84d228b217e214b68adbbda1f76c255a9;hpb=690470307a7995cc3117eb54545f921eedcecba5;p=picoclvr.git

diff --git a/tasks.py b/tasks.py
index 38c85ed..29f1e5a 100755
--- 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"),
     ):
@@ -1885,7 +1891,7 @@ class Escape(Task):
         self.width = width
 
         states, actions, rewards = escape.generate_episodes(
-            nb_train_samples + nb_test_samples, height, width, 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]
@@ -1912,39 +1918,55 @@ 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)
+        result = self.test_input[:250].clone()
+        t = torch.arange(result.size(1), device=result.device)[None, :]
+
         state_len = self.height * self.width
-        iteration_len = state_len + 3
+        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():
+        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(
-            iteration_len, result.size(1) - iteration_len + 1, iteration_len
+        # Generate iteration after iteration
+
+        optimistic_bias = result.new_zeros(self.nb_codes, device=result.device)
+        optimistic_bias[escape.lookahead_reward2code(-1)] = -math.log(1e1)
+        optimistic_bias[escape.lookahead_reward2code(1)] = math.log(1e1)
+
+        for u in tqdm.tqdm(
+            range(it_len, result.size(1) - it_len + 1, it_len), desc="thinking"
         ):
-            # Put a lookahead reward to -1, sample the next state
-            result[:, u - 1] = (-1) + 1 + escape.first_lookahead_rewards_code
+            # Re-generate the lookahead_reward pessimistically in the
+            # previous iterations
+            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_mask = ar_mask[None, :]
-            ar_mask = ar_mask.expand_as(result)
-            result *= 1 - ar_mask
-            ar()
+            ar(result, ar_mask)
 
-            # Put a lookahead reward to +1, sample the action and reward
-            result[:, u - 1] = (1) + 1 + escape.first_lookahead_rewards_code
-            ar_mask = (t >= state_len).long() * (t < state_len + 2).long()
-            ar_mask = ar_mask[None, :]
-            ar_mask = ar_mask.expand_as(result)
-            result *= 1 - ar_mask
-            ar()
+            # 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)
+
+            # 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
 
@@ -1963,7 +1985,7 @@ 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