X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=tasks.py;h=51538366383be455f160ab6158392627ecea5190;hb=19ec7f3e4030ddece2647983dcf1bed5eb0d9544;hp=02c44bb426f741aec694205ddb384ca16a4cdb5e;hpb=9141338f022ff991ac91e448eda0fd1cb401fd84;p=picoclvr.git

diff --git a/tasks.py b/tasks.py
index 02c44bb..5153836 100755
--- a/tasks.py
+++ b/tasks.py
@@ -1874,6 +1874,7 @@ class Escape(Task):
         height,
         width,
         T,
+        nb_walls,
         logger=None,
         device=torch.device("cpu"),
     ):
@@ -1885,10 +1886,10 @@ 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)
 
@@ -1912,11 +1913,15 @@ 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, :]
+
+        state_len = self.height * self.width
+        it_len = state_len + 3  # state / action / reward / lookahead_reward
 
-        def ar():
+        def ar(result, ar_mask):
+            ar_mask = ar_mask.expand_as(result)
+            result *= 1 - ar_mask
             masked_inplace_autoregression(
                 model,
                 self.batch_size,
@@ -1924,24 +1929,45 @@ class Escape(Task):
                 ar_mask,
                 deterministic_synthesis,
                 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
+        # Generate iteration after iteration
+
+        for u in tqdm.tqdm(
+            range(it_len, result.size(1) - it_len + 1, it_len), desc="thinking"
+        ):
+            # Put the lookahead reward to either 0 or -1 for the
+            # current iteration, with a proba that depends with the
+            # sequence index, so that we have diverse examples, sample
+            # the next state
+            s = -(
+                torch.rand(result.size(0), device=result.device)
+                <= torch.linspace(0, 1, result.size(0), device=result.device)
             ).long()
-            ar_mask = ar_mask[None, :]
-            ar_mask = ar_mask.expand_as(result)
-            result *= 1 - ar_mask
-            ar()
+            result[:, u - 1] = s + 1 + escape.first_lookahead_rewards_code
+            ar_mask = (t >= u).long() * (t < u + state_len).long()
+            ar(result, ar_mask)
+
+            # Put the lookahead reward to +1 for the current
+            # iteration, sample the action and reward
+            s = 1
+            result[:, u - 1] = s + 1 + escape.first_lookahead_rewards_code
+            ar_mask = (t >= u + state_len).long() * (t < u + state_len + 2).long()
+            ar(result, ar_mask)
+
+            # Fix the previous lookahead rewards in a consistant state
+            for v in range(0, u, it_len):
+                # Extract the rewards
+                r = result[:, range(v + state_len + 1 + it_len, u + it_len - 1, it_len)]
+                r = r - escape.first_rewards_code - 1
+                r = r.clamp(min=-1, max=1)  # the reward is predicted hence can be weird
+                a = r.min(dim=1).values
+                b = r.max(dim=1).values
+                s = (a < 0).long() * a + (a >= 0).long() * b
+                result[:, v + state_len + 2] = (
+                    s + 1 + escape.first_lookahead_rewards_code
+                )
 
         # Saving the generated sequences
 
@@ -1960,7 +1986,7 @@ 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