X-Git-Url: https://www.fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=beaver.py;h=5abe39b767c13299d8dcffcc3369682bfbfff69f;hb=HEAD;hp=6e1eaf4137f5a3bc184f4bd8fd529adbd1791745;hpb=519b5419b30de82828a41b620e76d993a70423e9;p=beaver.git

diff --git a/beaver.py b/beaver.py
index 6e1eaf4..5abe39b 100755
--- a/beaver.py
+++ b/beaver.py
@@ -127,6 +127,8 @@ def log_string(s):
     sys.stdout.flush()
 
 
+log_string(f"cmd {' '.join(sys.argv)}")
+
 for n in vars(args):
     log_string(f"args.{n} {getattr(args, n)}")
 
@@ -236,8 +238,13 @@ def oneshot_trace_loss(mazes, output, policies, height, width):
 def oneshot(model, learning_rate_scheduler, task):
     t = model.training
     model.eval()
-    mazes = task.test_input[:32].clone()
+    mazes = task.test_input[:48].clone()
     mazes[:, task.height * task.width :] = 0
+    policies = task.test_policies[:48]
+    targets = maze.stationary_densities(
+        mazes[:, : task.height * task.width].view(-1, task.height, task.width),
+        policies.view(-1, 4, task.height, task.width),
+    ).flatten(-2)
     output = eval_mygpt(model, mazes, prompt_len=task.height * task.width)
     output = F.softmax(output, dim=2)
     print(f"{output.size()=}")
@@ -245,16 +252,17 @@ def oneshot(model, learning_rate_scheduler, task):
         -1, task.height, task.width
     )
     mazes = mazes[:, : task.height * task.width].reshape(-1, task.height, task.width)
-    paths = task.test_input[:32, task.height * task.width :].reshape(
+    targets = targets.reshape(-1, task.height, task.width)
+    paths = task.test_input[:48, task.height * task.width :].reshape(
         -1, task.height, task.width
     )
     filename = f"oneshot.png"
     maze.save_image(
         os.path.join(args.result_dir, filename),
         mazes=mazes,
-        target_paths=paths,
+        # target_paths=paths,
         score_paths=proba_path,
-        # score_truth=targets,
+        score_truth=targets,
     )
     log_string(f"wrote {filename}")
 
@@ -327,8 +335,8 @@ def oneshot_old(gpt, learning_rate_scheduler, task):
         )
 
         # -------------------
-        mazes = task.test_input[:32, : task.height * task.width]
-        policies = task.test_policies[:32]
+        mazes = task.test_input[:48, : task.height * task.width]
+        policies = task.test_policies[:48]
         output_gpt = eval_mygpt(
             gpt, mazes, mode=args.oneshot_input, prompt_len=task.height * task.width
         )
@@ -571,7 +579,7 @@ class TaskMaze(Task):
                 f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
             )
 
-            input = self.test_input[:32]
+            input = self.test_input[:48]
             result = input.clone()
             ar_mask = result.new_zeros(result.size())
             ar_mask[:, self.height * self.width :] = 1