X-Git-Url: https://www.fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=beaver.py;h=2cc214019f48c47da2fc29edd8f84cdc8a10e7ba;hb=41c7509dc3d2153da79ed09ecf4a3b592503f15e;hp=54510f023d28d861557c933a0767f0d81eb8fece;hpb=0e74c10ba17f2f969072f3989579c0d6f47f1cbb;p=beaver.git

diff --git a/beaver.py b/beaver.py
index 54510f0..2cc2140 100755
--- a/beaver.py
+++ b/beaver.py
@@ -178,7 +178,10 @@ def one_shot(gpt, task):
         nn.Linear(args.dim_model, 4),
     ).to(device)
 
+    print(f"{args.nb_epochs=}")
+
     for n_epoch in range(args.nb_epochs):
+        print(f"{n_epoch=}")
         learning_rate = learning_rate_schedule[n_epoch]
         optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 
@@ -212,6 +215,23 @@ def one_shot(gpt, task):
             f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
         )
 
+        # -------------------
+        input, targets = next(task.policy_batches(split="test"))
+        output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
+        output = model(output_gpt)
+        losses = (-output.log_softmax(-1) * targets + targets.xlogy(targets)).sum(-1)
+        losses = losses / losses.max()
+        print(f"{input.size()=} {losses.size()=} {losses.min()=} {losses.max()=}")
+        losses = losses * (input == 0)
+        losses = losses.reshape(-1, args.maze_height, args.maze_width)
+        input = input.reshape(-1, args.maze_height, args.maze_width)
+        maze.save_image(
+            os.path.join(args.result_dir, f"oneshot_{n_epoch:04d}.png"),
+            mazes=input,
+            score_paths=losses,
+        )
+        # -------------------
+
     gpt.train(t)
 
 
@@ -354,10 +374,10 @@ class TaskMaze(Task):
             _, predicted_paths = self.seq2map(result)
             maze.save_image(
                 os.path.join(args.result_dir, f"result_{n_epoch:04d}.png"),
-                mazes,
-                paths,
-                predicted_paths,
-                maze.path_correctness(mazes, predicted_paths),
+                mazes=mazes,
+                target_paths=paths,
+                predicted_paths=predicted_paths,
+                path_correct=maze.path_correctness(mazes, predicted_paths),
             )
 
             model.train(t)