Update

diff --git a/beaver.py b/beaver.py
index e22fc7b..afec61d 100755 (executable)
--- a/beaver.py
+++ b/beaver.py
@@ -191,12 +191,11 @@ def one_shot(gpt, task):
             output = model(output_gpt)
             targets = targets * (input.unsqueeze(-1) == maze.v_empty)
             output = output * (input.unsqueeze(-1) == maze.v_empty)
-            loss = (output.softmax(-1) - targets).abs().max(-1).values.mean()
-            # loss = (
-            # -(output.log_softmax(-1) * targets).sum()
-            # / (input == maze.v_empty).sum()
-            # + targets.xlogy(targets).sum() / (input == maze.v_empty).sum()
-            # )
+            # loss = (output.softmax(-1) - targets).abs().max(-1).values.mean()
+            loss = (
+                -(output.log_softmax(-1) * targets).sum()
+                / (input == maze.v_empty).sum()
+            )
             acc_train_loss += loss.item() * input.size(0)
             nb_train_samples += input.size(0)
 
@@ -210,12 +209,11 @@ def one_shot(gpt, task):
             output = model(output_gpt)
             targets = targets * (input.unsqueeze(-1) == maze.v_empty)
             output = output * (input.unsqueeze(-1) == maze.v_empty)
-            loss = (output.softmax(-1) - targets).abs().max(-1).values.mean()
-            # loss = (
-            # -(output.log_softmax(-1) * targets).sum()
-            # / (input == maze.v_empty).sum()
-            # + targets.xlogy(targets).sum() / (input == maze.v_empty).sum()
-            # )
+            # loss = (output.softmax(-1) - targets).abs().max(-1).values.mean()
+            loss = (
+                -(output.log_softmax(-1) * targets).sum()
+                / (input == maze.v_empty).sum()
+            )
             acc_test_loss += loss.item() * input.size(0)
             nb_test_samples += input.size(0)
 
@@ -231,8 +229,11 @@ def one_shot(gpt, task):
         # losses = (-output.log_softmax(-1) * targets + targets.xlogy(targets)).sum(-1)
         # losses = losses * (input == maze.v_empty)
         # losses = losses / losses.max()
-        losses = (output.softmax(-1) - targets).abs().max(-1).values
-        losses = (losses >= 0.05).float()
+        # losses = (output.softmax(-1) - targets).abs().max(-1).values
+        # losses = (losses >= 0.05).float()
+        losses = (
+            (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
+        ).float()
         losses = losses.reshape(-1, args.maze_height, args.maze_width)
         input = input.reshape(-1, args.maze_height, args.maze_width)
         maze.save_image(