Update

diff --git a/beaver.py b/beaver.py
index c3b7e09..e22fc7b 100755 (executable)
--- a/beaver.py
+++ b/beaver.py
@@ -172,6 +172,7 @@ def compute_perplexity(model, split="train"):
 def one_shot(gpt, task):
     t = gpt.training
     gpt.eval()
+
     model = nn.Sequential(
         nn.Linear(args.dim_model, args.dim_model),
         nn.ReLU(),
@@ -190,11 +191,12 @@ 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.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()
+            # + targets.xlogy(targets).sum() / (input == maze.v_empty).sum()
+            # )
             acc_train_loss += loss.item() * input.size(0)
             nb_train_samples += input.size(0)
 
@@ -208,11 +210,12 @@ 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.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()
+            # + targets.xlogy(targets).sum() / (input == maze.v_empty).sum()
+            # )
             acc_test_loss += loss.item() * input.size(0)
             nb_test_samples += input.size(0)
 
@@ -225,9 +228,11 @@ def one_shot(gpt, task):
         targets = task.test_policies[:32]
         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 * (input == maze.v_empty)
-        losses = losses / losses.max()
+        # 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 = losses.reshape(-1, args.maze_height, args.maze_width)
         input = input.reshape(-1, args.maze_height, args.maze_width)
         maze.save_image(