X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=beaver.py;h=c3b7e09c1714199729a737a70725b3944ab787b7;hb=a0e547917131af0b353e3bf31a062c9b35c8dd18;hp=1408f0b2b2b1a0705916b8ebb74676a47e7a4e7b;hpb=1e6f089e67087e8cf1bcb6865e8d405b0a50f372;p=beaver.git

diff --git a/beaver.py b/beaver.py
index 1408f0b..c3b7e09 100755
--- a/beaver.py
+++ b/beaver.py
@@ -175,7 +175,9 @@ def one_shot(gpt, task):
     model = nn.Sequential(
         nn.Linear(args.dim_model, args.dim_model),
         nn.ReLU(),
-        nn.Linear(args.dim_model, 4)
+        nn.Linear(args.dim_model, args.dim_model),
+        nn.ReLU(),
+        nn.Linear(args.dim_model, 4),
     ).to(device)
 
     for n_epoch in range(args.nb_epochs):
@@ -186,9 +188,12 @@ def one_shot(gpt, task):
         for input, targets in task.policy_batches(split="train"):
             output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
             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(-1).mean()
-                + targets.xlogy(targets).sum(-1).mean()
+                -(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)
@@ -201,9 +206,12 @@ def one_shot(gpt, task):
         for input, targets in task.policy_batches(split="test"):
             output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
             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(-1).mean()
-                + targets.xlogy(targets).sum(-1).mean()
+                -(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)
@@ -212,6 +220,23 @@ 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 = task.test_input[:32, : task.height * task.width]
+        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 = 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 +379,10 @@ 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)