Update

diff --git a/beaver.py b/beaver.py
index 33d174d..b505156 100755 (executable)
--- a/beaver.py
+++ b/beaver.py
@@ -68,8 +68,6 @@ parser.add_argument("--no_checkpoint", action="store_true", default=False)
 
 parser.add_argument("--overwrite_results", action="store_true", default=False)
 
-parser.add_argument("--one_shot", action="store_true", default=False)
-
 parser.add_argument("--checkpoint_name", type=str, default="checkpoint.pth")
 
 ##############################
@@ -81,7 +79,11 @@ parser.add_argument("--maze_width", type=int, default=21)
 
 parser.add_argument("--maze_nb_walls", type=int, default=15)
 
-parser.add_argument("--oneshot_mode", type=str, default="head")
+parser.add_argument("--oneshot", action="store_true", default=False)
+
+parser.add_argument("--oneshot_input", type=str, default="head")
+
+parser.add_argument("--oneshot_output", type=str, default="policy")
 
 ######################################################################
 
@@ -171,10 +173,34 @@ def compute_perplexity(model, split="train"):
 ######################################################################
 
 
-def one_shot(gpt, task):
+def oneshot_policy_loss(output, policies, mask):
+    targets = policies.permute(0, 2, 1) * mask.unsqueeze(-1)
+    output = output * mask.unsqueeze(-1)
+    return -(output.log_softmax(-1) * targets).sum() / mask.sum()
+
+
+# loss = (output.softmax(-1) - targets).abs().max(-1).values.mean()
+
+
+def oneshot(gpt, task):
     t = gpt.training
     gpt.eval()
-    dim_in = args.dim_model * (args.nb_blocks * 2 if args.oneshot_mode == "deep" else 1)
+
+    if args.oneshot_input == "head":
+        dim_in = args.dim_model
+    elif args.oneshot_input == "deep":
+        dim_in = args.dim_model * args.nb_blocks * 2
+    else:
+        raise ValueError(f"{args.oneshot_input=}")
+
+    if args.oneshot_output == "policy":
+        dim_out = 4
+        compute_loss = oneshot_policy_loss
+    elif args.oneshot_output == "trace":
+        dim_out = 1
+    else:
+        raise ValueError(f"{args.oneshot_output=}")
+
     model = nn.Sequential(
         nn.Linear(dim_in, args.dim_model),
         nn.ReLU(),
@@ -194,13 +220,11 @@ def one_shot(gpt, task):
             # s = maze.stationary_densities(
             # exit(0)
             ####
-            mask = input.unsqueeze(-1) == maze.v_empty
-            output_gpt = gpt(mygpt.BracketedSequence(input), mode=args.oneshot_mode).x
+            mask = input == maze.v_empty
+            output_gpt = gpt(mygpt.BracketedSequence(input), mode=args.oneshot_input).x
             output = model(output_gpt)
-            targets = policies.permute(0, 2, 1) * mask
-            output = output * mask
-            # loss = (output.softmax(-1) - targets).abs().max(-1).values.mean()
-            loss = -(output.log_softmax(-1) * targets).sum() / mask.sum()
+
+            loss = compute_loss(output, policies, mask)
             acc_train_loss += loss.item() * input.size(0)
             nb_train_samples += input.size(0)
 
@@ -210,13 +234,10 @@ def one_shot(gpt, task):
 
         acc_test_loss, nb_test_samples = 0, 0
         for input, policies in task.policy_batches(split="test"):
-            mask = input.unsqueeze(-1) == maze.v_empty
-            output_gpt = gpt(mygpt.BracketedSequence(input), mode=args.oneshot_mode).x
+            mask = input == maze.v_empty
+            output_gpt = gpt(mygpt.BracketedSequence(input), mode=args.oneshot_input).x
             output = model(output_gpt)
-            targets = policies.permute(0, 2, 1) * mask
-            output = output * mask
-            # loss = (output.softmax(-1) - targets).abs().max(-1).values.mean()
-            loss = -(output.log_softmax(-1) * targets).sum() / mask.sum()
+            loss = compute_loss(output, policies, mask)
             acc_test_loss += loss.item() * input.size(0)
             nb_test_samples += input.size(0)
 
@@ -227,24 +248,20 @@ def one_shot(gpt, task):
         # -------------------
         input = task.test_input[:32, : task.height * task.width]
         targets = task.test_policies[:32].permute(0, 2, 1)
-        output_gpt = gpt(mygpt.BracketedSequence(input), mode=args.oneshot_mode).x
+        output_gpt = gpt(mygpt.BracketedSequence(input), mode=args.oneshot_input).x
         output = model(output_gpt)
-        # losses = (-output.log_softmax(-1) * targets + targets.xlogy(targets)).sum(-1)
-        # losses = losses * mask
-        # losses = losses / losses.max()
-        # losses = (output.softmax(-1) - targets).abs().max(-1).values
-        # losses = (losses >= 0.05).float()
-        losses = (
+        scores = (
             (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)
+        scores = scores.reshape(-1, task.height, task.width)
+        input = input.reshape(-1, task.height, task.width)
         maze.save_image(
             os.path.join(
-                args.result_dir, f"oneshot_{args.oneshot_mode}_{n_epoch:04d}.png"
+                args.result_dir,
+                f"oneshot_{args.oneshot_input}_{args.oneshot_output}_{n_epoch:04d}.png",
             ),
             mazes=input,
-            score_paths=losses,
+            score_paths=scores,
         )
         # -------------------
 
@@ -498,8 +515,8 @@ log_string(f"learning_rate_schedule {learning_rate_schedule}")
 
 ##############################
 
-if args.one_shot:
-    one_shot(model, task)
+if args.oneshot:
+    oneshot(model, task)
     exit(0)
 
 ##############################
@@ -531,7 +548,7 @@ for n_epoch in range(nb_epochs_finished, args.nb_epochs):
     elif args.optim == "adamw":
         optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
     else:
-        raise ValueError(f"Unknown optimizer {args.optim}.")
+        raise ValueError(f"{args.optim=}")
 
     model.train()