X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=beaver.py;h=e7decd1a330a81e419f87d21ba61f191845aec47;hb=280f38363e34e202d38e6f7c00288329ab067a81;hp=2cc214019f48c47da2fc29edd8f84cdc8a10e7ba;hpb=41c7509dc3d2153da79ed09ecf4a3b592503f15e;p=beaver.git

diff --git a/beaver.py b/beaver.py
index 2cc2140..e7decd1 100755
--- a/beaver.py
+++ b/beaver.py
@@ -81,6 +81,8 @@ 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")
+
 ######################################################################
 
 args = parser.parse_args()
@@ -172,26 +174,29 @@ def compute_perplexity(model, split="train"):
 def one_shot(gpt, task):
     t = gpt.training
     gpt.eval()
+    dim_in = args.dim_model * (args.nb_blocks * 2 if args.oneshot_mode == "deep" else 1)
     model = nn.Sequential(
+        nn.Linear(dim_in, args.dim_model),
+        nn.ReLU(),
         nn.Linear(args.dim_model, args.dim_model),
         nn.ReLU(),
         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)
 
         acc_train_loss, nb_train_samples = 0, 0
         for input, targets in task.policy_batches(split="train"):
-            output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
+            output_gpt = gpt(mygpt.BracketedSequence(input), mode=args.oneshot_mode).x
             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(-1).mean()
-                + targets.xlogy(targets).sum(-1).mean()
+                -(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)
@@ -202,11 +207,14 @@ def one_shot(gpt, task):
 
         acc_test_loss, nb_test_samples = 0, 0
         for input, targets in task.policy_batches(split="test"):
-            output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
+            output_gpt = gpt(mygpt.BracketedSequence(input), mode=args.oneshot_mode).x
             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(-1).mean()
-                + targets.xlogy(targets).sum(-1).mean()
+                -(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)
@@ -216,17 +224,24 @@ def one_shot(gpt, task):
         )
 
         # -------------------
-        input, targets = next(task.policy_batches(split="test"))
-        output_gpt = gpt(mygpt.BracketedSequence(input), with_readout=False).x
+        input = task.test_input[:32, : task.height * task.width]
+        targets = task.test_policies[:32]
+        output_gpt = gpt(mygpt.BracketedSequence(input), mode=args.oneshot_mode).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 = (-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 = (
+            (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(
-            os.path.join(args.result_dir, f"oneshot_{n_epoch:04d}.png"),
+            os.path.join(
+                args.result_dir, f"oneshot_{args.oneshot_mode}_{n_epoch:04d}.png"
+            ),
             mazes=input,
             score_paths=losses,
         )