X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=beaver.py;h=e7decd1a330a81e419f87d21ba61f191845aec47;hb=280f38363e34e202d38e6f7c00288329ab067a81;hp=afec61d4a506161e0da2e449d2dfa3445e386110;hpb=71a5d04a1decec9d71be93cb816a15a8c0de83a2;p=beaver.git

diff --git a/beaver.py b/beaver.py
index afec61d..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()
@@ -172,9 +174,9 @@ 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(args.dim_model, args.dim_model),
+        nn.Linear(dim_in, args.dim_model),
         nn.ReLU(),
         nn.Linear(args.dim_model, args.dim_model),
         nn.ReLU(),
@@ -187,7 +189,7 @@ def one_shot(gpt, task):
 
         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)
@@ -205,7 +207,7 @@ 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)
@@ -224,7 +226,7 @@ def one_shot(gpt, task):
         # -------------------
         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_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 * (input == maze.v_empty)
@@ -237,7 +239,9 @@ def one_shot(gpt, task):
         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,
         )