Update

diff --git a/beaver.py b/beaver.py
index de38ff4..e7decd1 100755 (executable)
--- 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()
@@ -172,8 +174,7 @@ def compute_perplexity(model, split="train"):
 def one_shot(gpt, task):
     t = gpt.training
     gpt.eval()
-    mode='head'
-    dim_in=args.dim_model * (args.nb_blocks * 2 if mode=='deep' else 1)
+    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(),
@@ -188,7 +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), mode=mode).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)
@@ -206,7 +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), mode=mode).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)
@@ -225,7 +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), mode=mode).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)
@@ -238,7 +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,
         )

diff --git a/mygpt.py b/mygpt.py
index a1db2e3..0b63ac8 100755 (executable)
--- a/mygpt.py
+++ b/mygpt.py
@@ -246,19 +246,19 @@ class MyGPT(nn.Module):
                     m.bias.zero_()
                     m.weight.fill_(1.0)
 
-    def forward(self, bs, mode='standard'):
+    def forward(self, bs, mode="standard"):
         bs.x = F.pad(bs.x, (1, -1))
         bs = self.embedding(bs)
-        if mode=='standard':
+        if mode == "standard":
             bs = self.trunk(bs)
             bs = self.readout(bs)
-        elif mode=='head':
+        elif mode == "head":
             bs = self.trunk(bs)
-        elif mode=='deep':
+        elif mode == "deep":
             r = []
             for l in self.trunk:
                 bs = l(bs)
-                r += [ bs.slice() ]
+                r += [bs.slice()]
             bs = BracketedSequence(torch.cat(r, -1))
         else:
             raise ValueError