Update

diff --git a/beaver.py b/beaver.py
index dfbb7b6..c29dea5 100755 (executable)
--- a/beaver.py
+++ b/beaver.py
@@ -169,13 +169,50 @@ def compute_perplexity(model, split="train"):
 ######################################################################
 
 
+def nb_rank_error(output, targets):
+    output = output.reshape(-1, output.size(-1))
+    targets = targets.reshape(-1, targets.size(-1))
+    i = outputs.argmax(1)
+    # out=input.gather out[i][j]=input[i][index[i][j]]
+    # u[k]=targets[k][i[k]]
+    return output[targets.argmax(1)]
+
+
 def one_shot(gpt, task):
     t = gpt.training
     gpt.eval()
-    for input, targets in task.policy_batches():
-        output = gpt(mygpt.BracketedSequence(input), with_readout = False).x
+    model = nn.Linear(args.dim_model, 4).to(device)
+
+    for n_epoch in range(args.nb_epochs):
+        optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
+
+        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 = model(output_gpt)
+            loss = -(output.log_softmax(-1) * targets).sum(-1).mean()
+            acc_train_loss += loss.item() * input.size(0)
+            nb_train_samples += input.size(0)
+
+            optimizer.zero_grad()
+            loss.backward()
+            optimizer.step()
+
+        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 = model(output_gpt)
+            loss = -(output.log_softmax(-1) * targets).sum(-1).mean()
+            acc_test_loss += loss.item() * input.size(0)
+            nb_test_samples += input.size(0)
+
+        print(
+            f"{n_epoch=} {acc_train_loss/nb_train_samples=} {acc_test_loss/nb_test_samples=}"
+        )
+
     gpt.train(t)
 
+
 ######################################################################
 
 
@@ -226,7 +263,7 @@ class TaskMaze(Task):
             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
         )
         self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
-        self.train_policies = train_policies.to(device)
+        self.train_policies = train_policies.flatten(-2).permute(0, 2, 1).to(device)
 
         test_mazes, test_paths, test_policies = maze.create_maze_data(
             nb_test_samples,
@@ -236,7 +273,7 @@ class TaskMaze(Task):
             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
         )
         self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
-        self.test_policies = test_policies.to(device)
+        self.test_policies = test_policies.flatten(-2).permute(0, 2, 1).to(device)
 
         self.nb_codes = self.train_input.max() + 1
 
@@ -255,7 +292,7 @@ class TaskMaze(Task):
         input = self.train_input if split == "train" else self.test_input
         targets = self.train_policies if split == "train" else self.test_policies
         input = input[:, : self.height * self.width]
-        targets = targets.flatten(-2) * (input != maze.v_wall)[:,None]
+        targets = targets * (input != maze.v_wall)[:, :, None]
 
         if nb_to_use > 0:
             input = input[:nb_to_use]
@@ -390,8 +427,6 @@ else:
 
 ######################################################################
 
-nb_epochs = args.nb_epochs if args.nb_epochs > 0 else nb_epochs_default
-
 token_count = 0
 for input in task.batches(split="train"):
     token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
@@ -431,7 +466,7 @@ if args.one_shot:
 
 ##############################
 
-if nb_epochs_finished >= nb_epochs:
+if nb_epochs_finished >= args.nb_epochs:
     n_epoch = nb_epochs_finished
     train_perplexity = compute_perplexity(model, split="train")
     test_perplexity = compute_perplexity(model, split="test")
@@ -446,7 +481,7 @@ if nb_epochs_finished >= nb_epochs:
 
 ##############################
 
-for n_epoch in range(nb_epochs_finished, nb_epochs):
+for n_epoch in range(nb_epochs_finished, args.nb_epochs):
     learning_rate = learning_rate_schedule[n_epoch]
 
     log_string(f"learning_rate {learning_rate}")

diff --git a/maze.py b/maze.py
index d11ab6e..6c3fe94 100755 (executable)
--- a/maze.py
+++ b/maze.py
@@ -158,11 +158,11 @@ def create_maze_data(
 ):
     mazes = torch.empty(nb, height, width, dtype=torch.int64)
     paths = torch.empty(nb, height, width, dtype=torch.int64)
-    policies = torch.empty(nb, 4, height, width, dtype=torch.int64)
+    policies = torch.empty(nb, 4, height, width)
 
     for n in progress_bar(range(nb)):
         maze = create_maze(height, width, nb_walls)
-        i = (1 - maze).nonzero()
+        i = (maze == v_empty).nonzero()
         while True:
             start, goal = i[torch.randperm(i.size(0))[:2]]
             if (start - goal).abs().sum() >= dist_min:

diff --git a/mygpt.py b/mygpt.py
index d424eef..bd79676 100755 (executable)
--- a/mygpt.py
+++ b/mygpt.py
@@ -246,11 +246,12 @@ class MyGPT(nn.Module):
                     m.bias.zero_()
                     m.weight.fill_(1.0)
 
-    def forward(self, bs, with_readout = True):
+    def forward(self, bs, with_readout=True):
         bs.x = F.pad(bs.x, (1, -1))
         bs = self.embedding(bs)
         bs = self.trunk(bs)
-        if with_readout: bs = self.readout(bs)
+        if with_readout:
+            bs = self.readout(bs)
         return bs