Update

[beaver.git] / beaver.py

diff --git a/beaver.py b/beaver.py
index 4f694da..c29dea5 100755 (executable)
--- a/beaver.py
+++ b/beaver.py
@@ -169,8 +169,48 @@ 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):
-    pass
+    t = gpt.training
+    gpt.eval()
+    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)
 
 
 ######################################################################
@@ -215,25 +255,25 @@ class TaskMaze(Task):
         self.width = width
         self.device = device
 
-        mazes_train, paths_train = maze.create_maze_data(
+        train_mazes, train_paths, train_policies = maze.create_maze_data(
             nb_train_samples,
             height=height,
             width=width,
             nb_walls=nb_walls,
             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-train"),
         )
-        mazes_train, paths_train = mazes_train.to(device), paths_train.to(device)
-        self.train_input = self.map2seq(mazes_train, paths_train)
+        self.train_input = self.map2seq(train_mazes.to(device), train_paths.to(device))
+        self.train_policies = train_policies.flatten(-2).permute(0, 2, 1).to(device)
 
-        mazes_test, paths_test = maze.create_maze_data(
+        test_mazes, test_paths, test_policies = maze.create_maze_data(
             nb_test_samples,
             height=height,
             width=width,
             nb_walls=nb_walls,
             progress_bar=lambda x: tqdm.tqdm(x, dynamic_ncols=True, desc=f"data-test"),
         )
-        mazes_test, paths_test = mazes_test.to(device), paths_test.to(device)
-        self.test_input = self.map2seq(mazes_test, paths_test)
+        self.test_input = self.map2seq(test_mazes.to(device), test_paths.to(device))
+        self.test_policies = test_policies.flatten(-2).permute(0, 2, 1).to(device)
 
         self.nb_codes = self.train_input.max() + 1
 
@@ -247,6 +287,24 @@ class TaskMaze(Task):
         ):
             yield batch
 
+    def policy_batches(self, split="train", nb_to_use=-1):
+        assert split in {"train", "test"}
+        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 * (input != maze.v_wall)[:, :, None]
+
+        if nb_to_use > 0:
+            input = input[:nb_to_use]
+            targets = targets[:nb_to_use]
+
+        for batch in tqdm.tqdm(
+            zip(input.split(self.batch_size), targets.split(self.batch_size)),
+            dynamic_ncols=True,
+            desc=f"epoch-{split}",
+        ):
+            yield batch
+
     def vocabulary_size(self):
         return self.nb_codes
 
@@ -369,8 +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))
@@ -410,7 +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")
@@ -425,7 +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}")