Update

[beaver.git] / beaver.py

diff --git a/beaver.py b/beaver.py
index dfbb7b6..c3b7e09 100755 (executable)
--- a/beaver.py
+++ b/beaver.py
@@ -172,10 +172,74 @@ def compute_perplexity(model, split="train"):
 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.Sequential(
+        nn.Linear(args.dim_model, args.dim_model),
+        nn.ReLU(),
+        nn.Linear(args.dim_model, args.dim_model),
+        nn.ReLU(),
+        nn.Linear(args.dim_model, 4),
+    ).to(device)
+
+    for n_epoch in range(args.nb_epochs):
+        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 = model(output_gpt)
+            targets = targets * (input.unsqueeze(-1) == maze.v_empty)
+            output = output * (input.unsqueeze(-1) == maze.v_empty)
+            loss = (
+                -(output.log_softmax(-1) * targets).sum()
+                / (input == maze.v_empty).sum()
+                + targets.xlogy(targets).sum() / (input == maze.v_empty).sum()
+            )
+            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)
+            targets = targets * (input.unsqueeze(-1) == maze.v_empty)
+            output = output * (input.unsqueeze(-1) == maze.v_empty)
+            loss = (
+                -(output.log_softmax(-1) * targets).sum()
+                / (input == maze.v_empty).sum()
+                + targets.xlogy(targets).sum() / (input == maze.v_empty).sum()
+            )
+            acc_test_loss += loss.item() * input.size(0)
+            nb_test_samples += input.size(0)
+
+        log_string(
+            f"diff_ce {n_epoch} train {acc_train_loss/nb_train_samples} test {acc_test_loss/nb_test_samples}"
+        )
+
+        # -------------------
+        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 = model(output_gpt)
+        losses = (-output.log_softmax(-1) * targets + targets.xlogy(targets)).sum(-1)
+        losses = losses * (input == maze.v_empty)
+        losses = losses / losses.max()
+        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"),
+            mazes=input,
+            score_paths=losses,
+        )
+        # -------------------
+
     gpt.train(t)
 
+
 ######################################################################
 
 
@@ -226,7 +290,7 @@ 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 +300,7 @@ 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 +319,7 @@ 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]
@@ -315,10 +379,10 @@ class TaskMaze(Task):
             _, predicted_paths = self.seq2map(result)
             maze.save_image(
                 os.path.join(args.result_dir, f"result_{n_epoch:04d}.png"),
-                mazes,
-                paths,
-                predicted_paths,
-                maze.path_correctness(mazes, predicted_paths),
+                mazes=mazes,
+                target_paths=paths,
+                predicted_paths=predicted_paths,
+                path_correct=maze.path_correctness(mazes, predicted_paths),
             )
 
             model.train(t)
@@ -390,8 +454,6 @@ 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 +493,7 @@ 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 +508,7 @@ 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}")