+ 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(),
+ 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), 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)
+ # loss = (output.softmax(-1) - targets).abs().max(-1).values.mean()
+ loss = (
+ -(output.log_softmax(-1) * 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), 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)
+ # loss = (output.softmax(-1) - targets).abs().max(-1).values.mean()
+ loss = (
+ -(output.log_softmax(-1) * 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), 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)
+ # losses = losses / losses.max()
+ # losses = (output.softmax(-1) - targets).abs().max(-1).values
+ # losses = (losses >= 0.05).float()
+ losses = (
+ (F.one_hot(output.argmax(-1), num_classes=4) * targets).sum(-1) == 0
+ ).float()
+ 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_{args.oneshot_mode}_{n_epoch:04d}.png"
+ ),
+ mazes=input,
+ score_paths=losses,
+ )
+ # -------------------
+