parser.add_argument("--maze_nb_walls", type=int, default=15)
+parser.add_argument("--oneshot_mode", type=str, default="head")
+
######################################################################
args = parser.parse_args()
def one_shot(gpt, task):
- pass
+ t = gpt.training
+ gpt.eval()
+ 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,
+ )
+ # -------------------
+
+ gpt.train(t)
######################################################################
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
):
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
_, 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)
######################################################################
-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))
##############################
-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")
##############################
-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}")