parser.add_argument("--maze_nb_walls", type=int, default=15)
+##############################
+# Snake options
+
+parser.add_argument("--snake_height", type=int, default=6)
+
+parser.add_argument("--snake_width", type=int, default=8)
+
+parser.add_argument("--snake_nb_colors", type=int, default=3)
+
+parser.add_argument("--snake_length", type=int, default=100)
+
######################################################################
args = parser.parse_args()
0,
)
- image_name = os.path.join(args.result_dir, f"result_{n_epoch:04d}.png")
+ image_name = os.path.join(args.result_dir, f"picoclvr_result_{n_epoch:04d}.png")
torchvision.utils.save_image(
img / 255.0, image_name, nrow=nb_per_primer, padding=1, pad_value=1.0
)
log_string(f"wrote {image_name}")
+######################################################################
+
+
+class TaskMNIST(Task):
+ def __init__(self, batch_size, device=torch.device("cpu")):
+ self.device = device
+ self.batch_size = batch_size
+
+ def batches(self, split="train"):
+ assert split in {"train", "test"}
+ data_set = torchvision.datasets.MNIST(
+ root="./data", train=(split == "train"), download=True
+ )
+ data_input = data_set.data.view(-1, 28 * 28).long()
+ if args.nb_train_samples is not None:
+ data_input = data_input[: args.nb_train_samples]
+ for batch in tqdm.tqdm(
+ data_input.split(self.batch_size), desc=f"epoch-{split}"
+ ):
+ yield batch
+
+ def vocabulary_size(self):
+ return 256
+
+ def produce_results(self, n_epoch, model):
+ results = torch.empty(64, 28 * 28, device=self.device, dtype=torch.int64)
+ ar_mask = torch.full_like(results, 1)
+ masked_inplace_autoregression(
+ model, self.batch_size, results, ar_mask, device=self.device
+ )
+ image_name = os.path.join(args.result_dir, f"result_mnist_{n_epoch:04d}.png")
+ torchvision.utils.save_image(
+ 1 - results.reshape(-1, 1, 28, 28) / 255.0,
+ image_name,
+ nrow=16,
+ pad_value=0.8,
+ )
+ log_string(f"wrote {image_name}")
+
+
######################################################################
import maze
self.width = width
self.device = device
- train_mazes, train_paths, train_policies = maze.create_maze_data(
+ train_mazes, train_paths, _ = maze.create_maze_data(
nb_train_samples,
height=height,
width=width,
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.flatten(-2).to(device)
- test_mazes, test_paths, test_policies = maze.create_maze_data(
+ test_mazes, test_paths, _ = maze.create_maze_data(
nb_test_samples,
height=height,
width=width,
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.flatten(-2).to(device)
- self.nb_codes = self.train_input.max() + 1
+ self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
def batches(self, split="train", nb_to_use=-1, desc=None):
assert split in {"train", "test"}
):
yield batch
- def policy_batches(self, split="train", nb_to_use=-1, desc=None):
- assert split in {"train", "test"}
- input = self.train_input if split == "train" else self.test_input
- policies = self.train_policies if split == "train" else self.test_policies
- input = input[:, : self.height * self.width]
- policies = policies * (input != maze.v_wall)[:, None]
-
- if nb_to_use > 0:
- input = input[:nb_to_use]
- policies = policies[:nb_to_use]
-
- if desc is None:
- desc = f"epoch-{split}"
- for batch in tqdm.tqdm(
- zip(input.split(self.batch_size), policies.split(self.batch_size)),
- dynamic_ncols=True,
- desc=desc,
- ):
- yield batch
-
def vocabulary_size(self):
return self.nb_codes
mazes, paths = self.seq2map(input)
_, predicted_paths = self.seq2map(result)
- filename = f"result_{n_epoch:04d}.png"
+
+ filename = os.path.join(args.result_dir, f"result_{n_epoch:04d}.png")
maze.save_image(
- os.path.join(args.result_dir, filename),
+ filename,
mazes=mazes,
target_paths=paths,
predicted_paths=predicted_paths,
######################################################################
+def generate_snake_sequences(
+ nb, height, width, nb_colors, length, device=torch.device("cpu")
+):
+ worlds = torch.randint(nb_colors, (nb, height, width), device=device)
+ # nb x 2
+ snake_position = torch.cat(
+ (
+ torch.randint(height, (nb, 1), device=device),
+ torch.randint(width, (nb, 1), device=device),
+ ),
+ 1,
+ )
+ snake_direction = torch.randint(4, (nb,), device=device)
+ sequences = torch.empty(nb, 2 * length, device=device, dtype=torch.int64)
+ count = torch.arange(nb, device=device) # [:,None]
+
+ for l in range(length):
+ # nb x 3
+ snake_next_direction = torch.cat(
+ (
+ (snake_direction[:, None] - 1) % 4,
+ snake_direction[:, None],
+ (snake_direction[:, None] + 1) % 4,
+ ),
+ 1,
+ )
+
+ # nb x 3
+ vh = (snake_next_direction + 1) % 2 * (snake_next_direction - 1)
+ vw = snake_next_direction % 2 * (snake_next_direction - 2)
+
+ # nb x 3 x 2
+ snake_next_speed = torch.cat((vh[:, :, None], vw[:, :, None]), 2)
+ snake_next_position = snake_position[:, None, :] + snake_next_speed
+
+ # nb x 3
+ val = torch.logical_and(
+ torch.logical_and(
+ snake_next_position[:, :, 0] >= 0, snake_next_position[:, :, 0] < height
+ ),
+ torch.logical_and(
+ snake_next_position[:, :, 1] >= 0, snake_next_position[:, :, 1] < width
+ ),
+ ).float()
+ val = (
+ torch.rand_like(val) * val * torch.tensor([[1.0, 4.0, 1.0]], device=device)
+ )
+
+ # nb
+ i = torch.arange(val.size(0), device=device)
+ j = val.argmax(1)
+ snake_direction = snake_next_direction[i, j]
+
+ sequences[:, 2 * l] = worlds[count, snake_position[:, 0], snake_position[:, 1]]
+ sequences[:, 2 * l + 1] = snake_direction
+
+ # nb x 2
+ snake_position = snake_next_position[i, j]
+
+ return sequences, worlds
+
+ # print(snake_position)
+
+
+# generate_snake_sequences(nb=1, height=4, width=6, nb_colors=3, length=20)
+# exit(0)
+
+
+class TaskSnake(Task):
+ def __init__(
+ self,
+ nb_train_samples,
+ nb_test_samples,
+ batch_size,
+ height,
+ width,
+ nb_colors,
+ length,
+ device=torch.device("cpu"),
+ ):
+ self.batch_size = batch_size
+ self.height = height
+ self.width = width
+ self.device = device
+
+ self.train_input, self.train_worlds = generate_snake_sequences(
+ nb_train_samples, height, width, nb_colors, length, self.device
+ )
+ self.test_input, self.test_worlds = generate_snake_sequences(
+ nb_test_samples, height, width, nb_colors, length, self.device
+ )
+
+ self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
+
+ def batches(self, split="train", nb_to_use=-1, desc=None):
+ assert split in {"train", "test"}
+ input = self.train_input if split == "train" else self.test_input
+ if nb_to_use > 0:
+ input = input[:nb_to_use]
+ if desc is None:
+ desc = f"epoch-{split}"
+ for batch in tqdm.tqdm(
+ input.split(self.batch_size), dynamic_ncols=True, desc=desc
+ ):
+ yield batch
+
+ def vocabulary_size(self):
+ return self.nb_codes
+
+
+######################################################################
+
+
def picoclvr_pruner_horizontal_green(p):
return not ("green" in p and ("left" in p or "right" in p))
pruner_eval=picoclvr_pruner_eval,
)
+elif args.task == "mnist":
+ task = TaskMNIST(
+ batch_size=args.batch_size,
+ device=device,
+ )
+
elif args.task == "maze":
task = TaskMaze(
nb_train_samples=args.nb_train_samples,
device=device,
)
+elif args.task == "snake":
+ task = TaskSnake(
+ nb_train_samples=args.nb_train_samples,
+ nb_test_samples=args.nb_test_samples,
+ batch_size=args.batch_size,
+ height=args.snake_height,
+ width=args.snake_width,
+ nb_colors=args.snake_nb_colors,
+ length=args.snake_length,
+ device=device,
+ )
+
else:
raise ValueError(f"Unknown task {args.task}")
projects / picoclvr.git / blobdiff