+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_prior_visits = torch.zeros(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)
+ sequences_prior_visits = torch.zeros(
+ nb, 2 * length, device=device, dtype=torch.int64
+ )
+ i = 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 = (
+ # The multiplicative factors bias toward moving forward
+ torch.rand_like(val)
+ * val
+ * torch.tensor([[1.0, 2.0, 1.0]], device=device)
+ )
+
+ # nb
+ j = val.argmax(1)
+ snake_direction = snake_next_direction[i, j]
+
+ sequences[:, 2 * l] = worlds[i, snake_position[:, 0], snake_position[:, 1]] + 4
+ sequences_prior_visits[:, 2 * l] = nb_prior_visits[
+ i, snake_position[:, 0], snake_position[:, 1]
+ ]
+ nb_prior_visits[i, snake_position[:, 0], snake_position[:, 1]] += 1
+ sequences[:, 2 * l + 1] = snake_direction
+
+ # nb x 2
+ snake_position = snake_next_position[i, j]
+
+ return sequences, sequences_prior_visits
+
+
+# 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_prior_visits = generate_snake_sequences(
+ nb_train_samples, height, width, nb_colors, length, self.device
+ )
+ self.test_input, self.test_prior_visits = 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 produce_results(self, n_epoch, model):
+ with torch.autograd.no_grad():
+ t = model.training
+ model.eval()
+
+ def compute_nb_correct(input, prior_visits):
+ result = input.clone()
+ i = torch.arange(result.size(1), device=result.device)[None, :]
+ ar_mask = torch.logical_and(i >= i.size(0) // 2, i % 2 == 0).long()
+ result *= 1 - ar_mask
+ masked_inplace_autoregression(
+ model, self.batch_size, result, ar_mask, device=self.device
+ )
+
+ nb_total = (
+ (prior_visits > 0) * ar_mask
+ ).sum()
+
+ nb_correct = (
+ (result == input).long() * (prior_visits > 0) * ar_mask
+ ).sum()
+
+ # nb_total = result.size(0)
+ # nb_correct = ((result - input).abs().sum(1) == 0).sum()
+
+ return nb_total, nb_correct
+
+ train_nb_total, train_nb_correct = compute_nb_correct(
+ self.train_input, self.train_prior_visits
+ )
+
+ log_string(
+ f"accuracy_train nb_total {train_nb_total} nb_correct {train_nb_correct} accuracy {(100.0*train_nb_correct)/train_nb_total:.02f}%"
+ )
+
+ test_nb_total, test_nb_correct = compute_nb_correct(
+ self.test_input, self.test_prior_visits
+ )
+
+ log_string(
+ f"accuracy_test nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
+ )
+
+ model.train(t)
+
+
+######################################################################
+
+