+ test_nb_total, test_nb_correct = self.compute_error(
+ model, "test", nb_to_use=1000
+ )
+ 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}%"
+ )
+
+ input = self.test_input[:48]
+ result = input.clone()
+ ar_mask = result.new_zeros(result.size())
+ ar_mask[:, self.height * self.width :] = 1
+ result *= 1 - ar_mask
+ masked_inplace_autoregression(
+ model, self.batch_size, result, ar_mask, device=self.device
+ )
+
+ mazes, paths = self.seq2map(input)
+ _, predicted_paths = self.seq2map(result)
+
+ filename = os.path.join(args.result_dir, f"result_{n_epoch:04d}.png")
+ maze.save_image(
+ filename,
+ mazes=mazes,
+ target_paths=paths,
+ predicted_paths=predicted_paths,
+ path_correct=maze.path_correctness(mazes, predicted_paths),
+ )
+ log_string(f"wrote {filename}")
+
+ model.train(t)
+
+
+######################################################################
+
+
+def generate_snake_sequences(
+ nb, height, width, nb_colors, length, device=torch.device("cpu")
+):
+ world = 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, 1), device=device)
+ result = 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 - 1) % 4,
+ snake_direction,
+ (snake_direction + 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.,4.,1.]], device=device)
+
+ # nb
+ i = torch.arange(val.size(0), device=device)
+ j = val.argmax(1)
+
+ # nb x 1
+ snake_direction = snake_next_direction[i[:, None], j[:, None]]
+
+ result[:, 2*l] = world[count, snake_position[:, 0], snake_position[:, 1]]
+ result[:, 2*l+1] = snake_direction[:,0]
+
+ # nb x 2
+ snake_position = snake_next_position[i[:, None], j[:, None]].squeeze(1)
+
+ return result
+
+generate_snake_sequences(nb=2, height=4, width=5, nb_colors=3, length=10)
+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 = generate_snake_sequences(
+ nb_train_samples, height, width, nb_colors, length, self.device
+ )
+ self.test_input = 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):