class World(Task):
def save_image(self, input, result_dir, filename, logger):
- img = world.sample2img(self.train_input.to("cpu"), self.height, self.width)
+ img = world.sample2img(input.to("cpu"), self.height, self.width)
image_name = os.path.join(result_dir, filename)
torchvision.utils.save_image(img.float() / 255.0, image_name, nrow=8, padding=2)
logger(f"wrote {image_name}")
device=self.device,
)
- self.save_image(result, result_dir, f"world_result_{n_epoch:04d}.png", logger)
+ self.save_image(
+ result[:96], result_dir, f"world_result_{n_epoch:04d}.png", logger
+ )
return main_test_accuracy
input[:nb_kept] = input[-nb_kept:].clone()
input[nb_kept:] = new_quizzes
- def create_new_quizzes(self, n_epoch, result_dir, logger, nb, model, nb_runs):
+ def create_new_quizzes(
+ self, n_epoch, result_dir, logger, nb, models, other_models, nb_runs
+ ):
new_quizzes = torch.empty(
nb, self.height * self.width * 2 + 1, device=self.device, dtype=torch.int64
)
device=self.device,
)
- nb_correct = torch.empty(nb, device=self.device, dtype=torch.int64)
+ input = (
+ new_quizzes[:, None, :]
+ .expand(-1, nb_runs, -1)
+ .clone()
+ .reshape(-1, new_quizzes.size(-1))
+ )
+ result = input.clone()
- for n in tqdm.tqdm(
- range(new_quizzes.size(0)), dynamic_ncols=True, desc="checking quizzes"
- ):
- result = new_quizzes[n][None, :].expand(nb_runs, -1).clone()
- ar_mask = (
- (torch.arange(result.size(1), device=self.device) > result.size(1) // 2)
- .long()[None, :]
- .expand_as(result)
- )
+ ar_mask = (
+ (torch.arange(result.size(1), device=self.device) > result.size(1) // 2)
+ .long()[None, :]
+ .expand_as(result)
+ )
+ dispatch = torch.randint(len(other_models), (result.size(0),))
+
+ for n, m in enumerate(other_models):
masked_inplace_autoregression(
- model,
+ m,
self.batch_size,
- result,
- ar_mask,
+ result[dispatch == n],
+ ar_mask[dispatch == n],
deterministic_synthesis=False,
progress_bar_desc=None,
device=self.device,
)
- nb_correct[n] = (
- (new_quizzes[n][None, :] == result).long().min(dim=1).values.sum()
- )
+ nb_correct = (
+ (input == result).long().min(dim=-1).values.reshape(-1, nb_runs).sum(dim=-1)
+ )
return new_quizzes, nb_correct