# logger(f"wrote {filename}")
-######################################################################
-
-import world
-
-
-class World(Task):
- def __init__(
- self,
- nb_train_samples,
- nb_test_samples,
- batch_size,
- logger=None,
- device=torch.device("cpu"),
- ):
- super().__init__()
-
- self.batch_size = batch_size
- self.device = device
- self.height = 6
- self.width = 8
-
- self.train_input = world.generate(
- nb_train_samples, height=self.height, width=self.width
- )
- self.train_ar_mask = (
- (torch.arange(self.train_input.size(1)) > self.train_input.size(1) // 2)
- .long()[None, :]
- .expand_as(self.train_input)
- )
-
- self.test_input = world.generate(
- nb_test_samples, height=self.height, width=self.width
- )
- self.test_ar_mask = (
- (torch.arange(self.test_input.size(1)) > self.test_input.size(1) // 2)
- .long()[None, :]
- .expand_as(self.test_input)
- )
-
- self.train_input, self.train_ar_mask = self.train_input.to(
- device
- ), self.train_ar_mask.to(device)
- self.test_input, self.test_ar_mask = self.test_input.to(
- device
- ), self.test_ar_mask.to(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, result_dir, logger, deterministic_synthesis, nmax=1000
- ):
- def compute_accuracy(input, ar_mask, logger=None):
- input, ar_mask = input[:nmax], ar_mask[:nmax]
- result = input.clone() * (1 - ar_mask)
-
- masked_inplace_autoregression(
- model,
- self.batch_size,
- result,
- ar_mask,
- deterministic_synthesis,
- progress_bar_desc=None,
- device=self.device,
- )
-
- nb_total, nb_correct = (
- input.size(0),
- (input == result).long().min(dim=1).values.sum(),
- )
-
- return nb_total, nb_correct
-
- train_nb_total, train_nb_correct = compute_accuracy(
- self.train_input, self.train_ar_mask
- )
-
- logger(
- f"accuracy_train {n_epoch} 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_accuracy(
- self.test_input, self.test_ar_mask, logger
- )
-
- logger(
- f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
- )
-
- logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
-
- if save_attention_image is not None:
- for k in range(10):
- ns = torch.randint(self.test_input.size(0), (1,)).item()
- input = self.test_input[ns : ns + 1].clone()
-
- with torch.autograd.no_grad():
- t = model.training
- model.eval()
- # model.record_attention(True)
- model(BracketedSequence(input))
- model.train(t)
- # ram = model.retrieve_attention()
- # model.record_attention(False)
-
- # tokens_output = [c for c in self.problem.seq2str(input[0])]
- # tokens_input = ["n/a"] + tokens_output[:-1]
- # for n_head in range(ram[0].size(1)):
- # filename = os.path.join(
- # result_dir, f"sandbox_attention_{k}_h{n_head}.pdf"
- # )
- # attention_matrices = [m[0, n_head] for m in ram]
- # save_attention_image(
- # filename,
- # tokens_input,
- # tokens_output,
- # attention_matrices,
- # k_top=10,
- ##min_total_attention=0.9,
- # token_gap=12,
- # layer_gap=50,
- # )
- # logger(f"wrote {filename}")
-
-
######################################################################
import picoclvr
######################################################################
+######################################################################
+
+import world
+
+
+class World(Task):
+ def save_image(self, input, result_dir, filename, logger):
+ img = world.sample2img(self.train_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}")
+
+ def __init__(
+ self,
+ nb_train_samples,
+ nb_test_samples,
+ batch_size,
+ result_dir=None,
+ logger=None,
+ device=torch.device("cpu"),
+ ):
+ super().__init__()
+
+ self.batch_size = batch_size
+ self.device = device
+ self.height = 6
+ self.width = 8
+
+ self.train_input = world.generate(
+ nb_train_samples, height=self.height, width=self.width
+ )
+ self.train_ar_mask = (
+ (torch.arange(self.train_input.size(1)) > self.train_input.size(1) // 2)
+ .long()[None, :]
+ .expand_as(self.train_input)
+ )
+
+ self.test_input = world.generate(
+ nb_test_samples, height=self.height, width=self.width
+ )
+ self.test_ar_mask = (
+ (torch.arange(self.test_input.size(1)) > self.test_input.size(1) // 2)
+ .long()[None, :]
+ .expand_as(self.test_input)
+ )
+
+ self.train_input, self.train_ar_mask = self.train_input.to(
+ device
+ ), self.train_ar_mask.to(device)
+ self.test_input, self.test_ar_mask = self.test_input.to(
+ device
+ ), self.test_ar_mask.to(device)
+
+ self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
+
+ if result_dir is not None:
+ self.save_image(
+ self.train_input[:96], result_dir, f"world_train.png", logger
+ )
+
+ 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, result_dir, logger, deterministic_synthesis, nmax=1000
+ ):
+ def compute_accuracy(input, ar_mask, logger=None):
+ input, ar_mask = input[:nmax], ar_mask[:nmax]
+ result = input.clone() * (1 - ar_mask)
+
+ masked_inplace_autoregression(
+ model,
+ self.batch_size,
+ result,
+ ar_mask,
+ deterministic_synthesis,
+ progress_bar_desc=None,
+ device=self.device,
+ )
+
+ nb_total, nb_correct = (
+ input.size(0),
+ (input == result).long().min(dim=1).values.sum(),
+ )
+
+ return nb_total, nb_correct
+
+ train_nb_total, train_nb_correct = compute_accuracy(
+ self.train_input, self.train_ar_mask
+ )
+
+ logger(
+ f"accuracy_train {n_epoch} 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_accuracy(
+ self.test_input, self.test_ar_mask, logger
+ )
+
+ logger(
+ f"accuracy_test {n_epoch} nb_total {test_nb_total} nb_correct {test_nb_correct} accuracy {(100.0*test_nb_correct)/test_nb_total:.02f}%"
+ )
+
+ logger(f"main_test_accuracy {n_epoch} {test_nb_correct/test_nb_total}")
+
+ ##############################
+
+ input, ar_mask = self.test_input[:96], self.test_ar_mask[:96]
+ result = input.clone() * (1 - ar_mask)
+
+ masked_inplace_autoregression(
+ model,
+ self.batch_size,
+ result,
+ ar_mask,
+ deterministic_synthesis,
+ progress_bar_desc=None,
+ device=self.device,
+ )
+
+ self.save_image(result, result_dir, f"world_result_{n_epoch:04d}.png", logger)
+
+ def store_new_problems(self, new_problems):
+ nb_current = self.train_input.size(0)
+ nb_new = new_problems.size(0)
+ if nb_new >= nb_current:
+ self.train_input[...] = new_problems[:nb_current]
+ else:
+ nb_kept = nb_current - nb_new
+ self.train_input[:nb_kept] = self.train_input[-nb_kept:].clone()
+ self.train_input[nb_kept:] = new_problems
+
+ def create_new_problems(self, n_epoch, result_dir, logger, nb, model, nb_runs):
+ new_problems = torch.empty(
+ nb, self.height * self.width * 2 + 1, device=self.device, dtype=torch.int64
+ )
+ ar_mask = torch.full(new_problems.size(), 1, device=self.device)
+
+ masked_inplace_autoregression(
+ model,
+ self.batch_size,
+ new_problems,
+ ar_mask,
+ deterministic_synthesis=False,
+ progress_bar_desc="new problems",
+ device=self.device,
+ )
+
+ nb_correct = torch.empty(nb, device=self.device, dtype=torch.int64)
+
+ for n in tqdm.tqdm(
+ range(new_problems.size(0)), dynamic_ncols=True, desc="checking problems"
+ ):
+ result = new_problems[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)
+ )
+
+ masked_inplace_autoregression(
+ model,
+ self.batch_size,
+ result,
+ ar_mask,
+ deterministic_synthesis=False,
+ progress_bar_desc=None,
+ device=self.device,
+ )
+
+ nb_correct[n] = (
+ (new_problems[n][None, :] == result).long().min(dim=1).values.sum()
+ )
+
+ return new_problems, nb_correct