b = torch.arange(input.size(1), device=input.device) > input.size(1) // 2
return b.long()[None, :].expand_as(input)
+ def generate_token_sequences(self, nb):
+ prompts, answers = self.problem.generate_prompts_and_answers(nb)
+ result = []
+
+ for prompt, answer in zip(prompts, answers):
+ if torch.rand(1) < 0.5:
+ a = [torch.tensor([self.token_forward]), prompt, answer]
+ else:
+ a = [torch.tensor([self.token_backward]), answer, prompt]
+
+ result.append(torch.cat(a, dim=0)[None, :])
+
+ return torch.cat(result, dim=0)
+
def __init__(
self,
problem,
):
super().__init__()
+ v = problem.nb_token_values()
+ self.token_forward = v
+ self.token_backward = v + 1
+ self.nb_token_values = v + 2
+
self.problem = problem
self.batch_size = batch_size
self.device = device
self.logger = logger
- self.train_w_quizzes = self.problem.generate_token_sequences(
- nb_train_samples
- ).to(device)
- self.test_w_quizzes = self.problem.generate_token_sequences(nb_test_samples).to(
+ self.train_w_quizzes = self.generate_token_sequences(nb_train_samples).to(
device
)
- self.nb_codes = max(self.train_w_quizzes.max(), self.test_w_quizzes.max()) + 1
+ self.test_w_quizzes = self.generate_token_sequences(nb_test_samples).to(device)
self.train_c_quizzes = []
self.test_c_quizzes = []
if result_dir is not None:
- self.problem.save_quizzes(
- self.train_w_quizzes[:72], result_dir, "culture_w_quizzes"
+ self.save_quizzes(
+ result_dir, "culture_w_quizzes", self.train_w_quizzes[:72]
)
+ def save_quizzes(self, result_dir, filename_prefix, quizzes, prediction=False):
+ print(f"DEBUG {quizzes.size()=}")
+ l = (quizzes.size(1) - 1) // 2
+ forward = (quizzes[:, 0] == self.token_forward).long()
+ backward = (quizzes[:, 0] == self.token_backward).long()
+ assert forward.equal(1 - backward)
+ first = quizzes[:, 1 : 1 + l]
+ second = quizzes[:, 1 + l : 1 + 2 * l]
+ prompts = forward[:, None] * first + backward[:, None] * second
+ answers = forward[:, None] * second + backward[:, None] * first
+
+ if prediction:
+ predicted_prompts = backward
+ predicted_answers = forward
+ else:
+ predicted_prompts = None
+ predicted_answers = None
+
+ self.problem.save_quizzes(
+ result_dir,
+ filename_prefix,
+ prompts,
+ answers,
+ predicted_prompts,
+ predicted_answers,
+ )
+
def batches(self, split="train", desc=None):
assert split in {"train", "test"}
if split == "train":
yield batch
def vocabulary_size(self):
- return self.nb_codes
+ return self.nb_token_values
def produce_results(
self, n_epoch, model, result_dir, deterministic_synthesis, nmax=1000
device=self.device,
)
- self.problem.save_quizzes(
- result[:72], result_dir, f"culture_prediction_{n_epoch:04d}_{model.id:02d}"
+ self.save_quizzes(
+ result_dir,
+ f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
+ quizzes=result[:72],
+ prediction=True,
)
return main_test_accuracy
input = self.train_w_quizzes if for_train else self.test_w_quizzes
nb = min(nb, input.size(0))
input[:-nb] = input[nb:].clone()
- input[-nb:] = self.problem.generate_token_sequences(nb).to(self.device)
+ input[-nb:] = self.generate_token_sequences(nb).to(self.device)
def store_c_quizzes(self, new_c_quizzes, for_train=True):
if for_train:
self.test_c_quizzes.append(new_c_quizzes)
def reverse_time(self, c_quizzes):
- token_forward, token_backward = self.problem.direction_tokens()
-
l = (c_quizzes.size(1) - 1) // 2
- direction = c_quizzes[:, l : l + 1]
- direction = self.problem.token_forward * (
- direction == self.problem.token_backward
- ) + self.problem.token_backward * (direction == self.problem.token_forward)
+ direction = c_quizzes[:, 0:1]
+ direction = self.token_forward * (
+ direction == self.token_backward
+ ) + self.token_backward * (direction == self.token_forward)
- return torch.cat([c_quizzes[:, l + 1 :], direction, c_quizzes[:, :l]], dim=1)
+ return torch.cat(
+ [direction, c_quizzes[:, l + 1 :], c_quizzes[:, 1 : l + 1]], dim=1
+ )
def compute_correctness(
- self, c_quizzes, models_for_validation, both_direction=True
+ self, c_quizzes, models_for_validation, both_directions=True
):
reversed_c_quizzes = self.reverse_time(c_quizzes)
correct = (c_quizzes == result).long().min(dim=-1).values
- if both_direction:
+ if both_directions:
reversed_result = reversed_c_quizzes.clone()
masked_inplace_autoregression(