class QuizzMachine:
- def make_ar_mask(self, input, first, nb):
- i = torch.arange(input.size(1), device=input.device)
- b = torch.logical_and(i >= first, i < first + nb)
- return b.long()[None, :].expand_as(input)
+ def indices_forward_and_backward(self, quizzes):
+ i_forward = quizzes[:, 0] == self.token_forward
+ j_forward = quizzes[:, 1 + self.prompt_len] == self.token_forward
+ i_backward = quizzes[:, 0] == self.token_backward
+ j_backward = quizzes[:, 1 + self.answer_len] == self.token_backward
+ assert torch.logical_or(
+ torch.logical_and(i_forward, j_forward),
+ torch.logical_and(i_backward, j_backward),
+ ).all()
+ return i_forward, i_backward
+
+ def reverse_time(self, quizzes):
+ i_forward, i_backward = self.indices_forward_and_backward(quizzes)
+
+ forward_to_backward = torch.cat(
+ [
+ quizzes[:, 0:1],
+ quizzes[:, 2 + self.prompt_len : 2 + self.prompt_len + self.answer_len],
+ quizzes[:, 1 + self.prompt_len : 1 + self.prompt_len + 1],
+ quizzes[:, 1 : 1 + self.prompt_len],
+ ],
+ dim=1,
+ )
+
+ forward_to_backward[:, 0] = self.token_backward
+ forward_to_backward[:, 1 + self.answer_len] = self.token_backward
+
+ backward_to_forward = torch.cat(
+ [
+ quizzes[:, 0:1],
+ quizzes[:, 2 + self.answer_len :],
+ quizzes[:, 1 + self.answer_len : 2 + self.answer_len],
+ quizzes[:, 1 : 1 + self.answer_len],
+ ],
+ dim=1,
+ )
+
+ backward_to_forward[:, 0] = self.token_forward
+ backward_to_forward[:, 1 + self.prompt_len] = self.token_forward
+
+ m = i_forward.long()[:, None]
+
+ return m * forward_to_backward + (1 - m) * backward_to_forward
+
+ def make_ar_mask(self, quizzes, first=False):
+ i_forward, i_backward = self.indices_forward_and_backward(quizzes)
+
+ t = torch.arange(quizzes.size(1), device=quizzes.device)
+
+ if first:
+ m_forward = (t >= 1).long() * (t < 1 + self.prompt_len).long()
+ m_backward = (t >= 1).long() * (t < 1 + self.answer_len).long()
+ else:
+ m_forward = (t >= 2 + self.prompt_len).long()
+ m_backward = (t >= 2 + self.answer_len).long()
+
+ m = i_forward.long()[:, None]
+
+ return m * m_forward + (1 - m) * m_backward
def generate_token_sequences(self, nb):
prompts, answers = self.problem.generate_prompts_and_answers(nb)
problem,
nb_train_samples,
nb_test_samples,
+ back_accuracy,
batch_size,
result_dir,
logger,
self.nb_token_values = v + 2
self.problem = problem
+ self.back_accuracy = back_accuracy
self.batch_size = batch_size
self.device = device
self.logger = logger
if result_dir is not None:
self.save_quizzes(
- result_dir, "culture_w_quizzes", self.train_w_quizzes[:72]
+ result_dir,
+ "culture_w_quizzes",
+ self.train_w_quizzes[:72],
+ n_backward=self.train_w_quizzes[:72, 0] == self.token_backward,
)
- def save_quizzes(self, result_dir, filename_prefix, quizzes, prediction=False):
- 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:
+ def save_quizzes(
+ self,
+ result_dir,
+ filename_prefix,
+ quizzes,
+ n_backward=None,
+ mistakes=None,
+ ):
+ quizzes = quizzes.clone()
+ forward = quizzes[quizzes[:, 0] == self.token_forward]
+ ib = quizzes[:, 0] == self.token_backward
+ backward = quizzes[ib]
+ assert forward.size(0) + backward.size(0) == quizzes.size(0)
+ quizzes[ib] = self.reverse_time(quizzes[ib])
+
+ if n_backward is None:
predicted_prompts = None
predicted_answers = None
+ else:
+ predicted_prompts = n_backward.long()
+ predicted_answers = 1 - predicted_prompts
+ if mistakes is not None:
+ # 0/-1/+1 ~ not-to-predict / predicted wrong / predicted correct
+ predicted_prompts *= mistakes
+ predicted_answers *= mistakes
+ else:
+ # 0/2 ~ not-to-predict / to predict
+ predicted_prompts *= 2
+ predicted_answers *= 2
self.problem.save_quizzes(
result_dir,
filename_prefix,
- prompts,
- answers,
+ quizzes[:, 1 : 1 + self.prompt_len],
+ quizzes[:, 2 + self.prompt_len :],
predicted_prompts,
predicted_answers,
)
def produce_results(
self, n_epoch, model, result_dir, deterministic_synthesis, nmax=1000
):
- def compute_accuracy(input):
- input = input[:nmax]
- ar_mask = self.make_ar_mask(input, 2 + self.prompt_len, self.answer_len)
+ def compute_accuracy(input, log_prefix=None):
+ ar_mask = self.make_ar_mask(input)
result = input.clone() * (1 - ar_mask)
seq_logproba = torch.empty(input.size(0), device=self.device)
device=self.device,
)
- nb_total = input.size(0)
- nb_correct = (input == result).long().min(dim=1).values.sum()
+ correct = torch.empty(input.size(0), dtype=torch.int64, device=input.device)
- return nb_total, nb_correct
+ n_forward = input[:, 0] == self.token_forward
+ n_backward = input[:, 0] == self.token_backward
- train_nb_total, train_nb_correct = compute_accuracy(self.train_w_quizzes)
+ correct[n_forward] = (
+ (input[n_forward] == result[n_forward]).long().min(dim=1).values
+ )
- self.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}%"
- )
+ if self.back_accuracy and n_backward.any():
+ # accuracy of B->A*->B*=B instead of B->A*=A
+ back_input = self.reverse_time(result[n_backward])
+ back_input[:, 2 + self.prompt_len :] = input[
+ n_backward, 1 : 1 + self.answer_len
+ ]
+ result[n_backward], correct[n_backward] = compute_accuracy(back_input)
- test_nb_total, test_nb_correct = compute_accuracy(self.test_w_quizzes)
+ if log_prefix is not None:
+ forward_nb_correct = correct[n_forward].sum()
+ forward_nb_total = correct[n_forward].size(0)
+ backward_nb_correct = correct[n_backward].sum()
+ backward_nb_total = correct[n_backward].size(0)
- self.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}%"
- )
+ self.logger(
+ f"forward_accuracy {log_prefix} {n_epoch} {model.id=} {forward_nb_correct} / {forward_nb_total}"
+ )
- main_test_accuracy = test_nb_correct / test_nb_total
- self.logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
+ self.logger(
+ f"backward_accuracy {log_prefix} {n_epoch} {model.id=} {backward_nb_correct} / {backward_nb_total}"
+ )
- ##############################
+ return result, correct
- input = self.test_w_quizzes[:96]
- ar_mask = self.make_ar_mask(input, 2 + self.prompt_len, self.answer_len)
- result = input.clone() * (1 - ar_mask)
- seq_logproba = torch.empty(input.size(0), device=self.device)
+ compute_accuracy(self.train_w_quizzes[:nmax], log_prefix="train")
- masked_inplace_autoregression(
- model=model,
- batch_size=self.batch_size,
- input=result,
- ar_mask=ar_mask,
- seq_logproba=seq_logproba,
- temperature=1.0,
- deterministic_synthesis=deterministic_synthesis,
- progress_bar_desc=None,
- device=self.device,
+ test_result, test_correct = compute_accuracy(
+ self.test_w_quizzes[:nmax], log_prefix="test"
)
+ main_test_accuracy = test_correct.sum() / test_correct.size(0)
+ self.logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
+
+ ##############################
+
self.save_quizzes(
result_dir,
f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
- quizzes=result[:72],
- prediction=True,
+ quizzes=test_result[:72],
+ n_backward=self.test_w_quizzes[:72, 0] == self.token_backward,
+ mistakes=test_correct[:72] * 2 - 1,
)
return main_test_accuracy
else:
self.test_c_quizzes.append(new_c_quizzes)
- def forward_to_backward(self, c_quizzes):
- prompts = c_quizzes[:, 1 : 1 + self.prompt_len]
- answers = c_quizzes[:, 2 + self.prompt_len :]
- return torch.cat(
- [
- c_quizzes.new_full((c_quizzes, 1), self.token_backward),
- answers,
- c_quizzes.new_full((c_quizzes, 1), self.token_backward),
- prompts,
- ],
- dim=1,
- )
-
- def backward_to_forward(self, c_quizzes):
- answers = c_quizzes[:, 1 : 1 + self.answer_len :]
- prompts = c_quizzes[:, 2 + self.answer_len :]
- return torch.cat(
- [
- c_quizzes.new_full((c_quizzes.size(0), 1), self.token_forward),
- prompts,
- c_quizzes.new_full((c_quizzes.size(0), 1), self.token_forward),
- answers,
- ],
- dim=1,
- )
-
def compute_correctness(
self,
c_quizzes,
seq_logproba[...] = 0.0
- ar_mask = self.make_ar_mask(result, 2 + self.prompt_len, self.answer_len)
+ ar_mask = self.make_ar_mask(result)
masked_inplace_autoregression(
model=model,
if bidirectional_validation:
backward_result = backward_c_quizzes.clone()
- ar_mask = self.make_ar_mask(
- backward_result, 2 + self.answer_len, self.prompt_len
- )
+ ar_mask = self.make_ar_mask(backward_result)
masked_inplace_autoregression(
model=model,
nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
)
- ar_mask_first = torch.zeros(c_quizzes.size(), device=self.device)
- ar_mask_first[:, : ar_mask_first.size(1) // 2 + 1] = 1
- ar_mask_second = 1 - ar_mask_first
- ar_mask_first[:, 0] = 0
- ar_mask_second[:, 0] = 0
-
- seq_logproba = torch.zeros(ar_mask_first.size(0), device=self.device)
+ seq_logproba = torch.zeros(nb, device=self.device)
# First, we generate the answer at high temperature
c_quizzes[:, 0] = self.token_backward
+ c_quizzes[:, 1 + self.answer_len] = self.token_backward
masked_inplace_autoregression(
model=model_for_generation,
batch_size=self.batch_size,
input=c_quizzes,
- ar_mask=ar_mask_first,
+ ar_mask=self.make_ar_mask(c_quizzes, first=True),
seq_logproba=seq_logproba,
temperature=temperature,
deterministic_synthesis=False,
model=model_for_generation,
batch_size=self.batch_size,
input=c_quizzes,
- ar_mask=ar_mask_second,
+ ar_mask=self.make_ar_mask(c_quizzes),
seq_logproba=seq_logproba,
temperature=1 / temperature,
deterministic_synthesis=False,
# Then we return the quizz, and re-generate the response, now
# at low temperature
- c_quizzes = self.backward_to_forward(c_quizzes)
+ c_quizzes = self.reverse_time(c_quizzes)
masked_inplace_autoregression(
model=model_for_generation,
batch_size=self.batch_size,
input=c_quizzes,
- ar_mask=ar_mask_second,
+ ar_mask=self.make_ar_mask(c_quizzes),
seq_logproba=seq_logproba,
temperature=1 / temperature,
deterministic_synthesis=False,