X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=quizz_machine.py;h=92b579980ec1a10aec128f8a91927f303417163c;hb=d3d4ce7bb2b799f4bf81a936987e3a8938514af8;hp=de855201c03cc0115ecdddb873e17da0adfd1ed3;hpb=eaed6307836d88abe7c0f4be733a38364ba20e2f;p=culture.git diff --git a/quizz_machine.py b/quizz_machine.py index de85520..92b5799 100755 --- a/quizz_machine.py +++ b/quizz_machine.py @@ -105,19 +105,94 @@ def masked_inplace_autoregression( class QuizzMachine: - def make_ar_mask(self, input): - b = torch.arange(input.size(1), device=input.device) > input.size(1) // 2 - 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) + + if self.prompt_len is None: + self.prompt_len = prompts.size(1) + + if self.answer_len is None: + self.answer_len = answers.size(1) + + assert prompts.size(1) == self.prompt_len and answers.size(1) == self.answer_len + result = [] for prompt, answer in zip(prompts, answers): if torch.rand(1) < 0.5: - a = [torch.tensor([self.token_forward]), prompt, answer] + a = [ + torch.tensor([self.token_forward]), + prompt, + torch.tensor([self.token_forward]), + answer, + ] else: - a = [torch.tensor([self.token_backward]), answer, prompt] + a = [ + torch.tensor([self.token_backward]), + answer, + torch.tensor([self.token_backward]), + prompt, + ] result.append(torch.cat(a, dim=0)[None, :]) @@ -128,6 +203,7 @@ class QuizzMachine: problem, nb_train_samples, nb_test_samples, + back_accuracy, batch_size, result_dir, logger, @@ -141,9 +217,12 @@ class QuizzMachine: 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 + self.prompt_len = None + self.answer_len = None self.train_w_quizzes = self.generate_token_sequences(nb_train_samples).to( device @@ -156,31 +235,47 @@ class QuizzMachine: 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, ) @@ -230,8 +325,7 @@ class QuizzMachine: def produce_results( self, n_epoch, model, result_dir, deterministic_synthesis, nmax=1000 ): - def compute_accuracy(input): - input = input[:nmax] + 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) @@ -248,52 +342,56 @@ class QuizzMachine: device=self.device, ) - nb_total, nb_correct = ( - input.size(0), - (input == result).long().min(dim=1).values.sum(), + correct = torch.empty(input.size(0), dtype=torch.int64, device=input.device) + + n_forward = input[:, 0] == self.token_forward + n_backward = input[:, 0] == self.token_backward + + correct[n_forward] = ( + (input[n_forward] == result[n_forward]).long().min(dim=1).values ) - return nb_total, nb_correct + 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) + + 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"forward_accuracy {log_prefix} {n_epoch} {model.id=} {forward_nb_correct} / {forward_nb_total}" + ) - train_nb_total, train_nb_correct = compute_accuracy(self.train_w_quizzes) + self.logger( + f"backward_accuracy {log_prefix} {n_epoch} {model.id=} {backward_nb_correct} / {backward_nb_total}" + ) - 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}%" - ) + return result, correct - test_nb_total, test_nb_correct = compute_accuracy(self.test_w_quizzes) + compute_accuracy(self.train_w_quizzes[:nmax], log_prefix="train") - 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}%" + test_result, test_correct = compute_accuracy( + self.test_w_quizzes[:nmax], log_prefix="test" ) - main_test_accuracy = test_nb_correct / test_nb_total + main_test_accuracy = test_correct.sum() / test_correct.size(0) self.logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}") ############################## - input = self.test_w_quizzes[:96] - ar_mask = self.make_ar_mask(input) - result = input.clone() * (1 - ar_mask) - seq_logproba = torch.empty(input.size(0), device=self.device) - - 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, - ) - 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 @@ -310,17 +408,6 @@ class QuizzMachine: else: self.test_c_quizzes.append(new_c_quizzes) - def reverse_time(self, c_quizzes): - l = (c_quizzes.size(1) - 1) // 2 - direction = c_quizzes[:, 0:1] - direction = self.token_forward * ( - direction == self.token_backward - ) + self.token_backward * (direction == self.token_forward) - - return torch.cat( - [direction, c_quizzes[:, l + 1 :], c_quizzes[:, 1 : l + 1]], dim=1 - ) - def compute_correctness( self, c_quizzes, @@ -328,17 +415,15 @@ class QuizzMachine: bidirectional_validation=False, deterministic_validation=True, ): - reversed_c_quizzes = self.reverse_time(c_quizzes) + if bidirectional_validation: + backward_c_quizzes = self.forward_to_backward(c_quizzes) - ar_mask = self.make_ar_mask(c_quizzes) seq_logproba = torch.zeros( c_quizzes.size(0), max([m.id for m in models_for_validation]) + 1, device=self.device, ) - # Check how many of models can solve the quizzes in both directions - nb_correct = 0 for model in models_for_validation: @@ -346,6 +431,8 @@ class QuizzMachine: seq_logproba[...] = 0.0 + ar_mask = self.make_ar_mask(result) + masked_inplace_autoregression( model=model, batch_size=self.batch_size, @@ -361,25 +448,27 @@ class QuizzMachine: correct = (c_quizzes == result).long().min(dim=-1).values if bidirectional_validation: - reversed_result = reversed_c_quizzes.clone() + backward_result = backward_c_quizzes.clone() + + ar_mask = self.make_ar_mask(backward_result) masked_inplace_autoregression( model=model, batch_size=self.batch_size, - input=reversed_result, + input=backward_result, ar_mask=ar_mask, seq_logproba=seq_logproba[:, model.id], temperature=1.0, deterministic_synthesis=deterministic_validation, - # progress_bar_desc="solving reversed c_quizzes", + # progress_bar_desc="solving backward c_quizzes", device=self.device, ) - reversed_correct = ( - (reversed_c_quizzes == reversed_result).long().min(dim=-1).values + backward_correct = ( + (backward_c_quizzes == backward_result).long().min(dim=-1).values ) - correct *= reversed_correct + correct *= backward_correct # endif @@ -394,23 +483,18 @@ class QuizzMachine: 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, @@ -423,7 +507,7 @@ class QuizzMachine: 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, @@ -439,7 +523,7 @@ class QuizzMachine: 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,