3 # Any copyright is dedicated to the Public Domain.
4 # https://creativecommons.org/publicdomain/zero/1.0/
6 # Written by Francois Fleuret <francois@fleuret.org>
8 import math, os, tqdm, warnings
10 import torch, torchvision
13 from torch.nn import functional as F
16 from mygpt import BracketedSequence
18 ######################################################################
20 # ar_mask is a tensor with 0s and 1s, of same shape as input, with
21 # 1s where tokens should be generated. The others are kept
25 def one_batch_masked_inplace_autoregression(
31 deterministic_synthesis=False,
33 to_generate = (ar_mask.sum(0) > 0).nonzero()
35 if to_generate.min() > 0:
37 BracketedSequence(input, 0, to_generate.min())
38 ) # Needed to initialize the model's cache
39 for s in range(to_generate.min(), to_generate.max() + 1):
40 output = model(BracketedSequence(input, s, 1)).x
44 logits = (logits / temperature).log_softmax(dim=-1)
46 if deterministic_synthesis:
47 t_next = logits.argmax(-1)
49 dist = torch.distributions.categorical.Categorical(logits=logits)
50 t_next = dist.sample()
52 all_n = torch.arange(t_next.size(0))
53 seq_logproba += logits[all_n, t_next].sum(dim=-1)
55 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
58 def masked_inplace_autoregression(
65 deterministic_synthesis,
66 forbidden_tokens=None,
68 progress_bar_desc=None,
69 device=torch.device("cpu"),
71 assert input.size() == ar_mask.size()
74 input.split(batch_size),
75 ar_mask.split(batch_size),
76 seq_logproba.split(batch_size),
79 if progress_bar_desc is not None:
83 desc=progress_bar_desc,
84 total=(input.size(0) + batch_size - 1) // batch_size,
87 with torch.autograd.no_grad():
91 for input, ar_mask, seq_logproba in batches:
92 one_batch_masked_inplace_autoregression(
96 seq_logproba=seq_logproba,
97 temperature=temperature,
98 deterministic_synthesis=deterministic_synthesis,
104 ######################################################################
108 def indices_forward_and_backward(self, quizzes):
109 i_forward = quizzes[:, 0] == self.token_forward
110 j_forward = quizzes[:, 1 + self.prompt_len] == self.token_forward
111 i_backward = quizzes[:, 0] == self.token_backward
112 j_backward = quizzes[:, 1 + self.answer_len] == self.token_backward
113 assert torch.logical_or(
114 torch.logical_and(i_forward, j_forward),
115 torch.logical_and(i_backward, j_backward),
117 return i_forward, i_backward
119 def reverse_time(self, quizzes):
120 i_forward, i_backward = self.indices_forward_and_backward(quizzes)
122 forward_to_backward = torch.cat(
125 quizzes[:, 2 + self.prompt_len : 2 + self.prompt_len + self.answer_len],
126 quizzes[:, 1 + self.prompt_len : 1 + self.prompt_len + 1],
127 quizzes[:, 1 : 1 + self.prompt_len],
132 forward_to_backward[:, 0] = self.token_backward
133 forward_to_backward[:, 1 + self.answer_len] = self.token_backward
135 backward_to_forward = torch.cat(
138 quizzes[:, 2 + self.answer_len :],
139 quizzes[:, 1 + self.answer_len : 2 + self.answer_len],
140 quizzes[:, 1 : 1 + self.answer_len],
145 backward_to_forward[:, 0] = self.token_forward
146 backward_to_forward[:, 1 + self.prompt_len] = self.token_forward
148 m = i_forward.long()[:, None]
150 return m * forward_to_backward + (1 - m) * backward_to_forward
152 def make_ar_mask(self, quizzes, first=False):
153 i_forward, i_backward = self.indices_forward_and_backward(quizzes)
155 t = torch.arange(quizzes.size(1), device=quizzes.device)
158 m_forward = (t >= 1).long() * (t < 1 + self.prompt_len).long()
159 m_backward = (t >= 1).long() * (t < 1 + self.answer_len).long()
161 m_forward = (t >= 2 + self.prompt_len).long()
162 m_backward = (t >= 2 + self.answer_len).long()
164 m = i_forward.long()[:, None]
166 return m * m_forward + (1 - m) * m_backward
168 def generate_token_sequences(self, nb):
169 prompts, answers = self.problem.generate_prompts_and_answers(nb)
171 if self.prompt_len is None:
172 self.prompt_len = prompts.size(1)
174 if self.answer_len is None:
175 self.answer_len = answers.size(1)
177 assert prompts.size(1) == self.prompt_len and answers.size(1) == self.answer_len
181 for prompt, answer in zip(prompts, answers):
182 if torch.rand(1) < 0.5:
184 torch.tensor([self.token_forward]),
186 torch.tensor([self.token_forward]),
191 torch.tensor([self.token_backward]),
193 torch.tensor([self.token_backward]),
197 result.append(torch.cat(a, dim=0)[None, :])
199 return torch.cat(result, dim=0)
210 device=torch.device("cpu"),
214 v = problem.nb_token_values()
215 self.token_forward = v
216 self.token_backward = v + 1
217 self.nb_token_values = v + 2
219 self.problem = problem
220 self.back_accuracy = back_accuracy
221 self.batch_size = batch_size
224 self.prompt_len = None
225 self.answer_len = None
227 self.train_w_quizzes = self.generate_token_sequences(nb_train_samples).to(
231 self.test_w_quizzes = self.generate_token_sequences(nb_test_samples).to(device)
233 self.train_c_quizzes = []
234 self.test_c_quizzes = []
236 if result_dir is not None:
240 self.train_w_quizzes[:72],
241 show_to_be_predicted=True,
249 show_to_be_predicted=False,
252 quizzes = quizzes.clone()
253 forward = quizzes[quizzes[:, 0] == self.token_forward]
254 ib = quizzes[:, 0] == self.token_backward
255 backward = quizzes[ib]
256 assert forward.size(0) + backward.size(0) == quizzes.size(0)
257 quizzes[ib] = self.reverse_time(quizzes[ib])
259 if show_to_be_predicted:
260 predicted_prompts = ib.long()
261 predicted_answers = 1 - predicted_prompts
262 if mistakes is not None:
263 # 0/-1/+1 ~ not-to-predict / predicted wrong / predicted correct
264 predicted_prompts *= mistakes
265 predicted_answers *= mistakes
267 # 0/2 ~ not-to-predict / to predict
268 predicted_prompts *= 2
269 predicted_answers *= 2
271 predicted_prompts = None
272 predicted_answers = None
274 self.problem.save_quizzes(
277 quizzes[:, 1 : 1 + self.prompt_len],
278 quizzes[:, 2 + self.prompt_len :],
283 def batches(self, split="train", desc=None):
284 assert split in {"train", "test"}
286 w_quizzes = self.train_w_quizzes
287 c_quizzes = self.train_c_quizzes
289 w_quizzes = self.test_w_quizzes
290 c_quizzes = self.test_c_quizzes
292 if len(c_quizzes) > 0:
293 c_quizzes = torch.cat(c_quizzes, dim=0)
294 if c_quizzes.size(0) > w_quizzes.size(0) // 2:
295 i = torch.randperm(c_quizzes.size(0))[: w_quizzes.size(0) // 2]
296 c_quizzes = c_quizzes[i]
298 i = torch.randperm(w_quizzes.size(0))[
299 : w_quizzes.size(0) - c_quizzes.size(0)
301 w_quizzes = w_quizzes[i]
303 self.nb_batch_w_quizzes = w_quizzes.size(0)
304 self.nb_batch_c_quizzes = c_quizzes.size(0)
306 input = torch.cat([w_quizzes, c_quizzes], dim=0)
309 self.nb_batch_w_quizzes = w_quizzes.size(0)
310 self.nb_batch_c_quizzes = 0
313 input = input[torch.randperm(input.size(0))]
316 desc = f"epoch-{split}"
317 for batch in tqdm.tqdm(
318 input.split(self.batch_size), dynamic_ncols=True, desc=desc
322 def vocabulary_size(self):
323 return self.nb_token_values
326 self, n_epoch, model, result_dir, deterministic_synthesis, nmax=1000
328 def compute_accuracy(input, log_prefix=None):
329 ar_mask = self.make_ar_mask(input)
330 result = input.clone() * (1 - ar_mask)
331 seq_logproba = torch.empty(input.size(0), device=self.device)
333 masked_inplace_autoregression(
335 batch_size=self.batch_size,
338 seq_logproba=seq_logproba,
340 deterministic_synthesis=deterministic_synthesis,
341 progress_bar_desc=None,
345 correct = torch.empty(input.size(0), dtype=torch.int64, device=input.device)
347 n_forward = input[:, 0] == self.token_forward
348 n_backward = input[:, 0] == self.token_backward
350 correct[n_forward] = (
351 (input[n_forward] == result[n_forward]).long().min(dim=1).values
354 if self.back_accuracy and n_backward.any():
355 # accuracy of B->A*->B*=B instead of B->A*=A
356 back_input = self.reverse_time(result[n_backward])
357 back_input[:, 2 + self.prompt_len :] = input[
358 n_backward, 1 : 1 + self.answer_len
360 result[n_backward], correct[n_backward] = compute_accuracy(back_input)
362 if log_prefix is not None:
363 forward_nb_correct = correct[n_forward].sum()
364 forward_nb_total = correct[n_forward].size(0)
365 backward_nb_correct = correct[n_backward].sum()
366 backward_nb_total = correct[n_backward].size(0)
369 f"forward_accuracy {log_prefix} {n_epoch} {model.id=} {forward_nb_correct} / {forward_nb_total}"
373 f"backward_accuracy {log_prefix} {n_epoch} {model.id=} {backward_nb_correct} / {backward_nb_total}"
376 return result, correct
378 compute_accuracy(self.train_w_quizzes[:nmax], log_prefix="train")
380 test_result, test_correct = compute_accuracy(
381 self.test_w_quizzes[:nmax], log_prefix="test"
384 main_test_accuracy = test_correct.sum() / test_correct.size(0)
385 self.logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
387 ##############################
391 f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
392 quizzes=test_result[:72],
393 show_to_be_predicted=True,
394 mistakes=test_correct[:72] * 2 - 1,
397 return main_test_accuracy
399 def renew_w_quizzes(self, nb, for_train=True):
400 input = self.train_w_quizzes if for_train else self.test_w_quizzes
401 nb = min(nb, input.size(0))
402 input[:-nb] = input[nb:].clone()
403 input[-nb:] = self.generate_token_sequences(nb).to(self.device)
405 def store_c_quizzes(self, new_c_quizzes, for_train=True):
407 self.train_c_quizzes.append(new_c_quizzes)
409 self.test_c_quizzes.append(new_c_quizzes)
411 def compute_correctness(
414 models_for_validation,
415 bidirectional_validation=False,
416 deterministic_validation=True,
418 if bidirectional_validation:
419 backward_c_quizzes = self.forward_to_backward(c_quizzes)
421 seq_logproba = torch.zeros(
423 max([m.id for m in models_for_validation]) + 1,
429 for model in models_for_validation:
430 result = c_quizzes.clone()
432 seq_logproba[...] = 0.0
434 ar_mask = self.make_ar_mask(result)
436 masked_inplace_autoregression(
438 batch_size=self.batch_size,
441 seq_logproba=seq_logproba[:, model.id],
443 deterministic_synthesis=deterministic_validation,
444 # progress_bar_desc="solving c_quizzes",
448 correct = (c_quizzes == result).long().min(dim=-1).values
450 if bidirectional_validation:
451 backward_result = backward_c_quizzes.clone()
453 ar_mask = self.make_ar_mask(backward_result)
455 masked_inplace_autoregression(
457 batch_size=self.batch_size,
458 input=backward_result,
460 seq_logproba=seq_logproba[:, model.id],
462 deterministic_synthesis=deterministic_validation,
463 # progress_bar_desc="solving backward c_quizzes",
468 (backward_c_quizzes == backward_result).long().min(dim=-1).values
471 correct *= backward_correct
475 nb_correct += correct
477 return nb_correct, seq_logproba
479 ###############################################################
481 def generate_quizzes(self, nb, model_for_generation, temperature=1.0):
482 c_quizzes = torch.empty(
483 nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
486 seq_logproba = torch.zeros(nb, device=self.device)
488 # First, we generate the answer at high temperature
490 c_quizzes[:, 0] = self.token_backward
491 c_quizzes[:, 1 + self.answer_len] = self.token_backward
493 masked_inplace_autoregression(
494 model=model_for_generation,
495 batch_size=self.batch_size,
497 ar_mask=self.make_ar_mask(c_quizzes, first=True),
498 seq_logproba=seq_logproba,
499 temperature=temperature,
500 deterministic_synthesis=False,
504 # Then, we generate the prompt at low temperature
506 masked_inplace_autoregression(
507 model=model_for_generation,
508 batch_size=self.batch_size,
510 ar_mask=self.make_ar_mask(c_quizzes),
511 seq_logproba=seq_logproba,
512 temperature=1 / temperature,
513 deterministic_synthesis=False,
517 # Then we return the quizz, and re-generate the response, now
520 c_quizzes = self.reverse_time(c_quizzes)
522 masked_inplace_autoregression(
523 model=model_for_generation,
524 batch_size=self.batch_size,
526 ar_mask=self.make_ar_mask(c_quizzes),
527 seq_logproba=seq_logproba,
528 temperature=1 / temperature,
529 deterministic_synthesis=False,