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
20 ######################################################################
22 # ar_mask is a tensor with 0s and 1s, of same shape as input, with
23 # 1s where tokens should be generated. The others are kept
27 def one_batch_masked_inplace_autoregression(
33 deterministic_synthesis,
35 to_generate = (ar_mask.sum(0) > 0).nonzero()
37 if to_generate.min() > 0:
39 BracketedSequence(input, 0, to_generate.min())
40 ) # Needed to initialize the model's cache
41 for s in range(to_generate.min(), to_generate.max() + 1):
42 output = model(BracketedSequence(input, s, 1)).x
46 logits = (logits / temperature).log_softmax(dim=-1)
48 if deterministic_synthesis:
49 t_next = logits.argmax(-1)
51 dist = torch.distributions.categorical.Categorical(logits=logits)
52 t_next = dist.sample()
54 all_n = torch.arange(t_next.size(0))
56 seq_logproba += logits[all_n, t_next]
58 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
61 def masked_inplace_autoregression(
68 deterministic_synthesis,
69 forbidden_tokens=None,
71 progress_bar_desc=None,
72 device=torch.device("cpu"),
74 assert input.size() == ar_mask.size()
77 input.split(batch_size),
78 ar_mask.split(batch_size),
79 seq_logproba.split(batch_size),
82 if progress_bar_desc is not None:
86 desc=progress_bar_desc,
87 total=(input.size(0) + batch_size - 1) // batch_size,
90 with torch.autograd.no_grad():
94 for input, ar_mask, seq_logproba in batches:
95 one_batch_masked_inplace_autoregression(
99 seq_logproba=seq_logproba,
100 temperature=temperature,
101 deterministic_synthesis=deterministic_synthesis,
107 ######################################################################
111 def indices_forward_and_backward(self, quizzes):
112 i_forward = quizzes[:, 0] == self.token_forward
113 j_forward = quizzes[:, 1 + self.prompt_len] == self.token_forward
114 i_backward = quizzes[:, 0] == self.token_backward
115 j_backward = quizzes[:, 1 + self.answer_len] == self.token_backward
116 assert torch.logical_or(
117 torch.logical_and(i_forward, j_forward),
118 torch.logical_and(i_backward, j_backward),
120 return i_forward, i_backward
122 def non_trivial(self, quizzes):
123 quizzes = quizzes.clone()
124 n_forward = quizzes[quizzes[:, 0] == self.token_forward]
125 n_backward = quizzes[:, 0] == self.token_backward
126 backward = quizzes[n_backward]
127 quizzes[n_backward] = self.reverse_time(quizzes[n_backward])
128 return torch.logical_not(
129 self.problem.trivial_prompts_and_answers(
130 quizzes[:, 1 : 1 + self.prompt_len],
131 quizzes[:, 2 + self.prompt_len :],
135 def reverse_time(self, quizzes):
136 i_forward, i_backward = self.indices_forward_and_backward(quizzes)
138 forward_to_backward = torch.cat(
141 quizzes[:, 2 + self.prompt_len : 2 + self.prompt_len + self.answer_len],
142 quizzes[:, 1 + self.prompt_len : 1 + self.prompt_len + 1],
143 quizzes[:, 1 : 1 + self.prompt_len],
148 forward_to_backward[:, 0] = self.token_backward
149 forward_to_backward[:, 1 + self.answer_len] = self.token_backward
151 backward_to_forward = torch.cat(
154 quizzes[:, 2 + self.answer_len :],
155 quizzes[:, 1 + self.answer_len : 2 + self.answer_len],
156 quizzes[:, 1 : 1 + self.answer_len],
161 backward_to_forward[:, 0] = self.token_forward
162 backward_to_forward[:, 1 + self.prompt_len] = self.token_forward
164 m = i_forward.long()[:, None]
166 return m * forward_to_backward + (1 - m) * backward_to_forward
168 def reverse_random_half_in_place(self, quizzes):
169 i = torch.rand(quizzes.size(0)) < 0.5
171 quizzes[i] = self.reverse_time(quizzes[i])
173 def make_ar_mask(self, quizzes, first=False):
174 i_forward, i_backward = self.indices_forward_and_backward(quizzes)
176 t = torch.arange(quizzes.size(1), device=quizzes.device)
179 m_forward = (t >= 1).long() * (t < 1 + self.prompt_len).long()
180 m_backward = (t >= 1).long() * (t < 1 + self.answer_len).long()
182 m_forward = (t >= 2 + self.prompt_len).long()
183 m_backward = (t >= 2 + self.answer_len).long()
185 m = i_forward.long()[:, None]
187 return m * m_forward + (1 - m) * m_backward
189 def generate_token_sequences(self, nb):
190 prompts, answers = self.problem.generate_prompts_and_answers(nb)
192 if self.prompt_len is None:
193 self.prompt_len = prompts.size(1)
195 if self.answer_len is None:
196 self.answer_len = answers.size(1)
198 assert prompts.size(1) == self.prompt_len and answers.size(1) == self.answer_len
202 for prompt, answer in zip(prompts, answers):
204 torch.tensor([self.token_forward]),
206 torch.tensor([self.token_forward]),
210 result.append(torch.cat(a, dim=0)[None, :])
212 return torch.cat(result, dim=0)
223 device=torch.device("cpu"),
227 v = problem.nb_token_values()
228 self.token_forward = v
229 self.token_backward = v + 1
230 self.nb_token_values = v + 2
232 self.problem = problem
233 self.back_accuracy = back_accuracy
234 self.batch_size = batch_size
237 self.prompt_len = None
238 self.answer_len = None
240 self.LOCK_C_QUIZZES = threading.Lock()
241 self.train_c_quizzes = []
242 self.test_c_quizzes = []
251 quizzes = quizzes.clone().to("cpu")
252 n_forward = quizzes[quizzes[:, 0] == self.token_forward]
253 n_backward = quizzes[:, 0] == self.token_backward
254 backward = quizzes[n_backward]
255 assert n_forward.size(0) + backward.size(0) == quizzes.size(0)
256 quizzes[n_backward] = self.reverse_time(quizzes[n_backward])
258 predicted_prompts = n_backward.long()
259 predicted_answers = 1 - predicted_prompts
260 if mistakes is not None:
261 # 0/-1/+1 ~ not-to-predict / predicted wrong / predicted correct
262 predicted_prompts *= mistakes.to("cpu")
263 predicted_answers *= mistakes.to("cpu")
265 # 0/2 ~ not-to-predict / to predict
266 predicted_prompts *= 2
267 predicted_answers *= 2
269 self.problem.save_quizzes(
272 quizzes[:, 1 : 1 + self.prompt_len],
273 quizzes[:, 2 + self.prompt_len :],
278 def vocabulary_size(self):
279 return self.nb_token_values
281 ######################################################################
283 def batches(self, model, split="train", desc=None):
284 assert split in {"train", "test"}
286 with self.LOCK_C_QUIZZES:
288 w_quizzes = model.train_w_quizzes
289 c_quizzes = self.train_c_quizzes
291 w_quizzes = model.test_w_quizzes
292 c_quizzes = self.test_c_quizzes
294 if len(c_quizzes) > 0:
295 c_quizzes = torch.cat(c_quizzes, dim=0)
296 if c_quizzes.size(0) > w_quizzes.size(0) // 2:
297 i = torch.randperm(c_quizzes.size(0))[: w_quizzes.size(0) // 2]
298 c_quizzes = c_quizzes[i]
300 i = torch.randperm(w_quizzes.size(0))[
301 : w_quizzes.size(0) - c_quizzes.size(0)
303 w_quizzes = w_quizzes[i]
305 self.nb_batch_w_quizzes = w_quizzes.size(0)
306 self.nb_batch_c_quizzes = c_quizzes.size(0)
308 input = torch.cat([w_quizzes, c_quizzes], dim=0)
311 self.nb_batch_w_quizzes = w_quizzes.size(0)
312 self.nb_batch_c_quizzes = 0
315 input = input[torch.randperm(input.size(0))]
318 desc = f"epoch-{split}"
319 for batch in tqdm.tqdm(
320 input.split(self.batch_size), dynamic_ncols=True, desc=desc
324 ######################################################################
327 self, n_epoch, model, result_dir, deterministic_synthesis, nmax=1000
329 def compute_accuracy(input, log_prefix=None):
330 input = input.to(self.device)
331 ar_mask = self.make_ar_mask(input)
332 result = input.clone() * (1 - ar_mask)
333 seq_logproba = torch.empty(input.size(0), device=self.device)
335 masked_inplace_autoregression(
337 batch_size=self.batch_size,
340 seq_logproba=seq_logproba,
342 deterministic_synthesis=deterministic_synthesis,
343 progress_bar_desc=None,
347 correct = torch.empty(input.size(0), dtype=torch.int64, device=input.device)
349 n_forward = input[:, 0] == self.token_forward
350 n_backward = input[:, 0] == self.token_backward
352 correct[n_forward] = (
353 (input[n_forward] == result[n_forward]).long().min(dim=1).values
356 if self.back_accuracy and n_backward.any():
357 # accuracy of B->A*->B*=B instead of B->A*=A
358 back_input = self.reverse_time(result[n_backward])
359 back_input[:, 2 + self.prompt_len :] = input[
360 n_backward, 1 : 1 + self.answer_len
362 _, correct[n_backward] = compute_accuracy(back_input)
364 if log_prefix is not None:
365 forward_nb_correct = correct[n_forward].sum()
366 forward_nb_total = correct[n_forward].size(0)
367 backward_nb_correct = correct[n_backward].sum()
368 backward_nb_total = correct[n_backward].size(0)
371 f"{log_prefix}_forward_accuracy {n_epoch} model {model.id} nb_correct {forward_nb_correct} / {forward_nb_total} ({forward_nb_correct*100/forward_nb_total} %)"
375 f"{log_prefix}_backward_accuracy {n_epoch} model {model.id} nb_correct {backward_nb_correct} / {backward_nb_total} ({backward_nb_correct*100/backward_nb_total} %)"
378 return result, correct
380 compute_accuracy(model.train_w_quizzes[:nmax], log_prefix="train")
382 test_result, test_correct = compute_accuracy(
383 model.test_w_quizzes[:nmax], log_prefix="test"
386 main_test_accuracy = test_correct.sum() / test_correct.size(0)
387 self.logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
389 ##############################
393 f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
394 quizzes=test_result[:72],
395 mistakes=test_correct[:72] * 2 - 1,
398 return main_test_accuracy
400 ######################################################################
402 def renew_w_quizzes(self, model, nb, for_train=True):
403 input = model.train_w_quizzes if for_train else model.test_w_quizzes
404 nb = min(nb, input.size(0))
405 input[:-nb] = input[nb:].clone()
406 fresh_w_quizzes = self.generate_token_sequences(nb)
407 self.reverse_random_half_in_place(fresh_w_quizzes)
408 input[-nb:] = fresh_w_quizzes.to("cpu")
410 ######################################################################
412 def store_c_quizzes(self, new_c_quizzes, for_train=True):
413 with self.LOCK_C_QUIZZES:
415 self.train_c_quizzes.append(new_c_quizzes.to("cpu"))
417 self.test_c_quizzes.append(new_c_quizzes.to("cpu"))
419 ######################################################################
421 def logproba_of_solutions(self, models, c_quizzes):
422 logproba = c_quizzes.new_zeros(
423 c_quizzes.size(0), len(models), device=self.device
427 with torch.autograd.no_grad():
432 c_quizzes.split(self.batch_size), logproba.split(self.batch_size)
434 input = input.to(self.device)
435 ar_mask = self.make_ar_mask(input)
436 output = model(mygpt.BracketedSequence(input)).x
438 F.cross_entropy(output.transpose(1, 2), input, reduction="none")
441 l[:, model.id] = -ce.sum(dim=-1)
445 return logproba.to("cpu")
447 ###############################################################
449 def compute_correctness(
452 models_for_validation,
453 bidirectional_validation=False,
454 deterministic_validation=True,
456 if bidirectional_validation:
457 backward_c_quizzes = self.forward_to_backward(c_quizzes)
459 seq_logproba = torch.zeros(
461 max([m.id for m in models_for_validation]) + 1,
467 seq_logproba[...] = 0.0
469 for model in models_for_validation:
470 result = c_quizzes.clone()
472 ar_mask = self.make_ar_mask(result)
474 masked_inplace_autoregression(
476 batch_size=self.batch_size,
479 seq_logproba=seq_logproba[:, model.id],
481 deterministic_synthesis=deterministic_validation,
482 # progress_bar_desc="solving c_quizzes",
486 correct = (c_quizzes == result).long().min(dim=-1).values
488 if bidirectional_validation:
489 backward_result = backward_c_quizzes.clone()
491 ar_mask = self.make_ar_mask(backward_result)
493 masked_inplace_autoregression(
495 batch_size=self.batch_size,
496 input=backward_result,
498 seq_logproba=seq_logproba[:, model.id],
500 deterministic_synthesis=deterministic_validation,
501 # progress_bar_desc="solving backward c_quizzes",
506 (backward_c_quizzes == backward_result).long().min(dim=-1).values
509 correct *= backward_correct
513 nb_correct += correct
515 return nb_correct, seq_logproba
517 ###############################################################
519 def generate_quizzes(self, nb, model_for_generation, temperature=1.0):
520 c_quizzes = torch.empty(
522 self.prompt_len + self.answer_len + 2,
527 seq_logproba = torch.zeros(nb, device=self.device)
529 # First, we generate the answer at high temperature
531 c_quizzes[:, 0] = self.token_backward
532 c_quizzes[:, 1 + self.answer_len] = self.token_backward
534 masked_inplace_autoregression(
535 model=model_for_generation,
536 batch_size=self.batch_size,
538 ar_mask=self.make_ar_mask(c_quizzes, first=True),
539 seq_logproba=seq_logproba,
540 temperature=temperature,
541 deterministic_synthesis=False,
545 # Then, we generate the prompt at low temperature
547 masked_inplace_autoregression(
548 model=model_for_generation,
549 batch_size=self.batch_size,
551 ar_mask=self.make_ar_mask(c_quizzes),
552 seq_logproba=seq_logproba,
553 temperature=1 / temperature,
554 deterministic_synthesis=False,
558 # Then we return the quizz, and re-generate the response, now
561 c_quizzes = self.reverse_time(c_quizzes)
563 masked_inplace_autoregression(
564 model=model_for_generation,
565 batch_size=self.batch_size,
567 ar_mask=self.make_ar_mask(c_quizzes),
568 seq_logproba=seq_logproba,
569 temperature=1 / temperature,
570 deterministic_synthesis=False,
574 return c_quizzes.to("cpu")