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,
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))
54 seq_logproba += logits[all_n, t_next]
56 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
59 def masked_inplace_autoregression(
66 deterministic_synthesis,
67 forbidden_tokens=None,
69 progress_bar_desc=None,
70 device=torch.device("cpu"),
72 assert input.size() == ar_mask.size()
75 input.split(batch_size),
76 ar_mask.split(batch_size),
77 seq_logproba.split(batch_size),
80 if progress_bar_desc is not None:
84 desc=progress_bar_desc,
85 total=(input.size(0) + batch_size - 1) // batch_size,
88 with torch.autograd.no_grad():
92 for input, ar_mask, seq_logproba in batches:
93 one_batch_masked_inplace_autoregression(
97 seq_logproba=seq_logproba,
98 temperature=temperature,
99 deterministic_synthesis=deterministic_synthesis,
105 ######################################################################
109 def indices_forward_and_backward(self, quizzes):
110 i_forward = quizzes[:, 0] == self.token_forward
111 j_forward = quizzes[:, 1 + self.prompt_len] == self.token_forward
112 i_backward = quizzes[:, 0] == self.token_backward
113 j_backward = quizzes[:, 1 + self.answer_len] == self.token_backward
114 assert torch.logical_or(
115 torch.logical_and(i_forward, j_forward),
116 torch.logical_and(i_backward, j_backward),
118 return i_forward, i_backward
120 def non_trivial(self, quizzes):
121 quizzes = quizzes.clone()
122 n_forward = quizzes[quizzes[:, 0] == self.token_forward]
123 n_backward = quizzes[:, 0] == self.token_backward
124 backward = quizzes[n_backward]
125 quizzes[n_backward] = self.reverse_time(quizzes[n_backward])
126 return torch.logical_not(
127 self.problem.trivial_prompts_and_answers(
128 quizzes[:, 1 : 1 + self.prompt_len],
129 quizzes[:, 2 + self.prompt_len :],
133 def reverse_time(self, quizzes):
134 i_forward, i_backward = self.indices_forward_and_backward(quizzes)
136 forward_to_backward = torch.cat(
139 quizzes[:, 2 + self.prompt_len : 2 + self.prompt_len + self.answer_len],
140 quizzes[:, 1 + self.prompt_len : 1 + self.prompt_len + 1],
141 quizzes[:, 1 : 1 + self.prompt_len],
146 forward_to_backward[:, 0] = self.token_backward
147 forward_to_backward[:, 1 + self.answer_len] = self.token_backward
149 backward_to_forward = torch.cat(
152 quizzes[:, 2 + self.answer_len :],
153 quizzes[:, 1 + self.answer_len : 2 + self.answer_len],
154 quizzes[:, 1 : 1 + self.answer_len],
159 backward_to_forward[:, 0] = self.token_forward
160 backward_to_forward[:, 1 + self.prompt_len] = self.token_forward
162 m = i_forward.long()[:, None]
164 return m * forward_to_backward + (1 - m) * backward_to_forward
166 def reverse_random_half_in_place(self, quizzes):
167 i = torch.rand(quizzes.size(0)) < 0.5
169 quizzes[i] = self.reverse_time(quizzes[i])
171 def make_ar_mask(self, quizzes, first=False):
172 i_forward, i_backward = self.indices_forward_and_backward(quizzes)
174 t = torch.arange(quizzes.size(1), device=quizzes.device)
177 m_forward = (t >= 1).long() * (t < 1 + self.prompt_len).long()
178 m_backward = (t >= 1).long() * (t < 1 + self.answer_len).long()
180 m_forward = (t >= 2 + self.prompt_len).long()
181 m_backward = (t >= 2 + self.answer_len).long()
183 m = i_forward.long()[:, None]
185 return m * m_forward + (1 - m) * m_backward
187 def generate_token_sequences(self, nb):
188 prompts, answers = self.problem.generate_prompts_and_answers(nb)
190 if self.prompt_len is None:
191 self.prompt_len = prompts.size(1)
193 if self.answer_len is None:
194 self.answer_len = answers.size(1)
196 assert prompts.size(1) == self.prompt_len and answers.size(1) == self.answer_len
200 for prompt, answer in zip(prompts, answers):
202 torch.tensor([self.token_forward]),
204 torch.tensor([self.token_forward]),
208 result.append(torch.cat(a, dim=0)[None, :])
210 return torch.cat(result, dim=0)
221 device=torch.device("cpu"),
225 v = problem.nb_token_values()
226 self.token_forward = v
227 self.token_backward = v + 1
228 self.nb_token_values = v + 2
230 self.problem = problem
231 self.back_accuracy = back_accuracy
232 self.batch_size = batch_size
235 self.prompt_len = None
236 self.answer_len = None
238 # self.train_w_quizzes = self.generate_token_sequences(nb_train_samples)
239 # self.reverse_random_half_in_place(self.train_w_quizzes)
241 # self.test_w_quizzes = self.generate_token_sequences(nb_test_samples).to(device)
242 # self.reverse_random_half_in_place(self.test_w_quizzes)
244 self.train_c_quizzes = []
245 self.test_c_quizzes = []
247 # if result_dir is not None:
250 # "culture_w_quizzes",
251 # self.train_w_quizzes[:72],
261 quizzes = quizzes.clone().to("cpu")
262 n_forward = quizzes[quizzes[:, 0] == self.token_forward]
263 n_backward = quizzes[:, 0] == self.token_backward
264 backward = quizzes[n_backward]
265 assert n_forward.size(0) + backward.size(0) == quizzes.size(0)
266 quizzes[n_backward] = self.reverse_time(quizzes[n_backward])
268 predicted_prompts = n_backward.long()
269 predicted_answers = 1 - predicted_prompts
270 if mistakes is not None:
271 # 0/-1/+1 ~ not-to-predict / predicted wrong / predicted correct
272 predicted_prompts *= mistakes.to("cpu")
273 predicted_answers *= mistakes.to("cpu")
275 # 0/2 ~ not-to-predict / to predict
276 predicted_prompts *= 2
277 predicted_answers *= 2
279 self.problem.save_quizzes(
282 quizzes[:, 1 : 1 + self.prompt_len],
283 quizzes[:, 2 + self.prompt_len :],
288 def batches(self, model, split="train", desc=None):
289 assert split in {"train", "test"}
291 w_quizzes = model.train_w_quizzes
292 c_quizzes = self.train_c_quizzes
294 w_quizzes = model.test_w_quizzes
295 c_quizzes = self.test_c_quizzes
297 if len(c_quizzes) > 0:
298 c_quizzes = torch.cat(c_quizzes, dim=0)
299 if c_quizzes.size(0) > w_quizzes.size(0) // 2:
300 i = torch.randperm(c_quizzes.size(0))[: w_quizzes.size(0) // 2]
301 c_quizzes = c_quizzes[i]
303 i = torch.randperm(w_quizzes.size(0))[
304 : w_quizzes.size(0) - c_quizzes.size(0)
306 w_quizzes = w_quizzes[i]
308 self.nb_batch_w_quizzes = w_quizzes.size(0)
309 self.nb_batch_c_quizzes = c_quizzes.size(0)
311 input = torch.cat([w_quizzes, c_quizzes], dim=0)
314 self.nb_batch_w_quizzes = w_quizzes.size(0)
315 self.nb_batch_c_quizzes = 0
318 input = input[torch.randperm(input.size(0))]
321 desc = f"epoch-{split}"
322 for batch in tqdm.tqdm(
323 input.split(self.batch_size), dynamic_ncols=True, desc=desc
327 def vocabulary_size(self):
328 return self.nb_token_values
331 self, n_epoch, model, result_dir, deterministic_synthesis, nmax=1000
333 def compute_accuracy(input, log_prefix=None):
334 ar_mask = self.make_ar_mask(input)
335 result = input.clone() * (1 - ar_mask)
336 seq_logproba = torch.empty(input.size(0), device=self.device)
338 masked_inplace_autoregression(
340 batch_size=self.batch_size,
343 seq_logproba=seq_logproba,
345 deterministic_synthesis=deterministic_synthesis,
346 progress_bar_desc=None,
350 correct = torch.empty(input.size(0), dtype=torch.int64, device=input.device)
352 n_forward = input[:, 0] == self.token_forward
353 n_backward = input[:, 0] == self.token_backward
355 correct[n_forward] = (
356 (input[n_forward] == result[n_forward]).long().min(dim=1).values
359 if self.back_accuracy and n_backward.any():
360 # accuracy of B->A*->B*=B instead of B->A*=A
361 back_input = self.reverse_time(result[n_backward])
362 back_input[:, 2 + self.prompt_len :] = input[
363 n_backward, 1 : 1 + self.answer_len
365 _, correct[n_backward] = compute_accuracy(back_input)
367 if log_prefix is not None:
368 forward_nb_correct = correct[n_forward].sum()
369 forward_nb_total = correct[n_forward].size(0)
370 backward_nb_correct = correct[n_backward].sum()
371 backward_nb_total = correct[n_backward].size(0)
374 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} %)"
378 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} %)"
381 return result, correct
383 compute_accuracy(model.train_w_quizzes[:nmax], log_prefix="train")
385 test_result, test_correct = compute_accuracy(
386 model.test_w_quizzes[:nmax], log_prefix="test"
389 main_test_accuracy = test_correct.sum() / test_correct.size(0)
390 self.logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
392 ##############################
396 f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
397 quizzes=test_result[:72],
398 mistakes=test_correct[:72] * 2 - 1,
401 return main_test_accuracy
403 def renew_w_quizzes(self, model, nb, for_train=True):
404 input = model.train_w_quizzes if for_train else model.test_w_quizzes
405 nb = min(nb, input.size(0))
406 input[:-nb] = input[nb:].clone()
407 fresh_w_quizzes = self.generate_token_sequences(nb)
408 self.reverse_random_half_in_place(fresh_w_quizzes)
409 input[-nb:] = fresh_w_quizzes.to(self.device)
411 def store_c_quizzes(self, new_c_quizzes, for_train=True):
413 self.train_c_quizzes.append(new_c_quizzes)
415 self.test_c_quizzes.append(new_c_quizzes)
417 def logproba_solution(self, models, c_quizzes):
418 logproba = c_quizzes.new_zeros(c_quizzes.size(0), len(models))
422 c_quizzes.split(self.batch_size), logproba.split(self.batch_size)
424 ar_mask = self.make_ar_mask(input)
425 output = model(mygpt.BracketedSequence(input)).x
427 F.cross_entropy(output.transpose(1, 2), input, reduction="none")
430 l[:, model.id] = -ce.sum(dim=-1)
434 ###############################################################
436 def compute_correctness(
439 models_for_validation,
440 bidirectional_validation=False,
441 deterministic_validation=True,
443 if bidirectional_validation:
444 backward_c_quizzes = self.forward_to_backward(c_quizzes)
446 seq_logproba = torch.zeros(
448 max([m.id for m in models_for_validation]) + 1,
454 seq_logproba[...] = 0.0
456 for model in models_for_validation:
457 result = c_quizzes.clone()
459 ar_mask = self.make_ar_mask(result)
461 masked_inplace_autoregression(
463 batch_size=self.batch_size,
466 seq_logproba=seq_logproba[:, model.id],
468 deterministic_synthesis=deterministic_validation,
469 # progress_bar_desc="solving c_quizzes",
473 correct = (c_quizzes == result).long().min(dim=-1).values
475 if bidirectional_validation:
476 backward_result = backward_c_quizzes.clone()
478 ar_mask = self.make_ar_mask(backward_result)
480 masked_inplace_autoregression(
482 batch_size=self.batch_size,
483 input=backward_result,
485 seq_logproba=seq_logproba[:, model.id],
487 deterministic_synthesis=deterministic_validation,
488 # progress_bar_desc="solving backward c_quizzes",
493 (backward_c_quizzes == backward_result).long().min(dim=-1).values
496 correct *= backward_correct
500 nb_correct += correct
502 return nb_correct, seq_logproba
504 ###############################################################
506 def generate_quizzes(self, nb, model_for_generation, temperature=1.0):
507 c_quizzes = torch.empty(
509 self.prompt_len + self.answer_len + 2,
514 seq_logproba = torch.zeros(nb, device=self.device)
516 # First, we generate the answer at high temperature
518 c_quizzes[:, 0] = self.token_backward
519 c_quizzes[:, 1 + self.answer_len] = self.token_backward
521 masked_inplace_autoregression(
522 model=model_for_generation,
523 batch_size=self.batch_size,
525 ar_mask=self.make_ar_mask(c_quizzes, first=True),
526 seq_logproba=seq_logproba,
527 temperature=temperature,
528 deterministic_synthesis=False,
532 # Then, we generate the prompt at low temperature
534 masked_inplace_autoregression(
535 model=model_for_generation,
536 batch_size=self.batch_size,
538 ar_mask=self.make_ar_mask(c_quizzes),
539 seq_logproba=seq_logproba,
540 temperature=1 / temperature,
541 deterministic_synthesis=False,
545 # Then we return the quizz, and re-generate the response, now
548 c_quizzes = self.reverse_time(c_quizzes)
550 masked_inplace_autoregression(
551 model=model_for_generation,
552 batch_size=self.batch_size,
554 ar_mask=self.make_ar_mask(c_quizzes),
555 seq_logproba=seq_logproba,
556 temperature=1 / temperature,
557 deterministic_synthesis=False,