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 reverse_random_half_in_place(self, quizzes):
153 i = torch.rand(quizzes.size(0)) < 0.5
155 quizzes[i] = self.reverse_time(quizzes[i])
157 def make_ar_mask(self, quizzes, first=False):
158 i_forward, i_backward = self.indices_forward_and_backward(quizzes)
160 t = torch.arange(quizzes.size(1), device=quizzes.device)
163 m_forward = (t >= 1).long() * (t < 1 + self.prompt_len).long()
164 m_backward = (t >= 1).long() * (t < 1 + self.answer_len).long()
166 m_forward = (t >= 2 + self.prompt_len).long()
167 m_backward = (t >= 2 + self.answer_len).long()
169 m = i_forward.long()[:, None]
171 return m * m_forward + (1 - m) * m_backward
173 def generate_token_sequences(self, nb):
174 prompts, answers = self.problem.generate_prompts_and_answers(nb)
176 if self.prompt_len is None:
177 self.prompt_len = prompts.size(1)
179 if self.answer_len is None:
180 self.answer_len = answers.size(1)
182 assert prompts.size(1) == self.prompt_len and answers.size(1) == self.answer_len
186 for prompt, answer in zip(prompts, answers):
188 torch.tensor([self.token_forward]),
190 torch.tensor([self.token_forward]),
194 result.append(torch.cat(a, dim=0)[None, :])
196 return torch.cat(result, dim=0)
207 device=torch.device("cpu"),
211 v = problem.nb_token_values()
212 self.token_forward = v
213 self.token_backward = v + 1
214 self.nb_token_values = v + 2
216 self.problem = problem
217 self.back_accuracy = back_accuracy
218 self.batch_size = batch_size
221 self.prompt_len = None
222 self.answer_len = None
224 self.train_w_quizzes = self.generate_token_sequences(nb_train_samples)
225 self.reverse_random_half_in_place(self.train_w_quizzes)
226 self.train_w_quizzes = self.train_w_quizzes.to(device)
228 self.test_w_quizzes = self.generate_token_sequences(nb_test_samples).to(device)
229 self.reverse_random_half_in_place(self.test_w_quizzes)
230 self.test_w_quizzes = self.test_w_quizzes.to(device)
232 self.train_c_quizzes = []
233 self.test_c_quizzes = []
235 if result_dir is not None:
239 self.train_w_quizzes[:72],
249 quizzes = quizzes.clone()
250 n_forward = quizzes[quizzes[:, 0] == self.token_forward]
251 n_backward = quizzes[:, 0] == self.token_backward
252 backward = quizzes[n_backward]
253 assert n_forward.size(0) + backward.size(0) == quizzes.size(0)
254 quizzes[n_backward] = self.reverse_time(quizzes[n_backward])
256 predicted_prompts = n_backward.long()
257 predicted_answers = 1 - predicted_prompts
258 if mistakes is not None:
259 # 0/-1/+1 ~ not-to-predict / predicted wrong / predicted correct
260 predicted_prompts *= mistakes
261 predicted_answers *= mistakes
263 # 0/2 ~ not-to-predict / to predict
264 predicted_prompts *= 2
265 predicted_answers *= 2
267 self.problem.save_quizzes(
270 quizzes[:, 1 : 1 + self.prompt_len],
271 quizzes[:, 2 + self.prompt_len :],
276 def batches(self, split="train", desc=None):
277 assert split in {"train", "test"}
279 w_quizzes = self.train_w_quizzes
280 c_quizzes = self.train_c_quizzes
282 w_quizzes = self.test_w_quizzes
283 c_quizzes = self.test_c_quizzes
285 if len(c_quizzes) > 0:
286 c_quizzes = torch.cat(c_quizzes, dim=0)
287 if c_quizzes.size(0) > w_quizzes.size(0) // 2:
288 i = torch.randperm(c_quizzes.size(0))[: w_quizzes.size(0) // 2]
289 c_quizzes = c_quizzes[i]
291 i = torch.randperm(w_quizzes.size(0))[
292 : w_quizzes.size(0) - c_quizzes.size(0)
294 w_quizzes = w_quizzes[i]
296 self.nb_batch_w_quizzes = w_quizzes.size(0)
297 self.nb_batch_c_quizzes = c_quizzes.size(0)
299 input = torch.cat([w_quizzes, c_quizzes], dim=0)
302 self.nb_batch_w_quizzes = w_quizzes.size(0)
303 self.nb_batch_c_quizzes = 0
306 input = input[torch.randperm(input.size(0))]
309 desc = f"epoch-{split}"
310 for batch in tqdm.tqdm(
311 input.split(self.batch_size), dynamic_ncols=True, desc=desc
315 def vocabulary_size(self):
316 return self.nb_token_values
319 self, n_epoch, model, result_dir, deterministic_synthesis, nmax=1000
321 def compute_accuracy(input, log_prefix=None):
322 ar_mask = self.make_ar_mask(input)
323 result = input.clone() * (1 - ar_mask)
324 seq_logproba = torch.empty(input.size(0), device=self.device)
326 masked_inplace_autoregression(
328 batch_size=self.batch_size,
331 seq_logproba=seq_logproba,
333 deterministic_synthesis=deterministic_synthesis,
334 progress_bar_desc=None,
338 correct = torch.empty(input.size(0), dtype=torch.int64, device=input.device)
340 n_forward = input[:, 0] == self.token_forward
341 n_backward = input[:, 0] == self.token_backward
343 correct[n_forward] = (
344 (input[n_forward] == result[n_forward]).long().min(dim=1).values
347 if self.back_accuracy and n_backward.any():
348 # accuracy of B->A*->B*=B instead of B->A*=A
349 back_input = self.reverse_time(result[n_backward])
350 back_input[:, 2 + self.prompt_len :] = input[
351 n_backward, 1 : 1 + self.answer_len
353 _, correct[n_backward] = compute_accuracy(back_input)
355 if log_prefix is not None:
356 forward_nb_correct = correct[n_forward].sum()
357 forward_nb_total = correct[n_forward].size(0)
358 backward_nb_correct = correct[n_backward].sum()
359 backward_nb_total = correct[n_backward].size(0)
362 f"{log_prefix}_forward_accuracy {n_epoch} model {model.id} nb_correct {forward_nb_correct} / {forward_nb_total}"
366 f"{log_prefix}_backward_accuracy {n_epoch} model {model.id} nb_correct {backward_nb_correct} / {backward_nb_total}"
369 return result, correct
371 compute_accuracy(self.train_w_quizzes[:nmax], log_prefix="train")
373 test_result, test_correct = compute_accuracy(
374 self.test_w_quizzes[:nmax], log_prefix="test"
377 main_test_accuracy = test_correct.sum() / test_correct.size(0)
378 self.logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
380 ##############################
384 f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
385 quizzes=test_result[:72],
386 mistakes=test_correct[:72] * 2 - 1,
389 return main_test_accuracy
391 def renew_w_quizzes(self, nb, for_train=True):
392 input = self.train_w_quizzes if for_train else self.test_w_quizzes
393 nb = min(nb, input.size(0))
394 input[:-nb] = input[nb:].clone()
395 fresh_w_quizzes = self.generate_token_sequences(nb)
396 self.reverse_random_half_in_place(fresh_w_quizzes)
397 input[-nb:] = fresh_w_quizzes.to(self.device)
399 def store_c_quizzes(self, new_c_quizzes, for_train=True):
401 self.train_c_quizzes.append(new_c_quizzes)
403 self.test_c_quizzes.append(new_c_quizzes)
405 def compute_correctness(
408 models_for_validation,
409 bidirectional_validation=False,
410 deterministic_validation=True,
412 if bidirectional_validation:
413 backward_c_quizzes = self.forward_to_backward(c_quizzes)
415 seq_logproba = torch.zeros(
417 max([m.id for m in models_for_validation]) + 1,
423 for model in models_for_validation:
424 result = c_quizzes.clone()
426 seq_logproba[...] = 0.0
428 ar_mask = self.make_ar_mask(result)
430 masked_inplace_autoregression(
432 batch_size=self.batch_size,
435 seq_logproba=seq_logproba[:, model.id],
437 deterministic_synthesis=deterministic_validation,
438 # progress_bar_desc="solving c_quizzes",
442 correct = (c_quizzes == result).long().min(dim=-1).values
444 if bidirectional_validation:
445 backward_result = backward_c_quizzes.clone()
447 ar_mask = self.make_ar_mask(backward_result)
449 masked_inplace_autoregression(
451 batch_size=self.batch_size,
452 input=backward_result,
454 seq_logproba=seq_logproba[:, model.id],
456 deterministic_synthesis=deterministic_validation,
457 # progress_bar_desc="solving backward c_quizzes",
462 (backward_c_quizzes == backward_result).long().min(dim=-1).values
465 correct *= backward_correct
469 nb_correct += correct
471 return nb_correct, seq_logproba
473 ###############################################################
475 def generate_quizzes(self, nb, model_for_generation, temperature=1.0):
476 c_quizzes = torch.empty(
477 nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
480 seq_logproba = torch.zeros(nb, device=self.device)
482 # First, we generate the answer at high temperature
484 c_quizzes[:, 0] = self.token_backward
485 c_quizzes[:, 1 + self.answer_len] = self.token_backward
487 masked_inplace_autoregression(
488 model=model_for_generation,
489 batch_size=self.batch_size,
491 ar_mask=self.make_ar_mask(c_quizzes, first=True),
492 seq_logproba=seq_logproba,
493 temperature=temperature,
494 deterministic_synthesis=False,
498 # Then, we generate the prompt at low temperature
500 masked_inplace_autoregression(
501 model=model_for_generation,
502 batch_size=self.batch_size,
504 ar_mask=self.make_ar_mask(c_quizzes),
505 seq_logproba=seq_logproba,
506 temperature=1 / temperature,
507 deterministic_synthesis=False,
511 # Then we return the quizz, and re-generate the response, now
514 c_quizzes = self.reverse_time(c_quizzes)
516 masked_inplace_autoregression(
517 model=model_for_generation,
518 batch_size=self.batch_size,
520 ar_mask=self.make_ar_mask(c_quizzes),
521 seq_logproba=seq_logproba,
522 temperature=1 / temperature,
523 deterministic_synthesis=False,