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 :],
126 quizzes[:, 1 + self.prompt_len : 2 + self.prompt_len],
127 quizzes[:, 1 : 1 + self.prompt_len],
131 forward_to_backward[:, 0] = self.token_backward
132 forward_to_backward[:, 1 + self.answer_len] = self.token_backward
134 backward_to_forward = torch.cat(
137 quizzes[:, 2 + self.answer_len :],
138 quizzes[:, 1 + self.answer_len : 2 + self.answer_len],
139 quizzes[:, 1 : 1 + self.answer_len],
144 backward_to_forward[:, 0] = self.token_forward
145 backward_to_forward[:, 1 + self.prompt_len] = self.token_forward
147 m = i_forward.long()[:, None]
149 return m * forward_to_backward + (1 - m) * backward_to_forward
151 def make_ar_mask(self, quizzes, first=False):
152 i_forward, i_backward = self.indices_forward_and_backward(quizzes)
154 t = torch.arange(quizzes.size(1), device=quizzes.device)
157 m_forward = (t >= 1).long() * (t < 1 + self.prompt_len).long()
158 m_backward = (t >= 1).long() * (t < 1 + self.answer_len).long()
160 m_forward = (t >= 2 + self.prompt_len).long()
161 m_backward = (t >= 2 + self.answer_len).long()
163 m = i_forward.long()[:, None]
165 return m * m_forward + (1 - m) * m_backward
167 def generate_token_sequences(self, nb):
168 prompts, answers = self.problem.generate_prompts_and_answers(nb)
170 if self.prompt_len is None:
171 self.prompt_len = prompts.size(1)
173 if self.answer_len is None:
174 self.answer_len = answers.size(1)
176 assert prompts.size(1) == self.prompt_len and answers.size(1) == self.answer_len
180 for prompt, answer in zip(prompts, answers):
181 if torch.rand(1) < 0.5:
183 torch.tensor([self.token_forward]),
185 torch.tensor([self.token_forward]),
190 torch.tensor([self.token_backward]),
192 torch.tensor([self.token_backward]),
196 result.append(torch.cat(a, dim=0)[None, :])
198 return torch.cat(result, dim=0)
208 device=torch.device("cpu"),
212 v = problem.nb_token_values()
213 self.token_forward = v
214 self.token_backward = v + 1
215 self.nb_token_values = v + 2
217 self.problem = problem
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).to(
228 self.test_w_quizzes = self.generate_token_sequences(nb_test_samples).to(device)
230 self.train_c_quizzes = []
231 self.test_c_quizzes = []
233 if result_dir is not None:
235 result_dir, "culture_w_quizzes", self.train_w_quizzes[:72]
238 # toto = self.reverse_time(self.train_w_quizzes[:72])
239 # self.save_quizzes(result_dir, "toto", toto)
242 def save_quizzes(self, result_dir, filename_prefix, quizzes, prediction=False):
243 forward = quizzes[quizzes[:, 0] == self.token_forward]
244 ib = quizzes[:, 0] == self.token_backward
245 backward = quizzes[ib]
246 assert forward.size(0) + backward.size(0) == quizzes.size(0)
247 quizzes[ib] = self.reverse_time(quizzes[ib])
250 predicted_prompts = ib
251 predicted_answers = torch.logical_not(ib)
253 predicted_prompts = None
254 predicted_answers = None
256 self.problem.save_quizzes(
259 quizzes[:, 1 : 1 + self.prompt_len],
260 quizzes[:, 2 + self.prompt_len :],
265 def batches(self, split="train", desc=None):
266 assert split in {"train", "test"}
268 w_quizzes = self.train_w_quizzes
269 c_quizzes = self.train_c_quizzes
271 w_quizzes = self.test_w_quizzes
272 c_quizzes = self.test_c_quizzes
274 if len(c_quizzes) > 0:
275 c_quizzes = torch.cat(c_quizzes, dim=0)
276 if c_quizzes.size(0) > w_quizzes.size(0) // 2:
277 i = torch.randperm(c_quizzes.size(0))[: w_quizzes.size(0) // 2]
278 c_quizzes = c_quizzes[i]
280 i = torch.randperm(w_quizzes.size(0))[
281 : w_quizzes.size(0) - c_quizzes.size(0)
283 w_quizzes = w_quizzes[i]
285 self.nb_batch_w_quizzes = w_quizzes.size(0)
286 self.nb_batch_c_quizzes = c_quizzes.size(0)
288 input = torch.cat([w_quizzes, c_quizzes], dim=0)
291 self.nb_batch_w_quizzes = w_quizzes.size(0)
292 self.nb_batch_c_quizzes = 0
295 input = input[torch.randperm(input.size(0))]
298 desc = f"epoch-{split}"
299 for batch in tqdm.tqdm(
300 input.split(self.batch_size), dynamic_ncols=True, desc=desc
304 def vocabulary_size(self):
305 return self.nb_token_values
308 self, n_epoch, model, result_dir, deterministic_synthesis, nmax=1000
310 def compute_accuracy(input):
312 ar_mask = self.make_ar_mask(input)
313 result = input.clone() * (1 - ar_mask)
314 seq_logproba = torch.empty(input.size(0), device=self.device)
316 masked_inplace_autoregression(
318 batch_size=self.batch_size,
321 seq_logproba=seq_logproba,
323 deterministic_synthesis=deterministic_synthesis,
324 progress_bar_desc=None,
328 nb_total = input.size(0)
329 nb_correct = (input == result).long().min(dim=1).values.sum()
331 return nb_total, nb_correct
333 train_nb_total, train_nb_correct = compute_accuracy(self.train_w_quizzes)
336 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}%"
339 test_nb_total, test_nb_correct = compute_accuracy(self.test_w_quizzes)
342 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}%"
345 main_test_accuracy = test_nb_correct / test_nb_total
346 self.logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
348 ##############################
350 input = self.test_w_quizzes[:96]
351 ar_mask = self.make_ar_mask(input)
352 result = input.clone() * (1 - ar_mask)
353 seq_logproba = torch.empty(input.size(0), device=self.device)
355 masked_inplace_autoregression(
357 batch_size=self.batch_size,
360 seq_logproba=seq_logproba,
362 deterministic_synthesis=deterministic_synthesis,
363 progress_bar_desc=None,
369 f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
374 return main_test_accuracy
376 def renew_w_quizzes(self, nb, for_train=True):
377 input = self.train_w_quizzes if for_train else self.test_w_quizzes
378 nb = min(nb, input.size(0))
379 input[:-nb] = input[nb:].clone()
380 input[-nb:] = self.generate_token_sequences(nb).to(self.device)
382 def store_c_quizzes(self, new_c_quizzes, for_train=True):
384 self.train_c_quizzes.append(new_c_quizzes)
386 self.test_c_quizzes.append(new_c_quizzes)
388 def compute_correctness(
391 models_for_validation,
392 bidirectional_validation=False,
393 deterministic_validation=True,
395 if bidirectional_validation:
396 backward_c_quizzes = self.forward_to_backward(c_quizzes)
398 seq_logproba = torch.zeros(
400 max([m.id for m in models_for_validation]) + 1,
406 for model in models_for_validation:
407 result = c_quizzes.clone()
409 seq_logproba[...] = 0.0
411 ar_mask = self.make_ar_mask(result)
413 masked_inplace_autoregression(
415 batch_size=self.batch_size,
418 seq_logproba=seq_logproba[:, model.id],
420 deterministic_synthesis=deterministic_validation,
421 # progress_bar_desc="solving c_quizzes",
425 correct = (c_quizzes == result).long().min(dim=-1).values
427 if bidirectional_validation:
428 backward_result = backward_c_quizzes.clone()
430 ar_mask = self.make_ar_mask(backward_result)
432 masked_inplace_autoregression(
434 batch_size=self.batch_size,
435 input=backward_result,
437 seq_logproba=seq_logproba[:, model.id],
439 deterministic_synthesis=deterministic_validation,
440 # progress_bar_desc="solving backward c_quizzes",
445 (backward_c_quizzes == backward_result).long().min(dim=-1).values
448 correct *= backward_correct
452 nb_correct += correct
454 return nb_correct, seq_logproba
456 ###############################################################
458 def generate_quizzes(self, nb, model_for_generation, temperature=1.0):
459 c_quizzes = torch.empty(
460 nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
463 seq_logproba = torch.zeros(nb, device=self.device)
465 # First, we generate the answer at high temperature
467 c_quizzes[:, 0] = self.token_backward
468 c_quizzes[:, 1 + self.answer_len] = self.token_backward
470 masked_inplace_autoregression(
471 model=model_for_generation,
472 batch_size=self.batch_size,
474 ar_mask=self.make_ar_mask(c_quizzes, first=True),
475 seq_logproba=seq_logproba,
476 temperature=temperature,
477 deterministic_synthesis=False,
481 # Then, we generate the prompt at low temperature
483 masked_inplace_autoregression(
484 model=model_for_generation,
485 batch_size=self.batch_size,
487 ar_mask=self.make_ar_mask(c_quizzes),
488 seq_logproba=seq_logproba,
489 temperature=1 / temperature,
490 deterministic_synthesis=False,
494 # Then we return the quizz, and re-generate the response, now
497 c_quizzes = self.reverse_time(c_quizzes)
499 masked_inplace_autoregression(
500 model=model_for_generation,
501 batch_size=self.batch_size,
503 ar_mask=self.make_ar_mask(c_quizzes),
504 seq_logproba=seq_logproba,
505 temperature=1 / temperature,
506 deterministic_synthesis=False,