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, sys
10 import torch, torchvision
13 from torch.nn import functional as F
16 from mygpt import BracketedSequence
20 ######################################################################
21 # if output is log(P(X=y)) and target is Y, returns -log P(X=Y) + H(X
25 # output is NxCxT and target is NxT
26 def confusion(output, target, reduction="mean"):
27 N, C, T = output.shape
28 output = output.permute(0, 2, 1).reshape(-1, C)
29 target = target.flatten()
30 all_t = torch.arange(N * T, device=output.device)
31 output = output.log_softmax(dim=-1)
32 result = -output[all_t, target]
34 output[all_t, target] = float("-inf")
35 output = output.log_softmax(dim=-1)
37 output[all_t, target] = 0
38 result = result - (output * e).sum(-1)
40 if reduction == "none":
41 return result.reshape(N, T)
42 elif reduction == "mean":
43 return result.reshape(N, T).mean()
44 elif reduction == "sum":
45 return result.reshape(N, T).sum()
47 raise ValueError(f"unknown reduction '{reduction}'.")
50 ######################################################################
52 # ar_mask is a tensor with 0s and 1s, of same shape as input, with
53 # 1s where tokens should be generated. The others are kept
57 def one_batch_masked_inplace_autoregression(
63 deterministic_synthesis,
65 to_generate = (ar_mask.sum(0) > 0).nonzero()
67 if to_generate.min() > 0:
69 BracketedSequence(input, 0, to_generate.min())
70 ) # Needed to initialize the model's cache
71 for s in range(to_generate.min(), to_generate.max() + 1):
72 output = model(BracketedSequence(input, s, 1)).x
76 logits = (logits / temperature).log_softmax(dim=-1)
78 if deterministic_synthesis:
79 t_next = logits.argmax(-1)
81 dist = torch.distributions.categorical.Categorical(logits=logits)
82 t_next = dist.sample()
84 all_n = torch.arange(t_next.size(0))
86 seq_logproba += logits[all_n, t_next]
88 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
91 def masked_inplace_autoregression(
98 deterministic_synthesis,
99 forbidden_tokens=None,
101 progress_bar_desc=None,
102 device=torch.device("cpu"),
104 assert input.size() == ar_mask.size()
107 input.split(batch_size),
108 ar_mask.split(batch_size),
109 seq_logproba.split(batch_size),
112 if progress_bar_desc is not None:
116 desc=progress_bar_desc,
117 total=(input.size(0) + batch_size - 1) // batch_size,
120 with torch.autograd.no_grad():
124 for input, ar_mask, seq_logproba in batches:
125 one_batch_masked_inplace_autoregression(
129 seq_logproba=seq_logproba,
130 temperature=temperature,
131 deterministic_synthesis=deterministic_synthesis,
137 ######################################################################
141 def indices_forward_and_backward(self, quizzes):
142 i_forward = quizzes[:, 0] == self.token_forward
143 j_forward = quizzes[:, 1 + self.prompt_len] == self.token_forward
144 i_backward = quizzes[:, 0] == self.token_backward
145 j_backward = quizzes[:, 1 + self.answer_len] == self.token_backward
146 assert torch.logical_or(
147 torch.logical_and(i_forward, j_forward),
148 torch.logical_and(i_backward, j_backward),
150 return i_forward, i_backward
152 def non_trivial(self, quizzes):
153 quizzes = quizzes.clone()
154 n_forward = quizzes[quizzes[:, 0] == self.token_forward]
155 n_backward = quizzes[:, 0] == self.token_backward
156 backward = quizzes[n_backward]
157 quizzes[n_backward] = self.reverse_time(quizzes[n_backward])
158 return torch.logical_not(
159 self.problem.trivial_prompts_and_answers(
160 quizzes[:, 1 : 1 + self.prompt_len],
161 quizzes[:, 2 + self.prompt_len :],
165 def reverse_time(self, quizzes):
166 i_forward, i_backward = self.indices_forward_and_backward(quizzes)
168 forward_to_backward = torch.cat(
171 quizzes[:, 2 + self.prompt_len : 2 + self.prompt_len + self.answer_len],
172 quizzes[:, 1 + self.prompt_len : 1 + self.prompt_len + 1],
173 quizzes[:, 1 : 1 + self.prompt_len],
178 forward_to_backward[:, 0] = self.token_backward
179 forward_to_backward[:, 1 + self.answer_len] = self.token_backward
181 backward_to_forward = torch.cat(
184 quizzes[:, 2 + self.answer_len :],
185 quizzes[:, 1 + self.answer_len : 2 + self.answer_len],
186 quizzes[:, 1 : 1 + self.answer_len],
191 backward_to_forward[:, 0] = self.token_forward
192 backward_to_forward[:, 1 + self.prompt_len] = self.token_forward
194 m = i_forward.long()[:, None]
196 return m * forward_to_backward + (1 - m) * backward_to_forward
198 def reverse_random_half_in_place(self, quizzes):
199 i = torch.rand(quizzes.size(0)) < 0.5
201 quizzes[i] = self.reverse_time(quizzes[i])
203 def make_ar_mask(self, quizzes, first=False):
204 i_forward, i_backward = self.indices_forward_and_backward(quizzes)
206 t = torch.arange(quizzes.size(1), device=quizzes.device)
209 m_forward = (t >= 1).long() * (t < 1 + self.prompt_len).long()
210 m_backward = (t >= 1).long() * (t < 1 + self.answer_len).long()
212 m_forward = (t >= 2 + self.prompt_len).long()
213 m_backward = (t >= 2 + self.answer_len).long()
215 m = i_forward.long()[:, None]
217 return m * m_forward + (1 - m) * m_backward
219 def generate_token_sequences(self, nb):
220 prompts, answers = self.problem.generate_prompts_and_answers(nb)
222 if self.prompt_len is None:
223 self.prompt_len = prompts.size(1)
225 if self.answer_len is None:
226 self.answer_len = answers.size(1)
228 assert prompts.size(1) == self.prompt_len and answers.size(1) == self.answer_len
232 for prompt, answer in zip(prompts, answers):
234 torch.tensor([self.token_forward]),
236 torch.tensor([self.token_forward]),
240 result.append(torch.cat(a, dim=0)[None, :])
242 return torch.cat(result, dim=0)
253 device=torch.device("cpu"),
257 v = problem.nb_token_values()
258 self.token_forward = v
259 self.token_backward = v + 1
260 self.nb_token_values = v + 2
262 self.problem = problem
263 self.back_accuracy = back_accuracy
264 self.batch_size = batch_size
267 self.prompt_len = None
268 self.answer_len = None
270 self.LOCK_C_QUIZZES = threading.Lock()
271 self.train_c_quizzes = []
272 self.test_c_quizzes = []
274 def save_quiz_illustrations(
281 quizzes = quizzes.clone().to("cpu")
282 n_forward = quizzes[quizzes[:, 0] == self.token_forward]
283 n_backward = quizzes[:, 0] == self.token_backward
284 backward = quizzes[n_backward]
285 assert n_forward.size(0) + backward.size(0) == quizzes.size(0)
286 quizzes[n_backward] = self.reverse_time(quizzes[n_backward])
288 predicted_prompts = n_backward.long()
289 predicted_answers = 1 - predicted_prompts
290 if mistakes is not None:
291 # 0/-1/+1 ~ not-to-predict / predicted wrong / predicted correct
292 predicted_prompts *= mistakes.to("cpu")
293 predicted_answers *= mistakes.to("cpu")
295 # 0/2 ~ not-to-predict / to predict
296 predicted_prompts *= 2
297 predicted_answers *= 2
299 self.problem.save_quiz_illustrations(
302 quizzes[:, 1 : 1 + self.prompt_len],
303 quizzes[:, 2 + self.prompt_len :],
308 def vocabulary_size(self):
309 return self.nb_token_values
311 ######################################################################
313 def batches(self, model, split="train", desc=None):
314 assert split in {"train", "test"}
316 with self.LOCK_C_QUIZZES:
318 w_quizzes = model.train_w_quizzes
319 c_quizzes = self.train_c_quizzes
321 w_quizzes = model.test_w_quizzes
322 c_quizzes = self.test_c_quizzes
324 if len(c_quizzes) > 0:
325 c_quizzes = torch.cat(c_quizzes, dim=0)
326 if c_quizzes.size(0) > w_quizzes.size(0) // 2:
327 i = torch.randperm(c_quizzes.size(0))[: w_quizzes.size(0) // 2]
328 c_quizzes = c_quizzes[i]
330 i = torch.randperm(w_quizzes.size(0))[
331 : w_quizzes.size(0) - c_quizzes.size(0)
333 w_quizzes = w_quizzes[i]
335 self.nb_batch_w_quizzes = w_quizzes.size(0)
336 self.nb_batch_c_quizzes = c_quizzes.size(0)
338 input = torch.cat([w_quizzes, c_quizzes], dim=0)
341 self.nb_batch_w_quizzes = w_quizzes.size(0)
342 self.nb_batch_c_quizzes = 0
345 input = input[torch.randperm(input.size(0))]
348 desc = f"epoch-{split}"
349 for batch in tqdm.tqdm(
350 input.split(self.batch_size), dynamic_ncols=True, desc=desc
354 ######################################################################
357 self, n_epoch, model, result_dir, deterministic_synthesis, nmax=1000
359 def compute_accuracy(input, log_prefix=None):
360 input = input.to(self.device)
361 ar_mask = self.make_ar_mask(input)
362 result = input.clone() * (1 - ar_mask)
363 seq_logproba = torch.empty(input.size(0), device=self.device)
365 masked_inplace_autoregression(
367 batch_size=self.batch_size,
370 seq_logproba=seq_logproba,
372 deterministic_synthesis=deterministic_synthesis,
373 progress_bar_desc=None,
377 correct = torch.empty(input.size(0), dtype=torch.int64, device=input.device)
379 n_forward = input[:, 0] == self.token_forward
380 n_backward = input[:, 0] == self.token_backward
382 correct[n_forward] = (
383 (input[n_forward] == result[n_forward]).long().min(dim=1).values
386 if self.back_accuracy and n_backward.any():
387 # accuracy of B->A*->B*=B instead of B->A*=A
388 back_input = self.reverse_time(result[n_backward])
389 back_input[:, 2 + self.prompt_len :] = input[
390 n_backward, 1 : 1 + self.answer_len
392 _, correct[n_backward] = compute_accuracy(back_input)
394 if log_prefix is not None:
395 forward_nb_correct = correct[n_forward].sum()
396 forward_nb_total = correct[n_forward].size(0)
397 backward_nb_correct = correct[n_backward].sum()
398 backward_nb_total = correct[n_backward].size(0)
401 f"{log_prefix}_accuracy {n_epoch} model {model.id} forward {forward_nb_correct} / {forward_nb_total} backward {backward_nb_correct} / {backward_nb_total}"
404 return result, correct
406 # compute_accuracy(model.train_w_quizzes[:nmax], log_prefix="train")
408 test_result, test_correct = compute_accuracy(
409 model.test_w_quizzes[:nmax], log_prefix="test"
412 main_test_accuracy = test_correct.sum() / test_correct.size(0)
413 self.logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
415 ##############################
417 self.save_quiz_illustrations(
419 f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
420 quizzes=test_result[:72],
421 mistakes=test_correct[:72] * 2 - 1,
424 return main_test_accuracy
426 ######################################################################
428 def renew_w_quizzes(self, model, nb, for_train=True):
429 input = model.train_w_quizzes if for_train else model.test_w_quizzes
430 nb = min(nb, input.size(0))
431 input[:-nb] = input[nb:].clone()
432 fresh_w_quizzes = self.generate_token_sequences(nb)
433 self.reverse_random_half_in_place(fresh_w_quizzes)
434 input[-nb:] = fresh_w_quizzes.to("cpu")
436 ######################################################################
438 def store_c_quizzes(self, new_c_quizzes, for_train=True):
439 with self.LOCK_C_QUIZZES:
441 self.train_c_quizzes.append(new_c_quizzes.to("cpu"))
443 self.test_c_quizzes.append(new_c_quizzes.to("cpu"))
445 def save_c_quizzes(self, filename):
446 torch.save((self.train_c_quizzes, self.test_c_quizzes), filename)
448 def load_c_quizzes(self, filename):
449 self.train_c_quizzes, self.test_c_quizzes = torch.load(filename)
451 ######################################################################
453 def logproba_of_solutions(self, models, c_quizzes):
454 logproba = c_quizzes.new_zeros(
455 c_quizzes.size(0), len(models), device=self.device, dtype=torch.float32
459 with torch.autograd.no_grad():
464 c_quizzes.split(self.batch_size), logproba.split(self.batch_size)
466 input = input.to(self.device)
467 ar_mask = self.make_ar_mask(input)
468 output = model(mygpt.BracketedSequence(input)).x
470 F.cross_entropy(output.transpose(1, 2), input, reduction="none")
473 l[:, model.id] = -ce.sum(dim=-1)
477 return logproba.to("cpu")
479 ###############################################################
481 def compute_correctness(
484 models_for_validation,
485 bidirectional_validation=False,
486 deterministic_validation=True,
488 if bidirectional_validation:
489 backward_c_quizzes = self.forward_to_backward(c_quizzes)
491 seq_logproba = torch.zeros(
493 max([m.id for m in models_for_validation]) + 1,
499 seq_logproba[...] = 0.0
501 for model in models_for_validation:
502 result = c_quizzes.clone()
504 ar_mask = self.make_ar_mask(result)
506 masked_inplace_autoregression(
508 batch_size=self.batch_size,
511 seq_logproba=seq_logproba[:, model.id],
513 deterministic_synthesis=deterministic_validation,
514 # progress_bar_desc="solving c_quizzes",
518 correct = (c_quizzes == result).long().min(dim=-1).values
520 if bidirectional_validation:
521 backward_result = backward_c_quizzes.clone()
523 ar_mask = self.make_ar_mask(backward_result)
525 masked_inplace_autoregression(
527 batch_size=self.batch_size,
528 input=backward_result,
530 seq_logproba=seq_logproba[:, model.id],
532 deterministic_synthesis=deterministic_validation,
533 # progress_bar_desc="solving backward c_quizzes",
538 (backward_c_quizzes == backward_result).long().min(dim=-1).values
541 correct *= backward_correct
545 nb_correct += correct
547 return nb_correct, seq_logproba
549 ###############################################################
551 def generate_quizzes(self, nb, model_for_generation, temperature=1.0):
552 c_quizzes = torch.empty(
554 self.prompt_len + self.answer_len + 2,
559 seq_logproba = torch.zeros(nb, device=self.device)
561 # First, we generate the answer at high temperature
563 c_quizzes[:, 0] = self.token_backward
564 c_quizzes[:, 1 + self.answer_len] = self.token_backward
566 masked_inplace_autoregression(
567 model=model_for_generation,
568 batch_size=self.batch_size,
570 ar_mask=self.make_ar_mask(c_quizzes, first=True),
571 seq_logproba=seq_logproba,
572 temperature=temperature,
573 deterministic_synthesis=False,
577 # Then, we generate the prompt at low temperature
579 masked_inplace_autoregression(
580 model=model_for_generation,
581 batch_size=self.batch_size,
583 ar_mask=self.make_ar_mask(c_quizzes),
584 seq_logproba=seq_logproba,
585 temperature=1 / temperature,
586 deterministic_synthesis=False,
590 # Then we return the quizz, and re-generate the response, now
593 c_quizzes = self.reverse_time(c_quizzes)
595 masked_inplace_autoregression(
596 model=model_for_generation,
597 batch_size=self.batch_size,
599 ar_mask=self.make_ar_mask(c_quizzes),
600 seq_logproba=seq_logproba,
601 temperature=1 / temperature,
602 deterministic_synthesis=False,
606 return c_quizzes.to("cpu")