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,
32 forbidden_tokens=None,
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 forbidden_tokens is not None:
49 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
51 if forced_biases is not None:
52 logits = logits + forced_biases[None, :]
54 if deterministic_synthesis:
55 t_next = logits.argmax(-1)
57 dist = torch.distributions.categorical.Categorical(logits=logits)
58 t_next = dist.sample()
60 all_n = torch.arange(t_next.size(0))
61 seq_logproba += logits[all_n, t_next].sum(dim=-1)
63 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
66 def masked_inplace_autoregression(
73 deterministic_synthesis,
74 forbidden_tokens=None,
76 progress_bar_desc=None,
77 device=torch.device("cpu"),
79 assert input.size() == ar_mask.size()
82 input.split(batch_size),
83 ar_mask.split(batch_size),
84 seq_logproba.split(batch_size),
87 if progress_bar_desc is not None:
91 desc=progress_bar_desc,
92 total=(input.size(0) + batch_size - 1) // batch_size,
95 with torch.autograd.no_grad():
99 for input, ar_mask, seq_logproba in batches:
100 one_batch_masked_inplace_autoregression(
104 seq_logproba=seq_logproba,
105 temperature=temperature,
106 deterministic_synthesis=deterministic_synthesis,
107 forbidden_tokens=forbidden_tokens,
108 forced_biases=logit_biases,
114 ######################################################################
118 def make_ar_mask(self, input):
119 b = torch.arange(input.size(1), device=input.device) > input.size(1) // 2
120 return b.long()[None, :].expand_as(input)
122 def generate_token_sequences(self, nb):
123 prompts, answers = self.problem.generate_prompts_and_answers(nb)
126 for prompt, answer in zip(prompts, answers):
127 if torch.rand(1) < 0.5:
128 a = [torch.tensor([self.token_forward]), prompt, answer]
130 a = [torch.tensor([self.token_backward]), answer, prompt]
132 result.append(torch.cat(a, dim=0)[None, :])
134 return torch.cat(result, dim=0)
144 device=torch.device("cpu"),
148 v = problem.nb_token_values()
149 self.token_forward = v
150 self.token_backward = v + 1
151 self.nb_token_values = v + 2
153 self.problem = problem
154 self.batch_size = batch_size
158 self.train_w_quizzes = self.generate_token_sequences(nb_train_samples).to(
162 self.test_w_quizzes = self.generate_token_sequences(nb_test_samples).to(device)
164 self.train_c_quizzes = []
165 self.test_c_quizzes = []
167 if result_dir is not None:
169 result_dir, "culture_w_quizzes", self.train_w_quizzes[:72]
172 def save_quizzes(self, result_dir, filename_prefix, quizzes, prediction=False):
173 print(f"DEBUG {quizzes.size()=}")
174 l = (quizzes.size(1) - 1) // 2
175 forward = (quizzes[:, 0] == self.token_forward).long()
176 backward = (quizzes[:, 0] == self.token_backward).long()
177 assert forward.equal(1 - backward)
178 first = quizzes[:, 1 : 1 + l]
179 second = quizzes[:, 1 + l : 1 + 2 * l]
180 prompts = forward[:, None] * first + backward[:, None] * second
181 answers = forward[:, None] * second + backward[:, None] * first
184 predicted_prompts = backward
185 predicted_answers = forward
187 predicted_prompts = None
188 predicted_answers = None
190 self.problem.save_quizzes(
199 def batches(self, split="train", desc=None):
200 assert split in {"train", "test"}
202 w_quizzes = self.train_w_quizzes
203 c_quizzes = self.train_c_quizzes
205 w_quizzes = self.test_w_quizzes
206 c_quizzes = self.test_c_quizzes
208 if len(c_quizzes) > 0:
209 c_quizzes = torch.cat(c_quizzes, dim=0)
210 if c_quizzes.size(0) > w_quizzes.size(0) // 2:
211 i = torch.randperm(c_quizzes.size(0))[: w_quizzes.size(0) // 2]
212 c_quizzes = c_quizzes[i]
214 i = torch.randperm(w_quizzes.size(0))[
215 : w_quizzes.size(0) - c_quizzes.size(0)
217 w_quizzes = w_quizzes[i]
219 self.nb_batch_w_quizzes = w_quizzes.size(0)
220 self.nb_batch_c_quizzes = c_quizzes.size(0)
222 input = torch.cat([w_quizzes, c_quizzes], dim=0)
225 self.nb_batch_w_quizzes = w_quizzes.size(0)
226 self.nb_batch_c_quizzes = 0
229 input = input[torch.randperm(input.size(0))]
232 desc = f"epoch-{split}"
233 for batch in tqdm.tqdm(
234 input.split(self.batch_size), dynamic_ncols=True, desc=desc
238 def vocabulary_size(self):
239 return self.nb_token_values
242 self, n_epoch, model, result_dir, deterministic_synthesis, nmax=1000
244 def compute_accuracy(input):
246 ar_mask = self.make_ar_mask(input)
247 result = input.clone() * (1 - ar_mask)
248 seq_logproba = torch.empty(input.size(0), device=self.device)
250 masked_inplace_autoregression(
252 batch_size=self.batch_size,
255 seq_logproba=seq_logproba,
257 deterministic_synthesis=deterministic_synthesis,
258 progress_bar_desc=None,
262 nb_total, nb_correct = (
264 (input == result).long().min(dim=1).values.sum(),
267 return nb_total, nb_correct
269 train_nb_total, train_nb_correct = compute_accuracy(self.train_w_quizzes)
272 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}%"
275 test_nb_total, test_nb_correct = compute_accuracy(self.test_w_quizzes)
278 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}%"
281 main_test_accuracy = test_nb_correct / test_nb_total
282 self.logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
284 ##############################
286 input = self.test_w_quizzes[:96]
287 ar_mask = self.make_ar_mask(input)
288 result = input.clone() * (1 - ar_mask)
289 seq_logproba = torch.empty(input.size(0), device=self.device)
291 masked_inplace_autoregression(
293 batch_size=self.batch_size,
296 seq_logproba=seq_logproba,
298 deterministic_synthesis=deterministic_synthesis,
299 progress_bar_desc=None,
305 f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
310 return main_test_accuracy
312 def renew_w_quizzes(self, nb, for_train=True):
313 input = self.train_w_quizzes if for_train else self.test_w_quizzes
314 nb = min(nb, input.size(0))
315 input[:-nb] = input[nb:].clone()
316 input[-nb:] = self.generate_token_sequences(nb).to(self.device)
318 def store_c_quizzes(self, new_c_quizzes, for_train=True):
320 self.train_c_quizzes.append(new_c_quizzes)
322 self.test_c_quizzes.append(new_c_quizzes)
324 def reverse_time(self, c_quizzes):
325 l = (c_quizzes.size(1) - 1) // 2
326 direction = c_quizzes[:, 0:1]
327 direction = self.token_forward * (
328 direction == self.token_backward
329 ) + self.token_backward * (direction == self.token_forward)
332 [direction, c_quizzes[:, l + 1 :], c_quizzes[:, 1 : l + 1]], dim=1
335 def compute_correctness(
336 self, c_quizzes, models_for_validation, both_directions=True
338 reversed_c_quizzes = self.reverse_time(c_quizzes)
340 ar_mask = self.make_ar_mask(c_quizzes)
341 seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
343 # Check how many of models can solve the quizzes in both directions
347 for model in models_for_validation:
348 result = c_quizzes.clone()
350 masked_inplace_autoregression(
352 batch_size=self.batch_size,
355 seq_logproba=seq_logproba,
357 deterministic_synthesis=True,
358 # progress_bar_desc="solving c_quizzes",
362 correct = (c_quizzes == result).long().min(dim=-1).values
365 reversed_result = reversed_c_quizzes.clone()
367 masked_inplace_autoregression(
369 batch_size=self.batch_size,
370 input=reversed_result,
372 seq_logproba=seq_logproba,
374 deterministic_synthesis=True,
375 # progress_bar_desc="solving reversed c_quizzes",
380 (reversed_c_quizzes == reversed_result).long().min(dim=-1).values
383 correct *= reversed_correct
387 nb_correct += correct
391 ###############################################################
393 def generate_quizzes(self, nb, model_for_generation, reverse_cleanup=False):
394 c_quizzes = torch.empty(
395 nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
398 c_quizzes[:, 0] = self.token_forward
400 ar_mask_first = torch.zeros(c_quizzes.size(), device=self.device)
401 ar_mask_first[:, : ar_mask_first.size(1) // 2 + 1] = 1
402 ar_mask_second = 1 - ar_mask_first
403 ar_mask_first[:, 0] = 0
404 ar_mask_second[:, 0] = 0
406 seq_logproba = torch.empty(ar_mask_first.size(0), device=self.device)
409 warnings.warn("very high temperature with reversed cleanup", RuntimeWarning)
414 # warnings.warn("noise injection", RuntimeWarning)
415 # noise_std = torch.rand(1).item()
416 # self.logger(f"{noise_std=}")
418 # mygpt.set_noise_injection(model_for_generation, noise_std)
420 masked_inplace_autoregression(
421 model=model_for_generation,
422 batch_size=self.batch_size,
424 ar_mask=ar_mask_first,
425 seq_logproba=seq_logproba,
426 temperature=temperature,
427 deterministic_synthesis=False,
431 # mygpt.set_noise_injection(model_for_generation, 0.0)
433 ave_seq_logproba = seq_logproba.mean()
435 masked_inplace_autoregression(
436 model=model_for_generation,
437 batch_size=self.batch_size,
439 ar_mask=ar_mask_second,
440 seq_logproba=seq_logproba,
441 temperature=temperature,
442 deterministic_synthesis=True,
447 c_quizzes = self.reverse_time(c_quizzes)
449 masked_inplace_autoregression(
450 model=model_for_generation,
451 batch_size=self.batch_size,
453 ar_mask=ar_mask_second,
454 seq_logproba=seq_logproba,
455 temperature=temperature,
456 deterministic_synthesis=True,
460 c_quizzes = self.reverse_time(c_quizzes)
462 masked_inplace_autoregression(
463 model=model_for_generation,
464 batch_size=self.batch_size,
466 ar_mask=ar_mask_second,
467 seq_logproba=seq_logproba,
468 temperature=temperature,
469 deterministic_synthesis=True,
473 return c_quizzes, seq_logproba.mean()