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
15 from mygpt import BracketedSequence
17 ######################################################################
20 class Gang(nn.Module):
21 def __init__(self, models, nb_models_for_generation, mode="groupthink"):
24 self.nb_models_for_generation = nb_models_for_generation
27 def forward(self, bs):
28 # If first = 0, we are re-starting an auto-regressive process,
29 # that's the right moment to randomize who gonna do it
31 self.models_to_use = [
33 for k in torch.randperm(len(self.models))[
34 : self.nb_models_for_generation
38 all_the_logits = torch.cat(
39 [model(bs).x[None] for model in self.models_to_use], dim=0
42 if self.mode == "groupthink":
43 y = all_the_logits.mean(dim=0)
44 elif self.mode == "groupwork":
45 m = torch.rand(all_the_logits.size(), device=all_the_logits.device)
46 m = (m.sort(dim=0).indices == 0).long()
47 y = (y * m).sum(dim=0)
49 raise ValueError(f"Invalid mode {self.mode}")
51 return BracketedSequence(y, bs.first, bs.nb)
54 ######################################################################
56 # ar_mask is a tensor with 0s and 1s, of same shape as input, with
57 # 1s where tokens should be generated. The others are kept
61 def one_batch_masked_inplace_autoregression(
67 deterministic_synthesis=False,
68 forbidden_tokens=None,
71 to_generate = (ar_mask.sum(0) > 0).nonzero()
73 if to_generate.min() > 0:
75 BracketedSequence(input, 0, to_generate.min())
76 ) # Needed to initialize the model's cache
77 for s in range(to_generate.min(), to_generate.max() + 1):
78 output = model(BracketedSequence(input, s, 1)).x
82 logits = (logits / temperature).log_softmax(dim=-1)
84 if forbidden_tokens is not None:
85 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
87 if forced_biases is not None:
88 logits = logits + forced_biases[None, :]
90 if deterministic_synthesis:
91 t_next = logits.argmax(-1)
93 dist = torch.distributions.categorical.Categorical(logits=logits)
94 t_next = dist.sample()
96 all_n = torch.arange(t_next.size(0))
97 seq_logproba += logits[all_n, t_next].sum(dim=-1)
99 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
102 def masked_inplace_autoregression(
109 deterministic_synthesis,
110 forbidden_tokens=None,
112 progress_bar_desc=None,
113 device=torch.device("cpu"),
115 assert input.size() == ar_mask.size()
118 input.split(batch_size),
119 ar_mask.split(batch_size),
120 seq_logproba.split(batch_size),
123 if progress_bar_desc is not None:
127 desc=progress_bar_desc,
128 total=(input.size(0) + batch_size - 1) // batch_size,
131 with torch.autograd.no_grad():
135 for input, ar_mask, seq_logproba in batches:
136 one_batch_masked_inplace_autoregression(
140 seq_logproba=seq_logproba,
141 temperature=temperature,
142 deterministic_synthesis=deterministic_synthesis,
143 forbidden_tokens=forbidden_tokens,
144 forced_biases=logit_biases,
150 ######################################################################
154 def make_ar_mask(self, input):
155 b = torch.arange(input.size(1), device=input.device) > input.size(1) // 2
156 return b.long()[None, :].expand_as(input)
166 device=torch.device("cpu"),
170 self.problem = problem
171 self.batch_size = batch_size
175 self.train_w_quizzes = self.problem.generate_token_sequences(
178 self.test_w_quizzes = self.problem.generate_token_sequences(nb_test_samples).to(
182 self.nb_codes = max(self.train_w_quizzes.max(), self.test_w_quizzes.max()) + 1
184 self.train_c_quizzes = []
185 self.test_c_quizzes = []
187 if result_dir is not None:
188 self.problem.save_quizzes(
189 self.train_w_quizzes[:72], result_dir, "culture_w_quizzes"
192 def batches(self, split="train", desc=None):
193 assert split in {"train", "test"}
195 w_quizzes = self.train_w_quizzes
196 c_quizzes = self.train_c_quizzes
198 w_quizzes = self.test_w_quizzes
199 c_quizzes = self.test_c_quizzes
201 if len(c_quizzes) > 0:
202 c_quizzes = torch.cat(c_quizzes, dim=0)
203 if c_quizzes.size(0) > w_quizzes.size(0) // 2:
204 i = torch.randperm(c_quizzes.size(0))[: w_quizzes.size(0) // 2]
205 c_quizzes = c_quizzes[i]
207 i = torch.randperm(w_quizzes.size(0))[
208 : w_quizzes.size(0) - c_quizzes.size(0)
210 w_quizzes = w_quizzes[i]
212 self.nb_batch_w_quizzes = w_quizzes.size(0)
213 self.nb_batch_c_quizzes = c_quizzes.size(0)
215 input = torch.cat([w_quizzes, c_quizzes], dim=0)
218 self.nb_batch_w_quizzes = w_quizzes.size(0)
219 self.nb_batch_c_quizzes = 0
222 input = input[torch.randperm(input.size(0))]
225 desc = f"epoch-{split}"
226 for batch in tqdm.tqdm(
227 input.split(self.batch_size), dynamic_ncols=True, desc=desc
231 def vocabulary_size(self):
235 self, n_epoch, model, result_dir, deterministic_synthesis, nmax=1000
237 def compute_accuracy(input):
239 ar_mask = self.make_ar_mask(input)
240 result = input.clone() * (1 - ar_mask)
241 seq_logproba = torch.empty(input.size(0), device=self.device)
243 masked_inplace_autoregression(
245 batch_size=self.batch_size,
248 seq_logproba=seq_logproba,
250 deterministic_synthesis=deterministic_synthesis,
251 progress_bar_desc=None,
255 nb_total, nb_correct = (
257 (input == result).long().min(dim=1).values.sum(),
260 return nb_total, nb_correct
262 train_nb_total, train_nb_correct = compute_accuracy(self.train_w_quizzes)
265 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}%"
268 test_nb_total, test_nb_correct = compute_accuracy(self.test_w_quizzes)
271 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}%"
274 main_test_accuracy = test_nb_correct / test_nb_total
275 self.logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
277 ##############################
279 input = self.test_w_quizzes[:96]
280 ar_mask = self.make_ar_mask(input)
281 result = input.clone() * (1 - ar_mask)
282 seq_logproba = torch.empty(input.size(0), device=self.device)
284 masked_inplace_autoregression(
286 batch_size=self.batch_size,
289 seq_logproba=seq_logproba,
291 deterministic_synthesis=deterministic_synthesis,
292 progress_bar_desc=None,
296 self.problem.save_quizzes(
297 result[:72], result_dir, f"culture_prediction_{n_epoch:04d}_{model.id:02d}"
300 return main_test_accuracy
302 def renew_w_quizzes(self, nb, for_train=True):
303 input = self.train_w_quizzes if for_train else self.test_w_quizzes
304 nb = min(nb, input.size(0))
305 input[:-nb] = input[nb:].clone()
306 input[-nb:] = self.problem.generate_token_sequences(nb).to(self.device)
308 def store_c_quizzes(self, new_c_quizzes, for_train=True):
310 self.train_c_quizzes.append(new_c_quizzes)
312 self.test_c_quizzes.append(new_c_quizzes)
314 def comput_correctness(self, c_quizzes, models_for_validation):
315 # Create the reverse quizzes
317 token_forward, token_backward = self.problem.direction_tokens()
319 l = (c_quizzes.size(1) - 1) // 2
320 direction = c_quizzes[:, l : l + 1]
321 direction = self.problem.token_forward * (
322 direction == self.problem.token_backward
323 ) + self.problem.token_backward * (direction == self.problem.token_forward)
324 reverse_c_quizzes = torch.cat(
325 [c_quizzes[:, l + 1 :], direction, c_quizzes[:, :l]], dim=1
328 ar_mask = self.make_ar_mask(c_quizzes)
329 seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
331 # Check how many of models can solve the quizzes in both directions
335 for model in models_for_validation:
336 result = c_quizzes.clone()
338 masked_inplace_autoregression(
340 batch_size=self.batch_size,
343 seq_logproba=seq_logproba,
345 deterministic_synthesis=True,
346 # progress_bar_desc="solving c_quizzes",
350 correct = (c_quizzes == result).long().min(dim=-1).values
352 reverse_result = reverse_c_quizzes.clone()
354 masked_inplace_autoregression(
356 batch_size=self.batch_size,
357 input=reverse_result,
359 seq_logproba=seq_logproba,
361 deterministic_synthesis=True,
362 # progress_bar_desc="solving reversed c_quizzes",
367 (reverse_c_quizzes == reverse_result).long().min(dim=-1).values
370 nb_correct += correct * reverse_correct
374 ###############################################################
376 def generate_quizzes(self, nb, model_for_generation, min_ave_seq_logproba):
377 c_quizzes = torch.empty(
378 nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
381 ar_mask_prompt = torch.zeros(c_quizzes.size(), device=self.device)
382 ar_mask_prompt[:, : ar_mask_prompt.size(1) // 2 + 1] = 1
383 ar_mask_solve = 1 - ar_mask_prompt
384 seq_logproba = torch.empty(ar_mask_prompt.size(0), device=self.device)
386 # bracketing of the temperature to get the target logproba if
387 # min_ave_seq_logproba is not None
390 d_temperature = 1 / 3
393 seq_logproba[...] = 0
395 masked_inplace_autoregression(
396 model=model_for_generation,
397 batch_size=self.batch_size,
399 ar_mask=ar_mask_prompt,
400 seq_logproba=seq_logproba,
401 temperature=temperature,
402 deterministic_synthesis=False,
403 # progress_bar_desc="sampling c_quizzes",
407 ave_seq_logproba = seq_logproba.mean()
409 masked_inplace_autoregression(
410 model=model_for_generation,
411 batch_size=self.batch_size,
413 ar_mask=ar_mask_solve,
414 seq_logproba=seq_logproba,
415 temperature=temperature,
416 deterministic_synthesis=True,
417 # progress_bar_desc="sampling c_quizzes",
421 # If we do not have target logprobs, get out now
422 if min_ave_seq_logproba is None:
426 if ave_seq_logproba < min_ave_seq_logproba:
427 if d_temperature > 0:
428 d_temperature *= -1 / 3
429 temperature += d_temperature
430 elif ave_seq_logproba > min_ave_seq_logproba * 0.99:
431 if d_temperature < 0:
432 d_temperature *= -1 / 3
433 temperature += d_temperature
437 self.logger(f"changing temperature to {temperature}")
439 return c_quizzes, seq_logproba.mean()
441 ######################################################################
443 def create_c_quizzes(
446 model_for_generation,
447 models_for_validation,
448 min_ave_seq_logproba,
452 c_quizzes, ave_seq_logproba = self.generate_quizzes(
453 nb, model_for_generation, min_ave_seq_logproba
456 nb_correct = self.comput_correctness(c_quizzes, models_for_validation)
458 return c_quizzes, nb_correct, ave_seq_logproba
460 ######################################################################
462 def gang_create_c_quizzes(
465 nb_models_for_generation,
468 min_ave_seq_logproba,
472 model_for_generation = Gang(models, nb_models_for_generation, mode)
473 models_for_validation = models
474 return self.create_c_quizzes(
476 model_for_generation,
477 models_for_validation,
478 min_ave_seq_logproba,