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 ######################################################################
21 class Gang(nn.Module):
22 def __init__(self, models, nb_models_for_generation, mode="groupthink"):
24 self.models = nn.ModuleList(models)
25 self.nb_models_for_generation = nb_models_for_generation
28 def forward(self, bs):
29 # If first = 0, we are re-starting an auto-regressive process,
30 # that's the right moment to randomize who gonna do it
32 self.models_to_use = [
34 for k in torch.randperm(len(self.models))[
35 : self.nb_models_for_generation
39 all_the_logits = torch.cat(
40 [model(bs).x[None] for model in self.models_to_use], dim=0
43 if self.mode == "groupthink":
44 y = all_the_logits.mean(dim=0)
45 elif self.mode == "groupwork":
46 m = torch.rand(all_the_logits.size(), device=all_the_logits.device)
47 m = (m.sort(dim=0).indices == 0).long()
48 y = (y * m).sum(dim=0)
50 raise ValueError(f"Invalid mode {self.mode}")
52 return BracketedSequence(y, bs.first, bs.nb)
55 ######################################################################
57 # ar_mask is a tensor with 0s and 1s, of same shape as input, with
58 # 1s where tokens should be generated. The others are kept
62 def one_batch_masked_inplace_autoregression(
68 deterministic_synthesis=False,
69 forbidden_tokens=None,
72 to_generate = (ar_mask.sum(0) > 0).nonzero()
74 if to_generate.min() > 0:
76 BracketedSequence(input, 0, to_generate.min())
77 ) # Needed to initialize the model's cache
78 for s in range(to_generate.min(), to_generate.max() + 1):
79 output = model(BracketedSequence(input, s, 1)).x
83 logits = (logits / temperature).log_softmax(dim=-1)
85 if forbidden_tokens is not None:
86 logits = logits.masked_fill(forbidden_tokens, float("-inf"))
88 if forced_biases is not None:
89 logits = logits + forced_biases[None, :]
91 if deterministic_synthesis:
92 t_next = logits.argmax(-1)
94 dist = torch.distributions.categorical.Categorical(logits=logits)
95 t_next = dist.sample()
97 all_n = torch.arange(t_next.size(0))
98 seq_logproba += logits[all_n, t_next].sum(dim=-1)
100 input[:, s] = ar_mask[:, s] * t_next + (1 - ar_mask[:, s]) * input[:, s]
103 def masked_inplace_autoregression(
110 deterministic_synthesis,
111 forbidden_tokens=None,
113 progress_bar_desc=None,
114 device=torch.device("cpu"),
116 assert input.size() == ar_mask.size()
119 input.split(batch_size),
120 ar_mask.split(batch_size),
121 seq_logproba.split(batch_size),
124 if progress_bar_desc is not None:
128 desc=progress_bar_desc,
129 total=(input.size(0) + batch_size - 1) // batch_size,
132 with torch.autograd.no_grad():
136 for input, ar_mask, seq_logproba in batches:
137 one_batch_masked_inplace_autoregression(
141 seq_logproba=seq_logproba,
142 temperature=temperature,
143 deterministic_synthesis=deterministic_synthesis,
144 forbidden_tokens=forbidden_tokens,
145 forced_biases=logit_biases,
151 ######################################################################
155 def make_ar_mask(self, input):
156 b = torch.arange(input.size(1), device=input.device) > input.size(1) // 2
157 return b.long()[None, :].expand_as(input)
167 device=torch.device("cpu"),
171 self.problem = problem
172 self.batch_size = batch_size
176 self.train_w_quizzes = self.problem.generate_token_sequences(
179 self.test_w_quizzes = self.problem.generate_token_sequences(nb_test_samples).to(
183 self.nb_codes = max(self.train_w_quizzes.max(), self.test_w_quizzes.max()) + 1
185 self.train_c_quizzes = []
186 self.test_c_quizzes = []
188 if result_dir is not None:
189 self.problem.save_quizzes(
190 self.train_w_quizzes[:72], result_dir, "culture_w_quizzes"
193 def batches(self, split="train", desc=None):
194 assert split in {"train", "test"}
196 w_quizzes = self.train_w_quizzes
197 c_quizzes = self.train_c_quizzes
199 w_quizzes = self.test_w_quizzes
200 c_quizzes = self.test_c_quizzes
202 if len(c_quizzes) > 0:
203 c_quizzes = torch.cat(c_quizzes, dim=0)
204 if c_quizzes.size(0) > w_quizzes.size(0) // 2:
205 i = torch.randperm(c_quizzes.size(0))[: w_quizzes.size(0) // 2]
206 c_quizzes = c_quizzes[i]
208 i = torch.randperm(w_quizzes.size(0))[
209 : w_quizzes.size(0) - c_quizzes.size(0)
211 w_quizzes = w_quizzes[i]
213 self.nb_batch_w_quizzes = w_quizzes.size(0)
214 self.nb_batch_c_quizzes = c_quizzes.size(0)
216 input = torch.cat([w_quizzes, c_quizzes], dim=0)
219 self.nb_batch_w_quizzes = w_quizzes.size(0)
220 self.nb_batch_c_quizzes = 0
223 input = input[torch.randperm(input.size(0))]
226 desc = f"epoch-{split}"
227 for batch in tqdm.tqdm(
228 input.split(self.batch_size), dynamic_ncols=True, desc=desc
232 def vocabulary_size(self):
236 self, n_epoch, model, result_dir, deterministic_synthesis, nmax=1000
238 def compute_accuracy(input):
240 ar_mask = self.make_ar_mask(input)
241 result = input.clone() * (1 - ar_mask)
242 seq_logproba = torch.empty(input.size(0), device=self.device)
244 masked_inplace_autoregression(
246 batch_size=self.batch_size,
249 seq_logproba=seq_logproba,
251 deterministic_synthesis=deterministic_synthesis,
252 progress_bar_desc=None,
256 nb_total, nb_correct = (
258 (input == result).long().min(dim=1).values.sum(),
261 return nb_total, nb_correct
263 train_nb_total, train_nb_correct = compute_accuracy(self.train_w_quizzes)
266 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}%"
269 test_nb_total, test_nb_correct = compute_accuracy(self.test_w_quizzes)
272 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}%"
275 main_test_accuracy = test_nb_correct / test_nb_total
276 self.logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
278 ##############################
280 input = self.test_w_quizzes[:96]
281 ar_mask = self.make_ar_mask(input)
282 result = input.clone() * (1 - ar_mask)
283 seq_logproba = torch.empty(input.size(0), device=self.device)
285 masked_inplace_autoregression(
287 batch_size=self.batch_size,
290 seq_logproba=seq_logproba,
292 deterministic_synthesis=deterministic_synthesis,
293 progress_bar_desc=None,
297 self.problem.save_quizzes(
298 result[:72], result_dir, f"culture_prediction_{n_epoch:04d}_{model.id:02d}"
301 return main_test_accuracy
303 def renew_w_quizzes(self, nb, for_train=True):
304 input = self.train_w_quizzes if for_train else self.test_w_quizzes
305 nb = min(nb, input.size(0))
306 input[:-nb] = input[nb:].clone()
307 input[-nb:] = self.problem.generate_token_sequences(nb).to(self.device)
309 def store_c_quizzes(self, new_c_quizzes, for_train=True):
311 self.train_c_quizzes.append(new_c_quizzes)
313 self.test_c_quizzes.append(new_c_quizzes)
315 def reverse_time(self, c_quizzes):
316 token_forward, token_backward = self.problem.direction_tokens()
318 l = (c_quizzes.size(1) - 1) // 2
319 direction = c_quizzes[:, l : l + 1]
320 direction = self.problem.token_forward * (
321 direction == self.problem.token_backward
322 ) + self.problem.token_backward * (direction == self.problem.token_forward)
324 return torch.cat([c_quizzes[:, l + 1 :], direction, c_quizzes[:, :l]], dim=1)
326 def comput_correctness(self, c_quizzes, models_for_validation):
327 reversed_c_quizzes = self.reverse_time(c_quizzes)
329 ar_mask = self.make_ar_mask(c_quizzes)
330 seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
332 # Check how many of models can solve the quizzes in both directions
336 for model in models_for_validation:
337 result = c_quizzes.clone()
339 masked_inplace_autoregression(
341 batch_size=self.batch_size,
344 seq_logproba=seq_logproba,
346 deterministic_synthesis=True,
347 # progress_bar_desc="solving c_quizzes",
351 correct = (c_quizzes == result).long().min(dim=-1).values
353 reversed_result = reversed_c_quizzes.clone()
355 masked_inplace_autoregression(
357 batch_size=self.batch_size,
358 input=reversed_result,
360 seq_logproba=seq_logproba,
362 deterministic_synthesis=True,
363 # progress_bar_desc="solving reversed c_quizzes",
368 (reversed_c_quizzes == reversed_result).long().min(dim=-1).values
371 nb_correct += correct * reversed_correct
375 ###############################################################
377 def generate_quizzes(
378 self, nb, model_for_generation, min_ave_seq_logproba, reverse_cleanup=False
380 c_quizzes = torch.empty(
381 nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
384 ar_mask_prompt = torch.zeros(c_quizzes.size(), device=self.device)
385 ar_mask_prompt[:, : ar_mask_prompt.size(1) // 2 + 1] = 1
386 ar_mask_solve = 1 - ar_mask_prompt
387 seq_logproba = torch.empty(ar_mask_prompt.size(0), device=self.device)
389 warnings.warn("very high temperature with reversed cleanup", RuntimeWarning)
392 # warnings.warn("noise injection", RuntimeWarning)
393 # noise_std = torch.rand(1).item()
394 # self.logger(f"{noise_std=}")
396 # mygpt.set_noise_injection(model_for_generation, noise_std)
398 masked_inplace_autoregression(
399 model=model_for_generation,
400 batch_size=self.batch_size,
402 ar_mask=ar_mask_prompt,
403 seq_logproba=seq_logproba,
404 temperature=temperature,
405 deterministic_synthesis=False,
406 # progress_bar_desc="sampling c_quizzes",
410 # mygpt.set_noise_injection(model_for_generation, 0.0)
412 ave_seq_logproba = seq_logproba.mean()
414 masked_inplace_autoregression(
415 model=model_for_generation,
416 batch_size=self.batch_size,
418 ar_mask=ar_mask_solve,
419 seq_logproba=seq_logproba,
420 temperature=temperature,
421 deterministic_synthesis=True,
422 # progress_bar_desc="sampling c_quizzes",
427 c_quizzes = self.reverse_time(c_quizzes)
428 masked_inplace_autoregression(
429 model=model_for_generation,
430 batch_size=self.batch_size,
432 ar_mask=ar_mask_solve,
433 seq_logproba=seq_logproba,
434 temperature=temperature,
435 deterministic_synthesis=True,
436 # progress_bar_desc="sampling c_quizzes",
440 return c_quizzes, seq_logproba.mean()
442 ######################################################################
444 def create_c_quizzes(
447 model_for_generation,
448 models_for_validation,
449 min_ave_seq_logproba,
454 c_quizzes, ave_seq_logproba = self.generate_quizzes(
456 model_for_generation=model_for_generation,
457 min_ave_seq_logproba=min_ave_seq_logproba,
458 reverse_cleanup=reverse_cleanup,
461 nb_correct = self.comput_correctness(c_quizzes, models_for_validation)
463 return c_quizzes, nb_correct, ave_seq_logproba
465 ######################################################################
467 def gang_create_c_quizzes(
470 nb_models_for_generation,
473 min_ave_seq_logproba,
477 model_for_generation = Gang(models, nb_models_for_generation, mode)
478 models_for_validation = models
479 return self.create_c_quizzes(
481 model_for_generation,
482 models_for_validation,
483 min_ave_seq_logproba,