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 def masked_inplace_autoregression(
27 deterministic_synthesis,
28 forbidden_tokens=None,
30 progress_bar_desc="autoregression",
31 device=torch.device("cpu"),
33 assert input.size() == ar_mask.size()
36 input.split(batch_size),
37 ar_mask.split(batch_size),
38 seq_logproba.split(batch_size),
41 if progress_bar_desc is not None:
45 desc=progress_bar_desc,
46 total=(input.size(0) + batch_size - 1) // batch_size,
49 with torch.autograd.no_grad():
53 for input, ar_mask, seq_logproba in batches:
54 model.masked_inplace_autoregression(
57 seq_logproba=seq_logproba,
58 temperature=temperature,
59 deterministic_synthesis=deterministic_synthesis,
60 forbidden_tokens=forbidden_tokens,
61 forced_biases=logit_biases,
67 ######################################################################
71 def batches(self, split="train", nb_to_use=-1, desc=None):
74 def vocabulary_size(self):
78 self, n_epoch, model, result_dir, logger, deterministic_synthesis
83 ######################################################################
88 class QuizzMachine(Task):
89 def save_image(self, input, result_dir, filename, logger):
90 img = world.seq2img(input.to("cpu"), self.height, self.width)
91 image_name = os.path.join(result_dir, filename)
92 torchvision.utils.save_image(img.float() / 255.0, image_name, nrow=6, padding=4)
93 logger(f"wrote {image_name}")
95 def save_quizzes(self, input, result_dir, filename_prefix, logger):
96 self.save_image(input, result_dir, filename_prefix + ".png", logger)
98 def make_ar_mask(self, input):
99 b = torch.arange(input.size(1), device=input.device) > input.size(1) // 2
100 return b.long()[None, :].expand_as(input)
109 device=torch.device("cpu"),
113 self.batch_size = batch_size
118 self.train_w_quizzes = world.generate_seq(
119 nb_train_samples, height=self.height, width=self.width
122 self.test_w_quizzes = world.generate_seq(
123 nb_test_samples, height=self.height, width=self.width
126 self.nb_codes = max(self.train_w_quizzes.max(), self.test_w_quizzes.max()) + 1
128 self.train_c_quizzes = []
129 self.test_c_quizzes = []
131 if result_dir is not None:
133 self.train_w_quizzes[:72], result_dir, f"culture_w_quizzes", logger
136 def batches(self, split="train", desc=None):
137 assert split in {"train", "test"}
139 w_quizzes = self.train_w_quizzes
140 c_quizzes = self.train_c_quizzes
142 w_quizzes = self.test_w_quizzes
143 c_quizzes = self.test_c_quizzes
145 if len(c_quizzes) > 0:
146 c_quizzes = torch.cat(c_quizzes, dim=0)
147 if c_quizzes.size(0) > w_quizzes.size(0) // 2:
148 i = torch.randperm(w_quizzes.size(0))[: w_quizzes.size(0) // 2]
149 c_quizzes = c_quizzes[i]
151 i = torch.randperm(w_quizzes.size(0))[
152 : w_quizzes.size(0) - c_quizzes.size(0)
154 w_quizzes = w_quizzes[i]
156 self.nb_batch_w_quizzes = w_quizzes.size(0)
157 self.nb_batch_c_quizzes = c_quizzes.size(0)
159 input = torch.cat([w_quizzes, c_quizzes], dim=0)
162 self.nb_batch_w_quizzes = w_quizzes.size(0)
163 self.nb_batch_c_quizzes = 0
166 input = input[torch.randperm(input.size(0))]
169 desc = f"epoch-{split}"
170 for batch in tqdm.tqdm(
171 input.split(self.batch_size), dynamic_ncols=True, desc=desc
175 def vocabulary_size(self):
179 self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
181 def compute_accuracy(input, logger=None):
183 ar_mask = self.make_ar_mask(input)
184 result = input.clone() * (1 - ar_mask)
185 seq_logproba = torch.empty(input.size(0), device=self.device)
187 masked_inplace_autoregression(
189 batch_size=self.batch_size,
192 seq_logproba=seq_logproba,
194 deterministic_synthesis=deterministic_synthesis,
195 progress_bar_desc=None,
199 nb_total, nb_correct = (
201 (input == result).long().min(dim=1).values.sum(),
204 return nb_total, nb_correct
206 train_nb_total, train_nb_correct = compute_accuracy(self.train_w_quizzes)
209 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}%"
212 test_nb_total, test_nb_correct = compute_accuracy(self.test_w_quizzes, logger)
215 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}%"
218 main_test_accuracy = test_nb_correct / test_nb_total
219 logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
221 ##############################
223 input = self.test_w_quizzes[:96]
224 ar_mask = self.make_ar_mask(input)
225 result = input.clone() * (1 - ar_mask)
226 seq_logproba = torch.empty(input.size(0), device=self.device)
228 masked_inplace_autoregression(
230 batch_size=self.batch_size,
233 seq_logproba=seq_logproba,
235 deterministic_synthesis=deterministic_synthesis,
236 progress_bar_desc=None,
243 f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
247 return main_test_accuracy
249 def renew_w_quizzes(self, nb, for_train=True):
250 input = self.train_w_quizzes if for_train else self.test_w_quizzes
251 nb = min(nb, input.size(0))
252 input[:-nb] = input[nb:].clone()
253 input[-nb:] = world.generate_seq(nb, height=self.height, width=self.width).to(
257 def store_c_quizzes(self, new_c_quizzes, for_train=True):
259 self.train_c_quizzes.append(new_c_quizzes)
261 self.test_c_quizzes.append(new_c_quizzes)
263 def create_c_quizzes(
271 min_ave_seq_logproba,
273 ###############################################################
274 # Generate quizzes with model
276 c_quizzes = torch.empty(
277 nb, self.height * self.width * 2 + 1, device=self.device, dtype=torch.int64
280 ar_mask = torch.full(c_quizzes.size(), 1, device=self.device)
281 seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
287 seq_logproba[...] = 0
289 masked_inplace_autoregression(
291 batch_size=self.batch_size,
294 seq_logproba=seq_logproba,
295 temperature=temperature,
296 deterministic_synthesis=False,
297 progress_bar_desc="sampling c_quizzes",
301 ave_seq_logproba = seq_logproba.mean()
303 logger(f"{ave_seq_logproba=} {min_ave_seq_logproba=}")
305 if min_ave_seq_logproba is None:
309 if ave_seq_logproba < min_ave_seq_logproba * 1.1:
310 if d_temperature > 0:
311 d_temperature *= -1 / 3
312 temperature += d_temperature
313 elif ave_seq_logproba > min_ave_seq_logproba:
314 if d_temperature < 0:
315 d_temperature *= -1 / 3
316 temperature += d_temperature
320 logger(f"chaging temperature to {temperature}")
322 ###############################################################
323 # Create the reverse quizzes
325 l = self.height * self.width
326 direction = c_quizzes[:, l : l + 1]
327 direction = world.token_forward * (
328 direction == world.token_backward
329 ) + world.token_backward * (direction == world.token_forward)
330 reverse_c_quizzes = torch.cat(
331 [c_quizzes[:, l + 1 :], direction, c_quizzes[:, :l]], dim=1
334 ar_mask = self.make_ar_mask(c_quizzes)
335 seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
337 ###############################################################
338 # Check how many of the other models can solve them in both
343 for m in other_models:
344 result = c_quizzes.clone()
346 masked_inplace_autoregression(
348 batch_size=self.batch_size,
351 seq_logproba=seq_logproba,
353 deterministic_synthesis=True,
354 progress_bar_desc="solving c_quizzes",
358 correct = (c_quizzes == result).long().min(dim=-1).values
360 reverse_result = reverse_c_quizzes.clone()
362 masked_inplace_autoregression(
364 batch_size=self.batch_size,
365 input=reverse_result,
367 seq_logproba=seq_logproba,
369 deterministic_synthesis=True,
370 progress_bar_desc="solving reversed c_quizzes",
375 (reverse_c_quizzes == reverse_result).long().min(dim=-1).values
378 nb_correct.append((correct * reverse_correct)[None, :])
380 nb_correct = torch.cat(nb_correct, dim=0).sum(dim=0)
382 return c_quizzes, nb_correct, seq_logproba.mean()