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 ######################################################################
73 def save_image(self, input, result_dir, filename, logger):
74 img = self.sky.seq2img(input.to("cpu"))
75 image_name = os.path.join(result_dir, filename)
76 torchvision.utils.save_image(img.float() / 255.0, image_name, nrow=6, padding=4)
77 logger(f"wrote {image_name}")
79 def save_quizzes(self, input, result_dir, filename_prefix, logger):
80 self.save_image(input, result_dir, filename_prefix + ".png", logger)
82 def make_ar_mask(self, input):
83 b = torch.arange(input.size(1), device=input.device) > input.size(1) // 2
84 return b.long()[None, :].expand_as(input)
93 device=torch.device("cpu"),
97 self.sky = sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2)
98 self.batch_size = batch_size
101 self.train_w_quizzes = self.sky.generate_seq(nb_train_samples).to(device)
102 self.test_w_quizzes = self.sky.generate_seq(nb_test_samples).to(device)
104 self.nb_codes = max(self.train_w_quizzes.max(), self.test_w_quizzes.max()) + 1
106 self.train_c_quizzes = []
107 self.test_c_quizzes = []
109 if result_dir is not None:
111 self.train_w_quizzes[:72], result_dir, f"culture_w_quizzes", logger
114 def batches(self, split="train", desc=None):
115 assert split in {"train", "test"}
117 w_quizzes = self.train_w_quizzes
118 c_quizzes = self.train_c_quizzes
120 w_quizzes = self.test_w_quizzes
121 c_quizzes = self.test_c_quizzes
123 if len(c_quizzes) > 0:
124 c_quizzes = torch.cat(c_quizzes, dim=0)
125 if c_quizzes.size(0) > w_quizzes.size(0) // 2:
126 i = torch.randperm(w_quizzes.size(0))[: w_quizzes.size(0) // 2]
127 c_quizzes = c_quizzes[i]
129 i = torch.randperm(w_quizzes.size(0))[
130 : w_quizzes.size(0) - c_quizzes.size(0)
132 w_quizzes = w_quizzes[i]
134 self.nb_batch_w_quizzes = w_quizzes.size(0)
135 self.nb_batch_c_quizzes = c_quizzes.size(0)
137 input = torch.cat([w_quizzes, c_quizzes], dim=0)
140 self.nb_batch_w_quizzes = w_quizzes.size(0)
141 self.nb_batch_c_quizzes = 0
144 input = input[torch.randperm(input.size(0))]
147 desc = f"epoch-{split}"
148 for batch in tqdm.tqdm(
149 input.split(self.batch_size), dynamic_ncols=True, desc=desc
153 def vocabulary_size(self):
157 self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
159 def compute_accuracy(input, logger=None):
161 ar_mask = self.make_ar_mask(input)
162 result = input.clone() * (1 - ar_mask)
163 seq_logproba = torch.empty(input.size(0), device=self.device)
165 masked_inplace_autoregression(
167 batch_size=self.batch_size,
170 seq_logproba=seq_logproba,
172 deterministic_synthesis=deterministic_synthesis,
173 progress_bar_desc=None,
177 nb_total, nb_correct = (
179 (input == result).long().min(dim=1).values.sum(),
182 return nb_total, nb_correct
184 train_nb_total, train_nb_correct = compute_accuracy(self.train_w_quizzes)
187 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}%"
190 test_nb_total, test_nb_correct = compute_accuracy(self.test_w_quizzes, logger)
193 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}%"
196 main_test_accuracy = test_nb_correct / test_nb_total
197 logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
199 ##############################
201 input = self.test_w_quizzes[:96]
202 ar_mask = self.make_ar_mask(input)
203 result = input.clone() * (1 - ar_mask)
204 seq_logproba = torch.empty(input.size(0), device=self.device)
206 masked_inplace_autoregression(
208 batch_size=self.batch_size,
211 seq_logproba=seq_logproba,
213 deterministic_synthesis=deterministic_synthesis,
214 progress_bar_desc=None,
221 f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
225 return main_test_accuracy
227 def renew_w_quizzes(self, nb, for_train=True):
228 input = self.train_w_quizzes if for_train else self.test_w_quizzes
229 nb = min(nb, input.size(0))
230 input[:-nb] = input[nb:].clone()
231 input[-nb:] = self.sky.generate_seq(nb).to(self.device)
233 def store_c_quizzes(self, new_c_quizzes, for_train=True):
235 self.train_c_quizzes.append(new_c_quizzes)
237 self.test_c_quizzes.append(new_c_quizzes)
239 def create_c_quizzes(
247 min_ave_seq_logproba,
249 ###############################################################
250 # Generate quizzes with model
252 c_quizzes = torch.empty(
253 nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
256 ar_mask = torch.full(c_quizzes.size(), 1, device=self.device)
257 seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
260 d_temperature = 1 / 3
263 seq_logproba[...] = 0
265 masked_inplace_autoregression(
267 batch_size=self.batch_size,
270 seq_logproba=seq_logproba,
271 temperature=temperature,
272 deterministic_synthesis=False,
273 progress_bar_desc="sampling c_quizzes",
277 ave_seq_logproba = seq_logproba.mean()
279 logger(f"{ave_seq_logproba=} {min_ave_seq_logproba=}")
281 if min_ave_seq_logproba is None:
285 if ave_seq_logproba < min_ave_seq_logproba * 1.1:
286 if d_temperature > 0:
287 d_temperature *= -1 / 3
288 temperature += d_temperature
289 elif ave_seq_logproba > min_ave_seq_logproba:
290 if d_temperature < 0:
291 d_temperature *= -1 / 3
292 temperature += d_temperature
296 logger(f"chaging temperature to {temperature}")
298 ###############################################################
299 # Create the reverse quizzes
301 l = (c_quizzes.size(1) - 1) // 2
302 direction = c_quizzes[:, l : l + 1]
303 direction = self.sky.token_forward * (
304 direction == self.sky.token_backward
305 ) + self.sky.token_backward * (direction == self.sky.token_forward)
306 reverse_c_quizzes = torch.cat(
307 [c_quizzes[:, l + 1 :], direction, c_quizzes[:, :l]], dim=1
310 ar_mask = self.make_ar_mask(c_quizzes)
311 seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
313 ###############################################################
314 # Check how many of the other models can solve them in both
319 for m in other_models:
320 result = c_quizzes.clone()
322 masked_inplace_autoregression(
324 batch_size=self.batch_size,
327 seq_logproba=seq_logproba,
329 deterministic_synthesis=True,
330 progress_bar_desc="solving c_quizzes",
334 correct = (c_quizzes == result).long().min(dim=-1).values
336 reverse_result = reverse_c_quizzes.clone()
338 masked_inplace_autoregression(
340 batch_size=self.batch_size,
341 input=reverse_result,
343 seq_logproba=seq_logproba,
345 deterministic_synthesis=True,
346 progress_bar_desc="solving reversed c_quizzes",
351 (reverse_c_quizzes == reverse_result).long().min(dim=-1).values
354 nb_correct.append((correct * reverse_correct)[None, :])
356 nb_correct = torch.cat(nb_correct, dim=0).sum(dim=0)
358 return c_quizzes, nb_correct, seq_logproba.mean()