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 make_ar_mask(self, input):
74 b = torch.arange(input.size(1), device=input.device) > input.size(1) // 2
75 return b.long()[None, :].expand_as(input)
84 device=torch.device("cpu"),
88 self.problem = sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2)
89 self.batch_size = batch_size
92 self.train_w_quizzes = self.problem.generate_seq(nb_train_samples).to(device)
93 self.test_w_quizzes = self.problem.generate_seq(nb_test_samples).to(device)
95 self.nb_codes = max(self.train_w_quizzes.max(), self.test_w_quizzes.max()) + 1
97 self.train_c_quizzes = []
98 self.test_c_quizzes = []
100 if result_dir is not None:
101 self.problem.save_quizzes(
102 self.train_w_quizzes[:72], result_dir, f"culture_w_quizzes", logger
105 def batches(self, split="train", desc=None):
106 assert split in {"train", "test"}
108 w_quizzes = self.train_w_quizzes
109 c_quizzes = self.train_c_quizzes
111 w_quizzes = self.test_w_quizzes
112 c_quizzes = self.test_c_quizzes
114 if len(c_quizzes) > 0:
115 c_quizzes = torch.cat(c_quizzes, dim=0)
116 if c_quizzes.size(0) > w_quizzes.size(0) // 2:
117 i = torch.randperm(w_quizzes.size(0))[: w_quizzes.size(0) // 2]
118 c_quizzes = c_quizzes[i]
120 i = torch.randperm(w_quizzes.size(0))[
121 : w_quizzes.size(0) - c_quizzes.size(0)
123 w_quizzes = w_quizzes[i]
125 self.nb_batch_w_quizzes = w_quizzes.size(0)
126 self.nb_batch_c_quizzes = c_quizzes.size(0)
128 input = torch.cat([w_quizzes, c_quizzes], dim=0)
131 self.nb_batch_w_quizzes = w_quizzes.size(0)
132 self.nb_batch_c_quizzes = 0
135 input = input[torch.randperm(input.size(0))]
138 desc = f"epoch-{split}"
139 for batch in tqdm.tqdm(
140 input.split(self.batch_size), dynamic_ncols=True, desc=desc
144 def vocabulary_size(self):
148 self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
150 def compute_accuracy(input, logger=None):
152 ar_mask = self.make_ar_mask(input)
153 result = input.clone() * (1 - ar_mask)
154 seq_logproba = torch.empty(input.size(0), device=self.device)
156 masked_inplace_autoregression(
158 batch_size=self.batch_size,
161 seq_logproba=seq_logproba,
163 deterministic_synthesis=deterministic_synthesis,
164 progress_bar_desc=None,
168 nb_total, nb_correct = (
170 (input == result).long().min(dim=1).values.sum(),
173 return nb_total, nb_correct
175 train_nb_total, train_nb_correct = compute_accuracy(self.train_w_quizzes)
178 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}%"
181 test_nb_total, test_nb_correct = compute_accuracy(self.test_w_quizzes, logger)
184 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}%"
187 main_test_accuracy = test_nb_correct / test_nb_total
188 logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
190 ##############################
192 input = self.test_w_quizzes[:96]
193 ar_mask = self.make_ar_mask(input)
194 result = input.clone() * (1 - ar_mask)
195 seq_logproba = torch.empty(input.size(0), device=self.device)
197 masked_inplace_autoregression(
199 batch_size=self.batch_size,
202 seq_logproba=seq_logproba,
204 deterministic_synthesis=deterministic_synthesis,
205 progress_bar_desc=None,
209 self.problem.save_quizzes(
212 f"culture_prediction_{n_epoch:04d}_{model.id:02d}",
216 return main_test_accuracy
218 def renew_w_quizzes(self, nb, for_train=True):
219 input = self.train_w_quizzes if for_train else self.test_w_quizzes
220 nb = min(nb, input.size(0))
221 input[:-nb] = input[nb:].clone()
222 input[-nb:] = self.problem.generate_seq(nb).to(self.device)
224 def store_c_quizzes(self, new_c_quizzes, for_train=True):
226 self.train_c_quizzes.append(new_c_quizzes)
228 self.test_c_quizzes.append(new_c_quizzes)
230 def create_c_quizzes(
238 min_ave_seq_logproba,
240 ###############################################################
241 # Generate quizzes with model
243 c_quizzes = torch.empty(
244 nb, self.train_w_quizzes.size(1), device=self.device, dtype=torch.int64
247 ar_mask = torch.full(c_quizzes.size(), 1, device=self.device)
248 seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
251 d_temperature = 1 / 3
254 seq_logproba[...] = 0
256 masked_inplace_autoregression(
258 batch_size=self.batch_size,
261 seq_logproba=seq_logproba,
262 temperature=temperature,
263 deterministic_synthesis=False,
264 progress_bar_desc="sampling c_quizzes",
268 ave_seq_logproba = seq_logproba.mean()
270 logger(f"{ave_seq_logproba=} {min_ave_seq_logproba=}")
272 if min_ave_seq_logproba is None:
276 if ave_seq_logproba < min_ave_seq_logproba * 1.1:
277 if d_temperature > 0:
278 d_temperature *= -1 / 3
279 temperature += d_temperature
280 elif ave_seq_logproba > min_ave_seq_logproba:
281 if d_temperature < 0:
282 d_temperature *= -1 / 3
283 temperature += d_temperature
287 logger(f"chaging temperature to {temperature}")
289 ###############################################################
290 # Create the reverse quizzes
292 token_forward, token_backward = self.problem.direction_tokens()
294 l = (c_quizzes.size(1) - 1) // 2
295 direction = c_quizzes[:, l : l + 1]
296 direction = self.problem.token_forward * (
297 direction == self.problem.token_backward
298 ) + self.problem.token_backward * (direction == self.problem.token_forward)
299 reverse_c_quizzes = torch.cat(
300 [c_quizzes[:, l + 1 :], direction, c_quizzes[:, :l]], dim=1
303 ar_mask = self.make_ar_mask(c_quizzes)
304 seq_logproba = torch.empty(ar_mask.size(0), device=self.device)
306 ###############################################################
307 # Check how many of the other models can solve them in both
312 for m in other_models:
313 result = c_quizzes.clone()
315 masked_inplace_autoregression(
317 batch_size=self.batch_size,
320 seq_logproba=seq_logproba,
322 deterministic_synthesis=True,
323 progress_bar_desc="solving c_quizzes",
327 correct = (c_quizzes == result).long().min(dim=-1).values
329 reverse_result = reverse_c_quizzes.clone()
331 masked_inplace_autoregression(
333 batch_size=self.batch_size,
334 input=reverse_result,
336 seq_logproba=seq_logproba,
338 deterministic_synthesis=True,
339 progress_bar_desc="solving reversed c_quizzes",
344 (reverse_c_quizzes == reverse_result).long().min(dim=-1).values
347 nb_correct.append((correct * reverse_correct)[None, :])
349 nb_correct = torch.cat(nb_correct, dim=0).sum(dim=0)
351 return c_quizzes, nb_correct, seq_logproba.mean()