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(
25 deterministic_synthesis,
26 forbidden_tokens=None,
28 progress_bar_desc="autoregression",
29 device=torch.device("cpu"),
31 assert input.size() == ar_mask.size()
33 batches = zip(input.split(batch_size), ar_mask.split(batch_size))
35 if progress_bar_desc is not None:
39 desc=progress_bar_desc,
40 total=(input.size(0) + batch_size - 1) // batch_size,
43 with torch.autograd.no_grad():
47 for input, ar_mask in batches:
48 model.masked_inplace_autoregression(
51 deterministic_synthesis,
59 ######################################################################
63 def batches(self, split="train", nb_to_use=-1, desc=None):
66 def vocabulary_size(self):
70 self, n_epoch, model, result_dir, logger, deterministic_synthesis
75 ######################################################################
81 def save_image(self, input, result_dir, filename, logger):
82 img = world.sample2img(input.to("cpu"), self.height, self.width)
83 image_name = os.path.join(result_dir, filename)
84 torchvision.utils.save_image(img.float() / 255.0, image_name, nrow=6, padding=4)
85 logger(f"wrote {image_name}")
87 def make_ar_mask(self, input):
88 b = torch.arange(input.size(1), device=input.device) > input.size(1) // 2
89 return b.long()[None, :].expand_as(input)
98 device=torch.device("cpu"),
102 self.batch_size = batch_size
107 self.train_input = world.generate_seq(
108 nb_train_samples, height=self.height, width=self.width
111 self.test_input = world.generate_seq(
112 nb_test_samples, height=self.height, width=self.width
115 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
117 self.train_quizzes = []
118 self.test_quizzes = []
120 if result_dir is not None:
122 self.train_input[:72], result_dir, f"world_train.png", logger
125 def batches(self, split="train", desc=None):
126 assert split in {"train", "test"}
128 input = self.train_input
129 quizzes = self.train_quizzes
131 input = self.test_input
132 quizzes = self.test_quizzes
135 quizzes = torch.cat(quizzes, dim=0)
136 if quizzes.size(0) > input.size(0) // 2:
137 i = torch.randperm(input.size(0))[: input.size(0) // 2]
140 i = torch.randperm(input.size(0))[: input.size(0) - quizzes.size(0)]
143 self.nb_batch_samples_world = input.size(0)
144 self.nb_batch_samples_quizzes = quizzes.size(0)
146 input = torch.cat([input, quizzes], dim=0)
148 self.nb_batch_samples_world = input.size(0)
149 self.nb_batch_samples_quizzes = 0
152 desc = f"epoch-{split}"
153 for batch in tqdm.tqdm(
154 input.split(self.batch_size), dynamic_ncols=True, desc=desc
158 def vocabulary_size(self):
162 self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
164 def compute_accuracy(input, logger=None):
166 ar_mask = self.make_ar_mask(input)
167 result = input.clone() * (1 - ar_mask)
169 masked_inplace_autoregression(
174 deterministic_synthesis,
175 progress_bar_desc=None,
179 nb_total, nb_correct = (
181 (input == result).long().min(dim=1).values.sum(),
184 return nb_total, nb_correct
186 train_nb_total, train_nb_correct = compute_accuracy(self.train_input)
189 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}%"
192 test_nb_total, test_nb_correct = compute_accuracy(self.test_input, logger)
195 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}%"
198 main_test_accuracy = test_nb_correct / test_nb_total
199 logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
201 ##############################
203 input = self.test_input[:96]
204 ar_mask = self.make_ar_mask(input)
205 result = input.clone() * (1 - ar_mask)
207 masked_inplace_autoregression(
212 deterministic_synthesis,
213 progress_bar_desc=None,
220 f"world_prediction_{n_epoch:04d}_{model.id:02d}.png",
224 return main_test_accuracy
226 def store_new_quizzes(self, new_quizzes, for_train=True):
228 self.train_quizzes.append(new_quizzes)
230 self.test_quizzes.append(new_quizzes)
232 def create_new_quizzes(
241 ###############################################################
242 # Generate quizzes with model
244 quizzes = torch.empty(
245 nb, self.height * self.width * 2 + 1, device=self.device, dtype=torch.int64
247 ar_mask = torch.full(quizzes.size(), 1, device=self.device)
249 masked_inplace_autoregression(
254 deterministic_synthesis=False,
255 progress_bar_desc="creating quizzes",
259 ###############################################################
260 # Create the reverse quizzes
262 l = self.height * self.width
263 direction = quizzes[:, l : l + 1]
264 direction = world.token_forward * (
265 direction == world.token_backward
266 ) + world.token_backward * (direction == world.token_forward)
267 reverse_quizzes = torch.cat(
268 [quizzes[:, l + 1 :], direction, quizzes[:, :l]], dim=1
271 ar_mask = self.make_ar_mask(quizzes)
273 ###############################################################
274 # Check how many of the other models can solve them in both
279 for m in other_models:
280 result = quizzes.clone()
282 masked_inplace_autoregression(
287 deterministic_synthesis=True,
288 progress_bar_desc="solving quizzes",
292 correct = (quizzes == result).long().min(dim=-1).values
294 reverse_result = reverse_quizzes.clone()
296 masked_inplace_autoregression(
301 deterministic_synthesis=True,
302 progress_bar_desc="solving reversed quizzes",
307 (reverse_quizzes == reverse_result).long().min(dim=-1).values
310 nb_correct.append((correct * reverse_correct)[None, :])
312 nb_correct = torch.cat(nb_correct, dim=0)
314 filename = os.path.join(result_dir, "correct_{n_epoch:04d}.dat")
315 with open(filename, "w") as f:
319 return quizzes, nb_correct.sum(dim=0)