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(
108 nb_train_samples, height=self.height, width=self.width
111 self.test_input = world.generate(
112 nb_test_samples, height=self.height, width=self.width
116 # for a in world.seq2str(self.train_input):
118 # for a in world.seq2str(self.test_input):
122 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
124 self.train_quizzes = []
125 self.test_quizzes = []
127 if result_dir is not None:
129 self.train_input[:72], result_dir, f"world_train.png", logger
132 def batches(self, split="train", desc=None):
133 assert split in {"train", "test"}
135 input = self.train_input
136 quizzes = self.train_quizzes
138 input = self.test_input
139 quizzes = self.test_quizzes
142 quizzes = torch.cat(quizzes, dim=0)
143 if quizzes.size(0) > input.size(0) // 2:
144 i = torch.randperm(input.size(0))[: input.size(0) // 2]
147 i = torch.randperm(input.size(0))[: input.size(0) - quizzes.size(0)]
150 self.nb_batch_samples_world = input.size(0)
151 self.nb_batch_samples_quizzes = quizzes.size(0)
153 input = torch.cat([input, quizzes], dim=0)
155 self.nb_batch_samples_world = input.size(0)
156 self.nb_batch_samples_quizzes = 0
159 desc = f"epoch-{split}"
160 for batch in tqdm.tqdm(
161 input.split(self.batch_size), dynamic_ncols=True, desc=desc
165 def vocabulary_size(self):
169 self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
171 def compute_accuracy(input, logger=None):
173 ar_mask = self.make_ar_mask(input)
174 result = input.clone() * (1 - ar_mask)
176 masked_inplace_autoregression(
181 deterministic_synthesis,
182 progress_bar_desc=None,
186 nb_total, nb_correct = (
188 (input == result).long().min(dim=1).values.sum(),
191 return nb_total, nb_correct
193 train_nb_total, train_nb_correct = compute_accuracy(self.train_input)
196 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}%"
199 test_nb_total, test_nb_correct = compute_accuracy(self.test_input, logger)
202 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}%"
205 main_test_accuracy = test_nb_correct / test_nb_total
206 logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
208 ##############################
210 input = self.test_input[:96]
211 ar_mask = self.make_ar_mask(input)
212 result = input.clone() * (1 - ar_mask)
214 masked_inplace_autoregression(
219 deterministic_synthesis,
220 progress_bar_desc=None,
227 f"world_prediction_{n_epoch:04d}_{model.id:02d}.png",
231 return main_test_accuracy
233 def store_new_quizzes(self, new_quizzes, for_train=True):
235 self.train_quizzes.append(new_quizzes)
237 self.test_quizzes.append(new_quizzes)
239 def create_new_quizzes(
248 ###############################################################
249 # Generate quizzes with model
251 quizzes = torch.empty(
252 nb, self.height * self.width * 2 + 1, device=self.device, dtype=torch.int64
254 ar_mask = torch.full(quizzes.size(), 1, device=self.device)
256 masked_inplace_autoregression(
261 deterministic_synthesis=False,
262 progress_bar_desc="creating quizzes",
266 ###############################################################
267 # Create the reverse quizzes
269 l = self.height * self.width
270 direction = quizzes[:, l : l + 1]
271 direction = world.token_forward * (
272 direction == world.token_backward
273 ) + world.token_backward * (direction == world.token_forward)
274 reverse_quizzes = torch.cat(
275 [quizzes[:, l + 1 :], direction, quizzes[:, :l]], dim=1
278 ar_mask = self.make_ar_mask(quizzes)
280 ###############################################################
281 # Check how many of the other models can solve them in both
286 for m in other_models:
287 result = quizzes.clone()
289 masked_inplace_autoregression(
294 deterministic_synthesis=True,
295 progress_bar_desc="solving quizzes",
299 correct = (quizzes == result).long().min(dim=-1).values
301 reverse_result = reverse_quizzes.clone()
303 masked_inplace_autoregression(
308 deterministic_synthesis=True,
309 progress_bar_desc="solving reversed quizzes",
314 (reverse_quizzes == reverse_result).long().min(dim=-1).values
317 nb_correct.append((correct * reverse_correct)[None, :])
319 nb_correct = torch.cat(nb_correct, dim=0)
321 filename = os.path.join(result_dir, "correct_{n_epoch:04d}.dat")
322 with open(filename, "w") as f:
326 return quizzes, nb_correct.sum(dim=0)