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(
26 deterministic_synthesis,
27 forbidden_tokens=None,
29 progress_bar_desc="autoregression",
30 device=torch.device("cpu"),
32 assert input.size() == ar_mask.size()
34 batches = zip(input.split(batch_size), ar_mask.split(batch_size))
36 if progress_bar_desc is not None:
40 desc=progress_bar_desc,
41 total=(input.size(0) + batch_size - 1) // batch_size,
46 with torch.autograd.no_grad():
50 for input, ar_mask in batches:
51 sum_logits += model.masked_inplace_autoregression(
54 temperature=temperature,
55 deterministic_synthesis=deterministic_synthesis,
56 forbidden_tokens=forbidden_tokens,
57 forced_biases=logit_biases,
65 ######################################################################
69 def batches(self, split="train", nb_to_use=-1, desc=None):
72 def vocabulary_size(self):
76 self, n_epoch, model, result_dir, logger, deterministic_synthesis
81 ######################################################################
87 def save_image(self, input, result_dir, filename, logger):
88 img = world.seq2img(input.to("cpu"), self.height, self.width)
89 image_name = os.path.join(result_dir, filename)
90 torchvision.utils.save_image(img.float() / 255.0, image_name, nrow=6, padding=4)
91 logger(f"wrote {image_name}")
93 def make_ar_mask(self, input):
94 b = torch.arange(input.size(1), device=input.device) > input.size(1) // 2
95 return b.long()[None, :].expand_as(input)
104 device=torch.device("cpu"),
108 self.batch_size = batch_size
113 self.train_input = world.generate_seq(
114 nb_train_samples, height=self.height, width=self.width
117 self.test_input = world.generate_seq(
118 nb_test_samples, height=self.height, width=self.width
121 self.nb_codes = max(self.train_input.max(), self.test_input.max()) + 1
123 self.train_quizzes = []
124 self.test_quizzes = []
126 if result_dir is not None:
128 self.train_input[:72], result_dir, f"world_train.png", logger
131 def batches(self, split="train", desc=None):
132 assert split in {"train", "test"}
134 input = self.train_input
135 quizzes = self.train_quizzes
137 input = self.test_input
138 quizzes = self.test_quizzes
141 quizzes = torch.cat(quizzes, dim=0)
142 if quizzes.size(0) > input.size(0) // 2:
143 i = torch.randperm(input.size(0))[: input.size(0) // 2]
146 i = torch.randperm(input.size(0))[: input.size(0) - quizzes.size(0)]
149 self.nb_batch_samples_world = input.size(0)
150 self.nb_batch_samples_quizzes = quizzes.size(0)
152 input = torch.cat([input, quizzes], dim=0)
154 self.nb_batch_samples_world = input.size(0)
155 self.nb_batch_samples_quizzes = 0
158 input = input[torch.randperm(input.size(0))]
161 desc = f"epoch-{split}"
162 for batch in tqdm.tqdm(
163 input.split(self.batch_size), dynamic_ncols=True, desc=desc
167 def vocabulary_size(self):
171 self, n_epoch, model, result_dir, logger, deterministic_synthesis, nmax=1000
173 def compute_accuracy(input, logger=None):
175 ar_mask = self.make_ar_mask(input)
176 result = input.clone() * (1 - ar_mask)
178 masked_inplace_autoregression(
180 batch_size=self.batch_size,
184 deterministic_synthesis=deterministic_synthesis,
185 progress_bar_desc=None,
189 nb_total, nb_correct = (
191 (input == result).long().min(dim=1).values.sum(),
194 return nb_total, nb_correct
196 train_nb_total, train_nb_correct = compute_accuracy(self.train_input)
199 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}%"
202 test_nb_total, test_nb_correct = compute_accuracy(self.test_input, logger)
205 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}%"
208 main_test_accuracy = test_nb_correct / test_nb_total
209 logger(f"main_test_accuracy {n_epoch} {main_test_accuracy}")
211 ##############################
213 input = self.test_input[:96]
214 ar_mask = self.make_ar_mask(input)
215 result = input.clone() * (1 - ar_mask)
217 masked_inplace_autoregression(
219 batch_size=self.batch_size,
223 deterministic_synthesis=deterministic_synthesis,
224 progress_bar_desc=None,
231 f"world_prediction_{n_epoch:04d}_{model.id:02d}.png",
235 return main_test_accuracy
237 def renew_samples(self, nb, for_train=True):
238 input = self.train_input if for_train else self.test_input
239 nb = min(nb, input.size(0))
240 input[:-nb] = input[nb:].clone()
241 input[-nb:] = world.generate_seq(nb, height=self.height, width=self.width).to(
245 def store_new_quizzes(self, new_quizzes, for_train=True):
247 self.train_quizzes.append(new_quizzes)
249 self.test_quizzes.append(new_quizzes)
251 def create_new_quizzes(
259 desired_average_logits=None,
261 ###############################################################
262 # Generate quizzes with model
264 quizzes = torch.empty(
265 nb, self.height * self.width * 2 + 1, device=self.device, dtype=torch.int64
268 ar_mask = torch.full(quizzes.size(), 1, device=self.device)
274 sum_logits = masked_inplace_autoregression(
276 batch_size=self.batch_size,
279 temperature=temperature,
280 deterministic_synthesis=False,
281 progress_bar_desc="creating quizzes",
285 average_logits = sum_logits / quizzes.size(0)
287 logger(f"{average_logits=} {desired_average_logits=}")
289 if desired_average_logits is None:
293 if average_logits > desired_average_logits:
294 if d_temperature < 0:
295 d_temperature *= -0.5
296 temperature += d_temperature
298 if d_temperature > 0:
299 d_temperature *= -0.5
300 temperature += d_temperature
301 logger(f"chaging temperature to {temperature}")
303 ###############################################################
304 # Create the reverse quizzes
306 l = self.height * self.width
307 direction = quizzes[:, l : l + 1]
308 direction = world.token_forward * (
309 direction == world.token_backward
310 ) + world.token_backward * (direction == world.token_forward)
311 reverse_quizzes = torch.cat(
312 [quizzes[:, l + 1 :], direction, quizzes[:, :l]], dim=1
315 ar_mask = self.make_ar_mask(quizzes)
317 ###############################################################
318 # Check how many of the other models can solve them in both
323 for m in other_models:
324 result = quizzes.clone()
326 masked_inplace_autoregression(
328 batch_size=self.batch_size,
332 deterministic_synthesis=True,
333 progress_bar_desc="solving quizzes",
337 correct = (quizzes == result).long().min(dim=-1).values
339 reverse_result = reverse_quizzes.clone()
341 masked_inplace_autoregression(
343 batch_size=self.batch_size,
344 input=reverse_result,
347 deterministic_synthesis=True,
348 progress_bar_desc="solving reversed quizzes",
353 (reverse_quizzes == reverse_result).long().min(dim=-1).values
356 nb_correct.append((correct * reverse_correct)[None, :])
358 nb_correct = torch.cat(nb_correct, dim=0)
360 # filename = os.path.join(result_dir, "correct_{n_epoch:04d}.dat")
361 # with open(filename, "w") as f:
362 # for k in nb_correct:
365 return quizzes, nb_correct.sum(dim=0), sum_logits