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, sys, argparse, time, tqdm, os, datetime, warnings
10 import torch, torchvision
12 from torch.nn import functional as F
17 # world quizzes vs. culture quizzes
19 ######################################################################
21 accuracy_to_make_c_quizzes = 0.975
22 nb_new_c_quizzes_for_train = 1000
23 nb_new_c_quizzes_for_test = 100
25 ######################################################################
27 if torch.cuda.is_available():
28 device = torch.device("cuda")
29 torch.backends.cuda.matmul.allow_tf32 = True
31 device = torch.device("cpu")
33 ######################################################################
35 parser = argparse.ArgumentParser(
36 description="An implementation of GPT with cache.",
37 formatter_class=argparse.ArgumentDefaultsHelpFormatter,
40 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
42 parser.add_argument("--result_dir", type=str, default=None)
44 parser.add_argument("--seed", type=int, default=0)
46 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
48 ########################################
50 parser.add_argument("--nb_epochs", type=int, default=10000)
52 parser.add_argument("--batch_size", type=int, default=None)
54 parser.add_argument("--physical_batch_size", type=int, default=None)
56 parser.add_argument("--nb_train_samples", type=int, default=None)
58 parser.add_argument("--nb_test_samples", type=int, default=None)
60 parser.add_argument("--learning_rate", type=float, default=1e-4)
62 ########################################
64 parser.add_argument("--model", type=str, default=None)
66 parser.add_argument("--dim_model", type=int, default=None)
68 parser.add_argument("--dim_keys", type=int, default=None)
70 parser.add_argument("--dim_hidden", type=int, default=None)
72 parser.add_argument("--nb_heads", type=int, default=None)
74 parser.add_argument("--nb_blocks", type=int, default=None)
76 parser.add_argument("--dropout", type=float, default=0.1)
78 ########################################
80 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
82 parser.add_argument("--nb_gpts", type=int, default=5)
84 parser.add_argument("--dirty_debug", action="store_true", default=False)
86 ######################################################################
88 args = parser.parse_args()
90 if args.result_dir is None:
91 args.result_dir = f"results_culture"
93 ######################################################################
96 accuracy_to_make_c_quizzes = 0.0
97 nb_new_c_quizzes_for_train = 100
98 nb_new_c_quizzes_for_test = 10
100 ######################################################################
105 "nb_train_samples": 250000,
106 "nb_test_samples": 10000,
109 for k, v in default_args.items():
110 if getattr(args, k) is None:
113 ######################################################################
115 default_model_args = {
153 if args.model in default_model_args:
154 for k, v in default_model_args[args.model].items():
155 if getattr(args, k) is None:
158 raise ValueError(f"Unknown model {args.model}")
160 ######################################################################
163 os.mkdir(args.result_dir)
164 except FileExistsError:
165 print(f"result directory {args.result_dir} already exists")
168 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
171 # torch.backends.cudnn.deterministic = True
172 # torch.backends.cudnn.benchmark = False
173 # torch.use_deterministic_algorithms(True)
174 torch.manual_seed(args.seed)
175 if torch.cuda.is_available():
176 torch.cuda.manual_seed_all(args.seed)
178 ######################################################################
182 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
184 if log_file is not None:
185 log_file.write(t + s + "\n")
192 log_string(f"argv {' '.join(sys.argv)}")
195 log_string(f"args.{n} {getattr(args, n)}")
198 ######################################################################
201 args.nb_train_samples = 2500
202 args.nb_test_samples = 100
204 if args.physical_batch_size is None:
205 args.physical_batch_size = args.batch_size
207 assert args.batch_size % args.physical_batch_size == 0
209 assert args.nb_train_samples % args.batch_size == 0
210 assert args.nb_test_samples % args.batch_size == 0
213 nb_train_samples=args.nb_train_samples,
214 nb_test_samples=args.nb_test_samples,
215 batch_size=args.physical_batch_size,
216 result_dir=args.result_dir,
221 ######################################################################
223 log_string(f"device {device}")
225 vocabulary_size = task.vocabulary_size()
227 log_string(f"vocabulary_size {vocabulary_size}")
229 ######################################################################
231 # Compute the entropy of the training tokens
234 for input in task.batches(split="train", desc="train-entropy"):
235 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
236 token_probas = token_count / token_count.sum()
237 entropy = -torch.xlogy(token_probas, token_probas).sum()
238 train_set_perplexity = math.exp(entropy)
240 ######################################################################
241 # A bit of paranoia never hurts
243 if args.max_percents_of_test_in_train >= 0:
245 def subsets_as_tuples(batches, cs):
247 for batch in batches:
249 s.add(tuple([v.item() for v in x]))
255 nb_test, nb_in_train = 0, 0
256 for test_subset in subsets_as_tuples(
257 task.batches(split="test", desc="test-check"), 25000
260 for train_subset in subsets_as_tuples(
261 task.batches(split="train", desc="train-check"), 25000
263 in_train.update(test_subset.intersection(train_subset))
264 nb_in_train += len(in_train)
265 nb_test += len(test_subset)
268 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
272 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
273 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
275 ##############################
278 def one_epoch(model, task):
279 optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
283 nb_train_samples, acc_train_loss = 0, 0.0
285 for input in task.batches(split="train"):
286 input = input.to(device)
288 if nb_train_samples % args.batch_size == 0:
289 optimizer.zero_grad()
291 output = model(mygpt.BracketedSequence(input)).x
292 loss = F.cross_entropy(output.transpose(1, 2), input)
293 acc_train_loss += loss.item() * input.size(0)
295 nb_train_samples += input.size(0)
299 if nb_train_samples % args.batch_size == 0:
302 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
304 log_string(f"train_perplexity {n_epoch} {train_perplexity}")
307 ######################################################################
310 def run_tests(model, task, deterministic_synthesis):
311 with torch.autograd.no_grad():
314 nb_test_samples, acc_test_loss = 0, 0.0
315 nb_samples_accumulated = 0
317 for input in task.batches(split="test"):
318 input = input.to(device)
320 bs = model(mygpt.BracketedSequence(input))
323 loss = F.cross_entropy(output.transpose(1, 2), input)
325 acc_test_loss += loss.item() * input.size(0)
327 nb_test_samples += input.size(0)
329 main_test_accuracy = task.produce_results(
332 result_dir=args.result_dir,
334 deterministic_synthesis=deterministic_synthesis,
337 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
339 log_string(f"test_perplexity {n_epoch} {test_perplexity}")
341 model.main_test_accuracy = main_test_accuracy
344 ######################################################################
347 def create_c_quizzes(
353 desired_average_logits=None,
357 sum_logits, sum_nb_c_quizzes = 0, 0
359 while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
360 nb_to_generate = 4 * (nb_for_train + nb_for_test)
362 new_c_quizzes, nb_correct, average_logits = task.create_c_quizzes(
364 result_dir=args.result_dir,
368 other_models=other_models,
369 desired_average_logits=desired_average_logits,
372 sum_logits += new_c_quizzes.size(0) * average_logits
373 sum_nb_c_quizzes += new_c_quizzes.size(0)
375 to_keep = new_c_quizzes[nb_correct == len(other_models) - 1]
378 to_keep = new_c_quizzes
381 f"keep {to_keep.size(0)}/{new_c_quizzes.size(0)} c_quizzes ({to_keep.size(0)*100/new_c_quizzes.size(0):.02f}%)"
386 new_c_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
388 task.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
389 task.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
394 f"culture_c_quiz_{n_epoch:04d}_{model.id:02d}",
398 return sum_logits / sum_nb_c_quizzes
401 ######################################################################
405 for k in range(args.nb_gpts):
407 vocabulary_size=vocabulary_size,
408 dim_model=args.dim_model,
409 dim_keys=args.dim_keys,
410 dim_hidden=args.dim_hidden,
411 nb_heads=args.nb_heads,
412 nb_blocks=args.nb_blocks,
414 dropout=args.dropout,
417 model.main_test_accuracy = 0.0
423 nb_parameters = sum(p.numel() for p in models[0].parameters())
424 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
426 ######################################################################
428 desired_average_logits = None
430 for n_epoch in range(args.nb_epochs):
431 log_string(f"--- epoch {n_epoch} ----------------------------------------")
433 a = [(model.id, float(model.main_test_accuracy)) for model in models]
434 a.sort(key=lambda p: p[0])
435 log_string(f"current accuracies {a}")
437 # select the model with lowest accuracy
438 models.sort(key=lambda model: model.main_test_accuracy)
442 f"training model {model.id} main_test_accuracy {model.main_test_accuracy}"
446 one_epoch(model, task)
448 task.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
451 f"train_set_composition w_quizzes {task.nb_batch_w_quizzes} c_quizzes {task.nb_batch_c_quizzes}"
455 run_tests(model, task, deterministic_synthesis=False)
458 f"test_set_composition w_quizzes {task.nb_batch_w_quizzes} c_quizzes {task.nb_batch_c_quizzes}"
461 if min([m.main_test_accuracy for m in models]) >= accuracy_to_make_c_quizzes:
462 other_models = models.copy()
463 other_models.remove(model)
465 average_logits = create_c_quizzes(
469 nb_for_train=nb_new_c_quizzes_for_train,
470 nb_for_test=nb_new_c_quizzes_for_test,
471 desired_average_logits=desired_average_logits,
474 # We keep the first average logits as a reference
475 if desired_average_logits is None:
476 desired_average_logits = average_logits
479 f"desired_average_logits {desired_average_logits} average_logits {average_logits}"
484 run_tests(model, task, deterministic_synthesis=False)
487 ######################################################################