X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=9c3d7f1ac607d2d47aa3702691c1fdc11e9300c5;hb=9ec709a2a08eb82dfc17ef1e24aa9a84751d63e0;hp=55379652a8766a72fa8b20035d4176702f791251;hpb=525bd24014786b53638dea78cfb88035a2b99d97;p=culture.git diff --git a/main.py b/main.py index 5537965..9c3d7f1 100755 --- a/main.py +++ b/main.py @@ -12,18 +12,15 @@ from torch import nn from torch.nn import functional as F import ffutils + import mygpt -import sky, wireworld, quizz_machine +import sky, grids, quiz_machine +from problem import MultiThreadProblem # world quizzes vs. culture quizzes ###################################################################### -nb_new_c_quizzes_for_train = 1000 -nb_new_c_quizzes_for_test = 100 - -###################################################################### - if torch.cuda.is_available(): device = torch.device("cuda") torch.backends.cuda.matmul.allow_tf32 = True @@ -57,7 +54,7 @@ parser.add_argument("--nb_train_samples", type=int, default=None) parser.add_argument("--nb_test_samples", type=int, default=None) -parser.add_argument("--learning_rate", type=float, default=1e-3) +parser.add_argument("--learning_rate", type=float, default=5e-4) ######################################## @@ -79,35 +76,50 @@ parser.add_argument("--dropout", type=float, default=0.1) parser.add_argument("--deterministic_synthesis", action="store_true", default=False) -parser.add_argument("--problem", type=str, default="sky") +parser.add_argument("--problem", type=str, default="grids") + +parser.add_argument("--multi_thread_problem", action="store_true", default=False) parser.add_argument("--nb_gpts", type=int, default=5) -parser.add_argument("--nb_models_for_generation", type=int, default=1) +parser.add_argument("--min_to_validate", type=int, default=None) -parser.add_argument("--generation_mode", type=str, default="groupthink") +parser.add_argument("--max_to_validate", type=int, default=None) -parser.add_argument("--min_to_validate", type=int, default=4) +parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975) -parser.add_argument("--max_to_validate", type=int, default=4) +parser.add_argument("--generation_temperature", type=float, default=2.0) -parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975) +parser.add_argument("--deterministic_validation", action="store_true", default=False) + +parser.add_argument("--bidirectional_validation", action="store_true", default=False) parser.add_argument("--dirty_debug", action="store_true", default=False) ###################################################################### -args = parser.parse_args() +parser.add_argument("--sky_height", type=int, default=6) -if args.result_dir is None: - args.result_dir = f"results_culture" +parser.add_argument("--sky_width", type=int, default=8) + +parser.add_argument("--sky_nb_birds", type=int, default=3) + +parser.add_argument("--sky_nb_iterations", type=int, default=2) + +parser.add_argument("--sky_speed", type=int, default=3) ###################################################################### -if args.dirty_debug: - args.accuracy_to_make_c_quizzes = 0.0 - nb_new_c_quizzes_for_train = 100 - nb_new_c_quizzes_for_test = 10 +args = parser.parse_args() + +if args.min_to_validate is None: + args.min_to_validate = args.nb_gpts - 1 + +if args.max_to_validate is None: + args.max_to_validate = args.nb_gpts - 1 + +if args.result_dir is None: + args.result_dir = f"results_culture" ###################################################################### @@ -222,16 +234,28 @@ assert args.nb_train_samples % args.batch_size == 0 assert args.nb_test_samples % args.batch_size == 0 if args.problem == "sky": - problem = (sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2, speed=2),) -elif args.problem == "wireworld": - problem = wireworld.Wireworld(height=10, width=15, nb_iterations=4) + problem = sky.Sky( + height=args.sky_height, + width=args.sky_width, + nb_birds=args.sky_nb_birds, + nb_iterations=args.sky_nb_iterations, + speed=args.sky_speed, + ) + back_accuracy = False +elif args.problem == "grids": + problem = grids.Grids(device=device) + back_accuracy = True else: raise ValueError -quizz_machine = quizz_machine.QuizzMachine( +if args.multi_thread_problem: + problem = MultiThreadProblem(problem, args.nb_train_samples, chunk_size=1000) + +quiz_machine = quiz_machine.QuizMachine( problem=problem, nb_train_samples=args.nb_train_samples, nb_test_samples=args.nb_test_samples, + back_accuracy=back_accuracy, batch_size=args.physical_batch_size, result_dir=args.result_dir, logger=log_string, @@ -242,7 +266,7 @@ quizz_machine = quizz_machine.QuizzMachine( log_string(f"device {device}") -vocabulary_size = quizz_machine.vocabulary_size() +vocabulary_size = quiz_machine.vocabulary_size() log_string(f"vocabulary_size {vocabulary_size}") @@ -251,8 +275,8 @@ log_string(f"vocabulary_size {vocabulary_size}") # Compute the entropy of the training tokens token_count = 0 -for input in quizz_machine.batches(split="train", desc="train-entropy"): - token_count += F.one_hot(input, num_classes=quizz_machine.vocabulary_size()).sum( +for input in quiz_machine.batches(split="train", desc="train-entropy"): + token_count += F.one_hot(input, num_classes=quiz_machine.vocabulary_size()).sum( (0, 1) ) token_probas = token_count / token_count.sum() @@ -276,11 +300,11 @@ if args.max_percents_of_test_in_train >= 0: nb_test, nb_in_train = 0, 0 for test_subset in subsets_as_tuples( - quizz_machine.batches(split="test", desc="test-check"), 25000 + quiz_machine.batches(split="test", desc="test-check"), 25000 ): in_train = set() for train_subset in subsets_as_tuples( - quizz_machine.batches(split="train", desc="train-check"), 25000 + quiz_machine.batches(split="train", desc="train-check"), 25000 ): in_train.update(test_subset.intersection(train_subset)) nb_in_train += len(in_train) @@ -297,14 +321,14 @@ if args.max_percents_of_test_in_train >= 0: ############################## -def one_epoch(model, quizz_machine): +def one_epoch(model, quiz_machine): optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate) model.train() nb_train_samples, acc_train_loss = 0, 0.0 - for input in quizz_machine.batches(split="train"): + for input in quiz_machine.batches(split="train"): input = input.to(device) if nb_train_samples % args.batch_size == 0: @@ -329,14 +353,14 @@ def one_epoch(model, quizz_machine): ###################################################################### -def run_tests(model, quizz_machine, deterministic_synthesis): +def run_tests(model, quiz_machine, deterministic_synthesis): with torch.autograd.no_grad(): model.eval() nb_test_samples, acc_test_loss = 0, 0.0 nb_samples_accumulated = 0 - for input in quizz_machine.batches(split="test"): + for input in quiz_machine.batches(split="test"): input = input.to(device) bs = model(mygpt.BracketedSequence(input)) @@ -348,18 +372,24 @@ def run_tests(model, quizz_machine, deterministic_synthesis): nb_test_samples += input.size(0) - main_test_accuracy = quizz_machine.produce_results( + test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples)) + + log_string(f"test_perplexity {n_epoch} {test_perplexity}") + + model.main_test_accuracy = quiz_machine.produce_results( n_epoch=n_epoch, model=model, result_dir=args.result_dir, deterministic_synthesis=deterministic_synthesis, ) - test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples)) - log_string(f"test_perplexity {n_epoch} {test_perplexity}") +###################################################################### + - model.main_test_accuracy = main_test_accuracy +def valid_c_quizzes(recorded, criteria): + result = [q[criteria(c)] for q, c in recorded] + return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([]) ###################################################################### @@ -367,97 +397,98 @@ def run_tests(model, quizz_machine, deterministic_synthesis): def create_c_quizzes( models, - quizz_machine, + quiz_machine, nb_for_train=1000, nb_for_test=100, - min_ave_seq_logproba=None, ): - # We will store the generated quizzes for each number of - # correct prediction - recorded = dict([(n, []) for n in range(len(models) + 1)]) - - model_indexes = [] - sum_logits, sum_nb_c_quizzes = 0, 0 - - def nb_generated(): - return sum([sum([x.size(0) for x in recorded[n]]) for n in recorded.keys()]) - - def nb_validated(): - return sum( - [ - sum([x.size(0) for x in recorded[n]]) - for n in range(args.min_to_validate, args.max_to_validate + 1) - ] - ) + quizzes_and_nb_correct_records = [] nb_to_create = nb_for_train + nb_for_test - while nb_validated() < nb_to_create: - ( - new_c_quizzes, - nb_correct, - ave_seq_logproba, - ) = quizz_machine.gang_create_c_quizzes( - nb=nb_to_create, - nb_models_for_generation=args.nb_models_for_generation, - models=models, - mode=args.generation_mode, - min_ave_seq_logproba=min_ave_seq_logproba, - n_epoch=n_epoch, - result_dir=args.result_dir, - ) + # ------------------------------------------------------------ - sum_logits += new_c_quizzes.size(0) * ave_seq_logproba - sum_nb_c_quizzes += new_c_quizzes.size(0) + standard_validity = lambda nb_correct: torch.logical_and( + nb_correct >= args.min_to_validate, nb_correct <= args.max_to_validate + ) + + file_name = os.path.join(args.result_dir, f"culture_c_quiz_{n_epoch:04d}_logp.dat") + + with open(file_name, "w") as logp_file: + while ( + valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity).size(0) + < nb_to_create + ): + # Select a model at random to generate the new quizzes + + model_for_generation = models[torch.randint(len(models), (1,))] - if args.dirty_debug: - nb_correct = torch.randint( - len(models) + 1, nb_correct.size(), device=new_c_quizzes.device + c_quizzes = quiz_machine.generate_quizzes( + nb_to_create, + model_for_generation=model_for_generation, + temperature=args.generation_temperature, ) - for n in range(nb_correct.max() + 1): - recorded[n].append(new_c_quizzes[nb_correct == n].clone()) + c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)] - log_string( - f"keep c_quizzes {nb_validated()*100/nb_generated():.02f}% kept total {nb_validated()} / {nb_to_create}" - ) + if c_quizzes.size(0) > 0: + nb_correct, seq_logproba = quiz_machine.compute_correctness( + c_quizzes, + models, + bidirectional_validation=args.bidirectional_validation, + deterministic_validation=args.deterministic_validation, + ) - # concatenate and shuffle - for n in recorded.keys(): - if len(recorded[n]) > 0: - q = torch.cat(recorded[n], dim=0) - q = q[torch.randperm(q.size(0), device=q.device)] - recorded[n] = q - else: - del recorded[n] - - new_c_quizzes = torch.cat( - [recorded[n] for n in range(args.min_to_validate, args.max_to_validate + 1)], - dim=0, - ) + for n, l in zip(nb_correct, seq_logproba): + s = " ".join([str(x.item()) for x in l]) + logp_file.write(f"{n} {s}\n") + + if args.dirty_debug: + nb_correct = torch.randint( + len(models) + 1, nb_correct.size(), device=c_quizzes.device + ) - new_c_quizzes = new_c_quizzes[ - torch.randperm(new_c_quizzes.size(0), device=new_c_quizzes.device)[ - : nb_for_train + nb_for_test - ] - ] + quizzes_and_nb_correct_records.append((c_quizzes, nb_correct)) - quizz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True) - quizz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False) + nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0) + nv = " ".join([str(x.item()) for x in nv]) - for n in recorded.keys(): + nb_validated = valid_c_quizzes( + quizzes_and_nb_correct_records, standard_validity + ).size(0) + + log_string( + f"keep c_quizzes model {model_for_generation.id} kept {nv} nb_accumulated {nb_validated} / {nb_to_create}" + ) + + # store the new c_quizzes which have been validated + + new_c_quizzes = valid_c_quizzes(quizzes_and_nb_correct_records, standard_validity) + + quiz_machine.reverse_random_half_in_place(new_c_quizzes) + + quiz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True) + quiz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False) + + # save a bunch of images to investigate what quizzes with a + # certain nb of correct predictions look like + + for n in range(len(models) + 1): s = ( "_validated" if n >= args.min_to_validate and n <= args.max_to_validate else "" ) - quizz_machine.problem.save_quizzes( - recorded[n][:72], - args.result_dir, - f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", - ) - return sum_logits / sum_nb_c_quizzes + q = valid_c_quizzes( + quizzes_and_nb_correct_records, criteria=lambda nb_correct: nb_correct == n + )[:72] + + quiz_machine.reverse_random_half_in_place(q) + + if q.size(0) > 0: + quiz_machine.save_quizzes( + args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q + ) ###################################################################### @@ -487,60 +518,69 @@ log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)") ###################################################################### -min_ave_seq_logproba = None +nb_new_c_quizzes_for_train = args.nb_train_samples // 50 +nb_new_c_quizzes_for_test = args.nb_test_samples // 50 + +log_string( + f"nb_new_c_quizzes_for_train {nb_new_c_quizzes_for_train} nb_new_c_quizzes_for_test {nb_new_c_quizzes_for_test}" +) + +###################################################################### + +if args.dirty_debug: + args.accuracy_to_make_c_quizzes = 0.0 + args.nb_gpts = 2 + nb_new_c_quizzes_for_train = 100 + nb_new_c_quizzes_for_test = 10 + +###################################################################### for n_epoch in range(args.nb_epochs): log_string(f"--- epoch {n_epoch} ----------------------------------------") - a = [(model.id, float(model.main_test_accuracy)) for model in models] - a.sort(key=lambda p: p[0]) - s = " ".join([f"{p[1]*100:.02f}%" for p in a]) - log_string(f"current accuracies {s}") + cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models]) + log_string(f"current_test_accuracies {cta}") - # select the model with lowest accuracy - models.sort(key=lambda model: model.main_test_accuracy) - model = models[0] + ################################################## + # Select, improve, and eval the worst model + + weakest_model = min(models, key=lambda m: float(m.main_test_accuracy)) log_string( - f"training model {model.id} main_test_accuracy {model.main_test_accuracy}" + f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}" ) - # improve it - one_epoch(model, quizz_machine) - - quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts) + one_epoch(weakest_model, quiz_machine) log_string( - f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}" + f"train_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}" ) - # test it - run_tests(model, quizz_machine, deterministic_synthesis=False) + run_tests(weakest_model, quiz_machine, deterministic_synthesis=False) log_string( - f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}" + f"test_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}" ) + ################################################## + # Replace a fraction of the w_quizzes with fresh ones + + quiz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts) + + ################################################## + # If all the models are good enough, generate new quizzes and + # re-compute the test errors + if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes: - ave_seq_logproba = create_c_quizzes( + create_c_quizzes( models, - quizz_machine, + quiz_machine, nb_for_train=nb_new_c_quizzes_for_train, nb_for_test=nb_new_c_quizzes_for_test, - min_ave_seq_logproba=min_ave_seq_logproba, ) - # We keep the first average logits as a reference - # if min_ave_seq_logproba is None: - # min_ave_seq_logproba = ave_seq_logproba - # else: - # log_string( - # f"min_ave_seq_logproba {min_ave_seq_logproba} ave_seq_logproba {ave_seq_logproba}" - # ) - - # We update everyone for model in models: - run_tests(model, quizz_machine, deterministic_synthesis=False) + run_tests(model, quiz_machine, deterministic_synthesis=False) ######################################################################