X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=6137834e97ed6fd85e444f99204b4a7f075edba0;hb=239a52ec7face6fcd4515916e80813702fbdf49b;hp=2d5b14891aa00487794868054a2630a6fead6e33;hpb=9f787901b2c7591a323f843ab973fe6abcf6b8ce;p=culture.git diff --git a/main.py b/main.py index 2d5b148..6137834 100755 --- a/main.py +++ b/main.py @@ -13,13 +13,12 @@ from torch.nn import functional as F import ffutils import mygpt -import sky, quizz_machine +import sky, wireworld, quizz_machine # world quizzes vs. culture quizzes ###################################################################### -accuracy_to_make_c_quizzes = 0.975 nb_new_c_quizzes_for_train = 1000 nb_new_c_quizzes_for_test = 100 @@ -38,7 +37,7 @@ parser = argparse.ArgumentParser( formatter_class=argparse.ArgumentDefaultsHelpFormatter, ) -parser.add_argument("--log_filename", type=str, default="train.log", help=" ") +parser.add_argument("--log_filename", type=str, default="train.log") parser.add_argument("--result_dir", type=str, default=None) @@ -58,7 +57,7 @@ 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-4) +parser.add_argument("--learning_rate", type=float, default=1e-3) ######################################## @@ -80,10 +79,34 @@ parser.add_argument("--dropout", type=float, default=0.1) parser.add_argument("--deterministic_synthesis", action="store_true", default=False) +parser.add_argument("--reverse_cleanup", action="store_true", default=False) + +parser.add_argument("--problem", type=str, default="sky") + parser.add_argument("--nb_gpts", type=int, default=5) +parser.add_argument("--nb_models_for_generation", type=int, default=1) + +parser.add_argument("--generation_mode", type=str, default="groupthink") + +parser.add_argument("--min_to_validate", type=int, default=4) + +parser.add_argument("--max_to_validate", type=int, default=4) + +parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975) + parser.add_argument("--dirty_debug", action="store_true", default=False) +parser.add_argument("--sky_height", type=int, default=6) + +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) + ###################################################################### args = parser.parse_args() @@ -94,7 +117,7 @@ if args.result_dir is None: ###################################################################### if args.dirty_debug: - accuracy_to_make_c_quizzes = 0.0 + args.accuracy_to_make_c_quizzes = 0.0 nb_new_c_quizzes_for_train = 100 nb_new_c_quizzes_for_test = 10 @@ -103,7 +126,7 @@ if args.dirty_debug: default_args = { "model": "37M", "batch_size": 100, - "nb_train_samples": 250000, + "nb_train_samples": 100000, "nb_test_samples": 10000, } @@ -210,8 +233,21 @@ else: 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=args.sky_height, + width=args.sky_width, + nb_birds=args.sky_nb_birds, + nb_iterations=args.sky_nb_iterations, + speed=args.sky_speed, + ) +elif args.problem == "wireworld": + problem = wireworld.Wireworld(height=8, width=10, nb_iterations=2, speed=5) +else: + raise ValueError + quizz_machine = quizz_machine.QuizzMachine( - sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2), + problem=problem, nb_train_samples=args.nb_train_samples, nb_test_samples=args.nb_test_samples, batch_size=args.physical_batch_size, @@ -334,7 +370,6 @@ def run_tests(model, quizz_machine, deterministic_synthesis): n_epoch=n_epoch, model=model, result_dir=args.result_dir, - logger=log_string, deterministic_synthesis=deterministic_synthesis, ) @@ -348,62 +383,79 @@ def run_tests(model, quizz_machine, deterministic_synthesis): ###################################################################### +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([]) + + +###################################################################### + + def create_c_quizzes( models, quizz_machine, nb_for_train=1000, nb_for_test=100, - min_ave_seq_logproba=None, ): - kept = [] - model_indexes = [] + recorded = [] + sum_logits, sum_nb_c_quizzes = 0, 0 - while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test: - nb_to_generate = nb_for_train + nb_for_test + nb_to_create = nb_for_train + nb_for_test - if len(model_indexes) == 0: - model_indexes = [i.item() for i in torch.randperm(len(models))] + # ------------------------------------------------------------ - model = models[model_indexes.pop()] + standard_validity = lambda nb_correct: torch.logical_and( + nb_correct >= args.min_to_validate, nb_correct <= args.max_to_validate + ) - new_c_quizzes, nb_correct, ave_seq_logproba = quizz_machine.create_c_quizzes( - nb=nb_to_generate, - model_for_generation=model, - models_for_validation=models, - min_ave_seq_logproba=min_ave_seq_logproba, - n_epoch=n_epoch, - result_dir=args.result_dir, - logger=log_string, + while valid_c_quizzes(recorded, standard_validity).size(0) < nb_to_create: + model_for_generation = models[torch.randint(len(models), (1,))] + + c_quizzes, ave_seq_logproba = quizz_machine.generate_quizzes( + nb_to_create, + model_for_generation=model_for_generation, + reverse_cleanup=args.reverse_cleanup, ) - sum_logits += new_c_quizzes.size(0) * ave_seq_logproba - sum_nb_c_quizzes += new_c_quizzes.size(0) + sum_logits += c_quizzes.size(0) * ave_seq_logproba + sum_nb_c_quizzes += c_quizzes.size(0) - to_keep = new_c_quizzes[nb_correct == len(models) - 1] + nb_correct = quizz_machine.comput_correctness(c_quizzes, models) if args.dirty_debug: - to_keep = new_c_quizzes[ - torch.randint(3, (new_c_quizzes.size(0),), device=new_c_quizzes.device) - == 0 - ] + nb_correct = torch.randint( + len(models) + 1, nb_correct.size(), device=c_quizzes.device + ) - kept.append(to_keep) + recorded.append((c_quizzes, nb_correct)) - log_string( - f"keep c_quizzes {to_keep.size(0)}/{new_c_quizzes.size(0)} ({to_keep.size(0)*100/new_c_quizzes.size(0):.02f}%) total {sum([ x.size(0) for x in kept])}/{nb_to_generate}" - ) + nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0) + nv = " ".join([str(x.item()) for x in nv]) - new_c_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test] + nb_validated = valid_c_quizzes(recorded, standard_validity).size(0) + + log_string(f"keep c_quizzes kept {nv} total {nb_validated} / {nb_to_create}") + + # ------------------------------------------------------------ + + new_c_quizzes = valid_c_quizzes(recorded, standard_validity) 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) - quizz_machine.problem.save_quizzes( - new_c_quizzes[:72], - args.result_dir, - f"culture_c_quiz_{n_epoch:04d}_{model.id:02d}", - ) + 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( + valid_c_quizzes(recorded, criteria=lambda nb_correct: nb_correct == n)[:72], + args.result_dir, + f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", + ) return sum_logits / sum_nb_c_quizzes @@ -435,56 +487,40 @@ log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)") ###################################################################### -min_ave_seq_logproba = None - 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]) - log_string(f"current accuracies {a}") - - # select the model with lowest accuracy - models.sort(key=lambda model: model.main_test_accuracy) - model = models[0] + 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, quizz_machine) log_string( f"train_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}" ) # test it - run_tests(model, quizz_machine, deterministic_synthesis=False) + run_tests(weakest_model, quizz_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}" ) - if min([m.main_test_accuracy for m in models]) >= accuracy_to_make_c_quizzes: - ave_seq_logproba = create_c_quizzes( + # replace a fraction of the w_quizzes with a fresh ones + quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts) + + if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes: + create_c_quizzes( models, quizz_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)