X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=aefc3a10b5c2d1e402ee97deba6310f5aa212485;hb=870d6808ac616b81cae00d9cb1f4de29bae23410;hp=597ec32f71f5232ddbbc7d26514ce1d238112816;hpb=ceddc8cc3adbb045fdef1ccb0b3df2b8fed9eb4c;p=culture.git diff --git a/main.py b/main.py index 597ec32..aefc3a1 100755 --- a/main.py +++ b/main.py @@ -12,18 +12,14 @@ from torch import nn from torch.nn import functional as F import ffutils + import mygpt -import sky, reasoning, quiz_machine +import sky, grids, quiz_machine # 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 @@ -79,7 +75,9 @@ 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("--nb_threads", type=int, default=-1) parser.add_argument("--nb_gpts", type=int, default=5) @@ -124,16 +122,9 @@ if args.result_dir is None: ###################################################################### -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 - -###################################################################### - default_args = { "model": "37M", - "batch_size": 100, + "batch_size": 25, "nb_train_samples": 100000, "nb_test_samples": 10000, } @@ -248,10 +239,18 @@ if args.problem == "sky": nb_birds=args.sky_nb_birds, nb_iterations=args.sky_nb_iterations, speed=args.sky_speed, + max_nb_cached_chunks=args.nb_train_samples // 100, + chunk_size=100, + nb_threads=args.nb_threads, ) back_accuracy = False -elif args.problem == "reasoning": - problem = reasoning.Reasoning(device=device) +elif args.problem == "grids": + problem = grids.Grids( + device=device, + max_nb_cached_chunks=args.nb_train_samples // 100, + chunk_size=100, + nb_threads=args.nb_threads, + ) back_accuracy = True else: raise ValueError @@ -392,8 +391,13 @@ def run_tests(model, quiz_machine, deterministic_synthesis): ###################################################################### +def standard_validity(logproba): + l = logproba.sort(dim=-1).values + return logical_and(l[0] < math.log(0.5), l[1] > math.log(0.95)) + + def valid_c_quizzes(recorded, criteria): - result = [q[criteria(c)] for q, c in recorded] + result = [q[criteria(lp)] for q, lp in recorded] return torch.cat(result, dim=0) if len(result) > 0 else torch.tensor([]) @@ -405,6 +409,80 @@ def create_c_quizzes( quiz_machine, nb_for_train=1000, nb_for_test=100, +): + quizzes_and_logproba_records = [] + + nb_to_create = nb_for_train + nb_for_test + + # ------------------------------------------------------------ + + 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_logproba_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,))] + + c_quizzes = quiz_machine.generate_quizzes( + nb_to_create, + model_for_generation=model_for_generation, + temperature=args.generation_temperature, + ) + + c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)] + + if c_quizzes.size(0) > 0: + logproba = c_quizzes.new(c_quizzes.size(0), len(models)) + for q, l in zip( + c_quizzes.split(args.batch_size), logits.split(args.batch_size) + ): + for model in models: + l[model.id] = F.cross_entropy(model(q)) + + for l in logproba: + s = " ".join([str(x.item()) for x in l]) + logp_file.write(s + "\n") + + quizzes_and_logproba_records.append((c_quizzes, logproba)) + + nb_validated = valid_c_quizzes( + quizzes_and_logproba_records, standard_validity + ).size(0) + + log_string( + f"keep c_quizzes model {model_for_generation.id} nb_accumulated {nb_validated} / {nb_to_create}" + ) + + # store the new c_quizzes which have been validated + + new_c_quizzes = valid_c_quizzes(quizzes_and_logproba_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 + + q = new_c_quizzes[:72] + + if q.size(0) > 0: + quiz_machine.save_quizzes(args.result_dir, f"culture_c_quiz_{n_epoch:04d}", q) + + +###################################################################### + + +def create_c_quizzes_( + models, + quiz_machine, + nb_for_train=1000, + nb_for_test=100, ): quizzes_and_nb_correct_records = [] @@ -417,6 +495,7 @@ def create_c_quizzes( ) 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) @@ -432,23 +511,34 @@ def create_c_quizzes( temperature=args.generation_temperature, ) - nb_correct, seq_logproba = quiz_machine.compute_correctness( - c_quizzes, - models, - bidirectional_validation=args.bidirectional_validation, - deterministic_validation=args.deterministic_validation, - ) + # if args.prediction_correctness: - 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") + # else: + # logproba = quiz_machine.new(quiz_machine.size(0), len(models)) + # for q,l in zip(quizzes.split(args.batch_size), logits.split(args.batch_size)): + # for model in models: + # l[...] = F.cross_entropy(model(q)) - if args.dirty_debug: - nb_correct = torch.randint( - len(models) + 1, nb_correct.size(), device=c_quizzes.device + c_quizzes = c_quizzes[quiz_machine.non_trivial(c_quizzes)] + + 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, ) - quizzes_and_nb_correct_records.append((c_quizzes, nb_correct)) + 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 + ) + + quizzes_and_nb_correct_records.append((c_quizzes, nb_correct)) nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0) nv = " ".join([str(x.item()) for x in nv]) @@ -519,12 +609,30 @@ log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)") ###################################################################### +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} ----------------------------------------") cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models]) log_string(f"current_test_accuracies {cta}") + ################################################## # Select, improve, and eval the worst model weakest_model = min(models, key=lambda m: float(m.main_test_accuracy)) @@ -545,10 +653,12 @@ for n_epoch in range(args.nb_epochs): 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