X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=918f75d8b1e0ff4ae3f04f3310026ab82afd81ea;hb=64abc9f3a07a8211f308271fde7d8f876a968ab5;hp=d412e6c60f6ac119ac9a706ddf27c88120be5b36;hpb=bc9ed7c97f932ebc81f573c4ecd7207b82a011d7;p=culture.git diff --git a/main.py b/main.py index d412e6c..918f75d 100755 --- a/main.py +++ b/main.py @@ -79,9 +79,7 @@ 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=True) - -parser.add_argument("--validation_forward_only", action="store_true", default=False) +parser.add_argument("--both_directions", action="store_true", default=False) parser.add_argument("--problem", type=str, default="sky") @@ -95,6 +93,12 @@ 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("--generation_temperature", type=float, default=1.0) + +parser.add_argument("--stochastic_validation", action="store_true", default=False) + +###################################################################### + parser.add_argument("--sky_height", type=int, default=6) parser.add_argument("--sky_width", type=int, default=8) @@ -364,6 +368,10 @@ def run_tests(model, quizz_machine, deterministic_synthesis): nb_test_samples += input.size(0) + 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 = quizz_machine.produce_results( n_epoch=n_epoch, model=model, @@ -371,10 +379,6 @@ def run_tests(model, quizz_machine, deterministic_synthesis): 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}") - ###################################################################### @@ -403,34 +407,45 @@ def create_c_quizzes( nb_correct >= args.min_to_validate, nb_correct <= args.max_to_validate ) - while valid_c_quizzes(recorded, standard_validity).size(0) < nb_to_create: - model_for_generation = models[torch.randint(len(models), (1,))] + 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(recorded, standard_validity).size(0) < nb_to_create: + # Select a model at random to generate the new quizzes - c_quizzes, ave_seq_logproba = quizz_machine.generate_quizzes( - nb_to_create, - model_for_generation=model_for_generation, - reverse_cleanup=args.reverse_cleanup, - ) + model_for_generation = models[torch.randint(len(models), (1,))] - nb_correct = quizz_machine.compute_correctness( - c_quizzes, models, both_directions=not args.validation_forward_only - ) + c_quizzes = quizz_machine.generate_quizzes( + nb_to_create, + model_for_generation=model_for_generation, + temperature=args.generation_temperature, + ) - if args.dirty_debug: - nb_correct = torch.randint( - len(models) + 1, nb_correct.size(), device=c_quizzes.device + nb_correct, seq_logproba = quizz_machine.compute_correctness( + c_quizzes, + models, + both_directions=args.both_directions, + deterministic_validation=not args.stochastic_validation, ) - recorded.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") - nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0) - nv = " ".join([str(x.item()) for x in nv]) + if args.dirty_debug: + nb_correct = torch.randint( + len(models) + 1, nb_correct.size(), device=c_quizzes.device + ) - nb_validated = valid_c_quizzes(recorded, standard_validity).size(0) + recorded.append((c_quizzes, nb_correct)) - log_string( - f"keep c_quizzes kept {nv} nb_accumulated {nb_validated} / {nb_to_create}" - ) + nv = F.one_hot(nb_correct, num_classes=len(models) + 1).sum(0) + nv = " ".join([str(x.item()) for x in nv]) + + nb_validated = valid_c_quizzes(recorded, 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 @@ -452,10 +467,8 @@ def create_c_quizzes( q = valid_c_quizzes(recorded, criteria=lambda nb_correct: nb_correct == n)[:72] if q.size(0) > 0: - quizz_machine.problem.save_quizzes( - q, - args.result_dir, - f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", + quizz_machine.save_quizzes( + args.result_dir, f"culture_c_quiz_{n_epoch:04d}_N{n}{s}", q ) @@ -489,20 +502,20 @@ log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)") for n_epoch in range(args.nb_epochs): log_string(f"--- epoch {n_epoch} ----------------------------------------") + # Select, improve, and eval the worst model + weakest_model = min(models, key=lambda m: float(m.main_test_accuracy)) log_string( f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}" ) - # improve it 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(weakest_model, quizz_machine, deterministic_synthesis=False) log_string( @@ -512,9 +525,13 @@ for n_epoch in range(args.nb_epochs): cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models]) log_string(f"current_test_accuracies {cta}") - # replace a fraction of the w_quizzes with fresh ones + # Replace a fraction of the w_quizzes with fresh ones + quizz_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: create_c_quizzes( models, @@ -523,7 +540,6 @@ for n_epoch in range(args.nb_epochs): nb_for_test=nb_new_c_quizzes_for_test, ) - # We update everyone for model in models: run_tests(model, quizz_machine, deterministic_synthesis=False)