X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;ds=sidebyside;f=main.py;h=7f9d5210a486971126dcbf948224d8f296f79bf1;hb=7ad1043be4c7f85625e164fd586bc71096f93e5b;hp=ebecad8a471353400fc7e6e472bdb95f594b48ab;hpb=15192743a5dee8d88650319d64610f1603d21472;p=culture.git diff --git a/main.py b/main.py index ebecad8..7f9d521 100755 --- a/main.py +++ b/main.py @@ -12,7 +12,8 @@ from torch import nn from torch.nn import functional as F import ffutils -import mygpt, tasks +import mygpt +import sky, quizz_machine # world quizzes vs. culture quizzes @@ -209,7 +210,8 @@ else: assert args.nb_train_samples % args.batch_size == 0 assert args.nb_test_samples % args.batch_size == 0 -task = tasks.World( +quizz_machine = quizz_machine.QuizzMachine( + problem=sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2), nb_train_samples=args.nb_train_samples, nb_test_samples=args.nb_test_samples, batch_size=args.physical_batch_size, @@ -222,7 +224,7 @@ task = tasks.World( log_string(f"device {device}") -vocabulary_size = task.vocabulary_size() +vocabulary_size = quizz_machine.vocabulary_size() log_string(f"vocabulary_size {vocabulary_size}") @@ -231,8 +233,10 @@ log_string(f"vocabulary_size {vocabulary_size}") # Compute the entropy of the training tokens token_count = 0 -for input in task.batches(split="train", desc="train-entropy"): - token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1)) +for input in quizz_machine.batches(split="train", desc="train-entropy"): + token_count += F.one_hot(input, num_classes=quizz_machine.vocabulary_size()).sum( + (0, 1) + ) token_probas = token_count / token_count.sum() entropy = -torch.xlogy(token_probas, token_probas).sum() train_set_perplexity = math.exp(entropy) @@ -254,11 +258,11 @@ if args.max_percents_of_test_in_train >= 0: nb_test, nb_in_train = 0, 0 for test_subset in subsets_as_tuples( - task.batches(split="test", desc="test-check"), 25000 + quizz_machine.batches(split="test", desc="test-check"), 25000 ): in_train = set() for train_subset in subsets_as_tuples( - task.batches(split="train", desc="train-check"), 25000 + quizz_machine.batches(split="train", desc="train-check"), 25000 ): in_train.update(test_subset.intersection(train_subset)) nb_in_train += len(in_train) @@ -275,14 +279,14 @@ if args.max_percents_of_test_in_train >= 0: ############################## -def one_epoch(model, task): +def one_epoch(model, quizz_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 task.batches(split="train"): + for input in quizz_machine.batches(split="train"): input = input.to(device) if nb_train_samples % args.batch_size == 0: @@ -307,14 +311,14 @@ def one_epoch(model, task): ###################################################################### -def run_tests(model, task, deterministic_synthesis): +def run_tests(model, quizz_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 task.batches(split="test"): + for input in quizz_machine.batches(split="test"): input = input.to(device) bs = model(mygpt.BracketedSequence(input)) @@ -326,7 +330,7 @@ def run_tests(model, task, deterministic_synthesis): nb_test_samples += input.size(0) - main_test_accuracy = task.produce_results( + main_test_accuracy = quizz_machine.produce_results( n_epoch=n_epoch, model=model, result_dir=args.result_dir, @@ -345,54 +349,60 @@ def run_tests(model, task, deterministic_synthesis): def create_c_quizzes( - model, - other_models, - task, + models, + quizz_machine, nb_for_train=1000, nb_for_test=100, - desired_average_logits=None, + min_ave_seq_logproba=None, ): kept = [] - + model_indexes = [] 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 = 4 * (nb_for_train + nb_for_test) + nb_to_generate = nb_for_train + nb_for_test + + if len(model_indexes) == 0: + model_indexes = [i.item() for i in torch.randperm(len(models))] - new_c_quizzes, nb_correct, average_logits = task.create_c_quizzes( + model = models[model_indexes.pop()] + + 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, - nb=nb_to_generate, - model=model, - other_models=other_models, - desired_average_logits=desired_average_logits, ) - sum_logits += new_c_quizzes.size(0) * average_logits + sum_logits += new_c_quizzes.size(0) * ave_seq_logproba sum_nb_c_quizzes += new_c_quizzes.size(0) - to_keep = new_c_quizzes[nb_correct == len(other_models) - 1] + to_keep = new_c_quizzes[nb_correct == len(models) - 1] if args.dirty_debug: - to_keep = new_c_quizzes + to_keep = new_c_quizzes[ + torch.randint(3, (new_c_quizzes.size(0),), device=new_c_quizzes.device) + == 0 + ] + + kept.append(to_keep) log_string( - f"keep {to_keep.size(0)}/{new_c_quizzes.size(0)} c_quizzes ({to_keep.size(0)*100/new_c_quizzes.size(0):.02f}%)" + 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}" ) - kept.append(to_keep) - new_c_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test] - task.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True) - task.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False) + 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) - task.save_quizzes( + quizz_machine.problem.save_quizzes( new_c_quizzes[:72], args.result_dir, f"culture_c_quiz_{n_epoch:04d}_{model.id:02d}", - log_string, ) return sum_logits / sum_nb_c_quizzes @@ -425,7 +435,7 @@ log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)") ###################################################################### -desired_average_logits = None +min_ave_seq_logproba = None for n_epoch in range(args.nb_epochs): log_string(f"--- epoch {n_epoch} ----------------------------------------") @@ -443,45 +453,41 @@ for n_epoch in range(args.nb_epochs): ) # improve it - one_epoch(model, task) + one_epoch(model, quizz_machine) - task.renew_w_quizzes(args.nb_train_samples // args.nb_gpts) + quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts) log_string( - f"train_set_composition w_quizzes {task.nb_batch_w_quizzes} c_quizzes {task.nb_batch_c_quizzes}" + 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, task, deterministic_synthesis=False) + run_tests(model, quizz_machine, deterministic_synthesis=False) log_string( - f"test_set_composition w_quizzes {task.nb_batch_w_quizzes} c_quizzes {task.nb_batch_c_quizzes}" + 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: - other_models = models.copy() - other_models.remove(model) - - average_logits = create_c_quizzes( - model, - other_models, - task, + ave_seq_logproba = create_c_quizzes( + models, + quizz_machine, nb_for_train=nb_new_c_quizzes_for_train, nb_for_test=nb_new_c_quizzes_for_test, - desired_average_logits=desired_average_logits, + min_ave_seq_logproba=min_ave_seq_logproba, ) # We keep the first average logits as a reference - if desired_average_logits is None: - desired_average_logits = average_logits + if min_ave_seq_logproba is None: + min_ave_seq_logproba = ave_seq_logproba else: log_string( - f"desired_average_logits {desired_average_logits} average_logits {average_logits}" + 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, task, deterministic_synthesis=False) + run_tests(model, quizz_machine, deterministic_synthesis=False) ######################################################################