X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;ds=inline;f=main.py;h=b88cbc4b545c553e36629d43b174cfc056250bd9;hb=7c79c0b140c88a529962945ec5b482fe90c55581;hp=0c193f71b7cdf04f6aac31a392d6ec22b81675fc;hpb=522b1a3210d9ad0246372551f028079bbed3da76;p=culture.git diff --git a/main.py b/main.py index 0c193f7..b88cbc4 100755 --- a/main.py +++ b/main.py @@ -16,6 +16,10 @@ import ffutils import mygpt import sky, grids, quiz_machine +import threading + +import torch.multiprocessing as mp + # world quizzes vs. culture quizzes ###################################################################### @@ -38,7 +42,7 @@ parser.add_argument("--result_dir", type=str, default=None) parser.add_argument("--seed", type=int, default=0) -parser.add_argument("--max_percents_of_test_in_train", type=int, default=1) +parser.add_argument("--max_percents_of_test_in_train", type=int, default=-1) ######################################## @@ -78,6 +82,8 @@ parser.add_argument("--problem", type=str, default="grids") parser.add_argument("--nb_threads", type=int, default=1) +parser.add_argument("--nb_gpus", type=int, default=1) + parser.add_argument("--nb_gpts", type=int, default=5) parser.add_argument("--min_to_validate", type=int, default=None) @@ -88,10 +94,6 @@ parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975) parser.add_argument("--generation_temperature", type=float, default=2.0) -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) ###################################################################### @@ -238,14 +240,14 @@ 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, + max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100, chunk_size=100, nb_threads=args.nb_threads, ) back_accuracy = False elif args.problem == "grids": problem = grids.Grids( - max_nb_cached_chunks=args.nb_train_samples // 100, + max_nb_cached_chunks=args.nb_gpus * args.nb_train_samples // 100, chunk_size=100, nb_threads=args.nb_threads, ) @@ -253,6 +255,8 @@ elif args.problem == "grids": else: raise ValueError +problem.save_some_examples(args.result_dir) + quiz_machine = quiz_machine.QuizMachine( problem=problem, nb_train_samples=args.nb_train_samples, @@ -273,50 +277,23 @@ vocabulary_size = quiz_machine.vocabulary_size() log_string(f"vocabulary_size {vocabulary_size}") ###################################################################### -############################## - - -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 quiz_machine.batches(model, split="train"): - input = input.to(device) - - if nb_train_samples % args.batch_size == 0: - optimizer.zero_grad() - - output = model(mygpt.BracketedSequence(input)).x - loss = F.cross_entropy(output.transpose(1, 2), input) - acc_train_loss += loss.item() * input.size(0) - - nb_train_samples += input.size(0) - - loss.backward() - - if nb_train_samples % args.batch_size == 0: - optimizer.step() - - train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples)) - - log_string(f"train_perplexity {n_epoch} {train_perplexity}") ###################################################################### -def run_tests(model, quiz_machine, deterministic_synthesis): +def run_tests(model, quiz_machine, deterministic_synthesis, local_device=None): + if local_device is None: + local_device = device + with torch.autograd.no_grad(): - model.eval() + model.eval().to(local_device) nb_test_samples, acc_test_loss = 0, 0.0 nb_samples_accumulated = 0 for input in quiz_machine.batches(model, split="test"): - input = input.to(device) + input = input.to(local_device) bs = model(mygpt.BracketedSequence(input)) output = bs.x @@ -339,15 +316,46 @@ def run_tests(model, quiz_machine, deterministic_synthesis): ) +def one_epoch(model, quiz_machine, local_device=None): + if local_device is None: + local_device = device + + optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate) + + model.to(local_device).train() + + nb_train_samples, acc_train_loss = 0, 0.0 + + for input in quiz_machine.batches(model, split="train"): + input = input.to(local_device) + + if nb_train_samples % args.batch_size == 0: + optimizer.zero_grad() + + output = model(mygpt.BracketedSequence(input)).x + loss = F.cross_entropy(output.transpose(1, 2), input) + acc_train_loss += loss.item() * input.size(0) + + nb_train_samples += input.size(0) + + loss.backward() + + if nb_train_samples % args.batch_size == 0: + optimizer.step() + + train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples)) + + log_string(f"train_perplexity {n_epoch} model.id {model.id} {train_perplexity}") + + run_tests(model, quiz_machine, deterministic_synthesis=False) + + ###################################################################### def standard_validity(logproba): l = logproba.sort(dim=-1).values return (l[:, 0] < math.log(0.5)) & (l[:, 1] > math.log(0.99)) - # warnings.warn("TEST!!!", RuntimeWarning) - # print(l.exp()) - # return (l[:, 0] < math.log(0.99)) def valid_c_quizzes(recorded, criteria): @@ -422,113 +430,6 @@ def create_c_quizzes( 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 = [] - - nb_to_create = nb_for_train + nb_for_test - - # ------------------------------------------------------------ - - standard_validity = lambda nb_correct: (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,))] - - c_quizzes = quiz_machine.generate_quizzes( - nb_to_create, - model_for_generation=model_for_generation, - temperature=args.generation_temperature, - ) - - # if args.prediction_correctness: - - # 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)) - - 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, - ) - - 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]) - - 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 "" - ) - - 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 - ) - - ###################################################################### models = [] @@ -549,13 +450,9 @@ for k in range(args.nb_gpts): model.main_test_accuracy = 0.0 model.id = k - model.train_w_quizzes = quiz_machine.generate_token_sequences( - args.nb_train_samples - ).to(device) + model.train_w_quizzes = quiz_machine.generate_token_sequences(args.nb_train_samples) quiz_machine.reverse_random_half_in_place(model.train_w_quizzes) - model.test_w_quizzes = quiz_machine.generate_token_sequences( - args.nb_test_samples - ).to(device) + model.test_w_quizzes = quiz_machine.generate_token_sequences(args.nb_test_samples) quiz_machine.reverse_random_half_in_place(model.test_w_quizzes) models.append(model) @@ -629,6 +526,11 @@ if args.dirty_debug: nb_new_c_quizzes_for_train = 100 nb_new_c_quizzes_for_test = 10 + def standard_validity(logproba): + l = logproba.sort(dim=-1).values + return l[:, 0] < math.log(0.5) + + ###################################################################### for n_epoch in range(args.nb_epochs): @@ -640,23 +542,25 @@ for n_epoch in range(args.nb_epochs): ################################################## # Select, improve, and eval the worst model - weakest_model = min(models, key=lambda m: float(m.main_test_accuracy)) + ranked_models = sorted(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}" - ) + weakest_models = ranked_models[: args.nb_gpus] - one_epoch(weakest_model, quiz_machine) + threads = [] - log_string( - f"train_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}" - ) + for gpu_id, model in enumerate(weakest_models): + log_string(f"training model {model.id}") - run_tests(weakest_model, quiz_machine, deterministic_synthesis=False) + t = threading.Thread( + target=one_epoch, daemon=True, args=(model, quiz_machine, f"cuda:{gpu_id}") + ) - log_string( - f"test_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}" - ) + threads.append(t) + + t.start() + + for t in threads: + t.join() ################################################## # Replace a fraction of the w_quizzes with fresh ones @@ -667,7 +571,8 @@ for n_epoch in range(args.nb_epochs): # Renew entirely the train set - quiz_machine.renew_w_quizzes(model, args.nb_train_samples) + for model in weakest_models: + quiz_machine.renew_w_quizzes(model, args.nb_train_samples) ################################################## # If all the models are good enough, generate new quizzes and @@ -681,8 +586,4 @@ for n_epoch in range(args.nb_epochs): nb_for_test=nb_new_c_quizzes_for_test, ) - for model in models: - run_tests(model, quiz_machine, deterministic_synthesis=False) - - ######################################################################