+ 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 = []
+
+ nb_to_create = nb_for_train + nb_for_test
+
+ # ------------------------------------------------------------
+
+ standard_validity = lambda nb_correct: torch.logical_and(
+ 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 = []
+
+for k in range(args.nb_gpts):
+ model = mygpt.MyGPT(
+ vocabulary_size=vocabulary_size,
+ dim_model=args.dim_model,
+ dim_keys=args.dim_keys,
+ dim_hidden=args.dim_hidden,
+ nb_heads=args.nb_heads,
+ nb_blocks=args.nb_blocks,
+ causal=True,
+ dropout=args.dropout,
+ ).to(device)
+
+ model.main_test_accuracy = 0.0
+ model.id = k
+
+ models.append(model)
+
+
+nb_parameters = sum(p.numel() for p in models[0].parameters())
+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))
+
+ log_string(
+ f"training model {weakest_model.id} main_test_accuracy {weakest_model.main_test_accuracy}"
+ )
+
+ one_epoch(weakest_model, quiz_machine)
+
+ log_string(
+ f"train_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
+ )
+
+ run_tests(weakest_model, quiz_machine, deterministic_synthesis=False)
+
+ log_string(
+ 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
+
+ if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
+ create_c_quizzes(
+ models,
+ quiz_machine,
+ nb_for_train=nb_new_c_quizzes_for_train,
+ nb_for_test=nb_new_c_quizzes_for_test,
+ )
+
+ for model in models:
+ run_tests(model, quiz_machine, deterministic_synthesis=False)