parser.add_argument("--max_to_validate", type=int, default=None)
-parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
+parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.9)
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)
######################################################################
else:
raise ValueError
+problem.save_some_examples(args.result_dir)
+
quiz_machine = quiz_machine.QuizMachine(
problem=problem,
nb_train_samples=args.nb_train_samples,
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 = []
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(0.99)
+
+
######################################################################
for n_epoch in range(args.nb_epochs):
log_string(f"current_test_accuracies {cta}")
##################################################
- # Select, improve, and eval the worst model
+ # Select, improve, and eval the worst models
ranked_models = sorted(models, key=lambda m: float(m.main_test_accuracy))
model.TRAINING_LOCK.release()
##################################################
- # Replace a fraction of the w_quizzes with fresh ones
+ # Renew the train sets
log_string(
f"cache_w_quizzes contains {quiz_machine.problem.nb_cached_quizzes()} quizzes"
)
- # Renew entirely the train set
-
for model in weakest_models:
quiz_machine.renew_w_quizzes(model, args.nb_train_samples)