######################################################################
-accuracy_to_make_c_quizzes = 0.975
nb_new_c_quizzes_for_train = 1000
nb_new_c_quizzes_for_test = 100
parser.add_argument("--nb_gpts", type=int, default=5)
-parser.add_argument("--nb_correct_to_validate", type=int, default=4)
+parser.add_argument("--nb_models_for_generation", type=int, default=1)
+
+parser.add_argument("--generation_mode", type=str, default="groupthink")
+
+parser.add_argument("--min_to_validate", type=int, default=4)
+
+parser.add_argument("--max_to_validate", type=int, default=4)
+
+parser.add_argument("--accuracy_to_make_c_quizzes", type=float, default=0.975)
parser.add_argument("--dirty_debug", action="store_true", default=False)
######################################################################
if args.dirty_debug:
- accuracy_to_make_c_quizzes = 0.0
+ args.accuracy_to_make_c_quizzes = 0.0
nb_new_c_quizzes_for_train = 100
nb_new_c_quizzes_for_test = 10
return sum(
[
sum([x.size(0) for x in recorded[n]])
- for n in range(args.nb_correct_to_validate, len(models))
+ for n in range(args.min_to_validate, args.max_to_validate + 1)
]
)
- while nb_validated() < nb_for_train + nb_for_test:
- nb_to_validate = nb_for_train + nb_for_test
-
- if len(model_indexes) == 0:
- model_indexes = [i.item() for i in torch.randperm(len(models))]
-
- model = models[model_indexes.pop()]
-
- new_c_quizzes, nb_correct, ave_seq_logproba = quizz_machine.create_c_quizzes(
- nb=nb_to_validate,
- model_for_generation=model,
- models_for_validation=models,
+ nb_to_create = nb_for_train + nb_for_test
+
+ while nb_validated() < nb_to_create:
+ (
+ new_c_quizzes,
+ nb_correct,
+ ave_seq_logproba,
+ ) = quizz_machine.gang_create_c_quizzes(
+ nb=nb_to_create,
+ nb_models_for_generation=args.nb_models_for_generation,
+ models=models,
+ mode=args.generation_mode,
min_ave_seq_logproba=min_ave_seq_logproba,
n_epoch=n_epoch,
result_dir=args.result_dir,
recorded[n].append(new_c_quizzes[nb_correct == n].clone())
log_string(
- f"keep c_quizzes {nb_validated()*100/nb_generated():.02f}% kept total {nb_validated()} / {nb_to_validate}"
+ f"keep c_quizzes {nb_validated()*100/nb_generated():.02f}% kept total {nb_validated()} / {nb_to_create}"
)
# concatenate and shuffle
del recorded[n]
new_c_quizzes = torch.cat(
- [recorded[n] for n in range(args.nb_correct_to_validate, len(models))], dim=0
+ [recorded[n] for n in range(args.min_to_validate, args.max_to_validate + 1)],
+ dim=0,
)
new_c_quizzes = new_c_quizzes[
quizz_machine.store_c_quizzes(new_c_quizzes[nb_for_train:], for_train=False)
for n in recorded.keys():
- s = "_validated" if n >= args.nb_correct_to_validate and n < len(models) else ""
+ s = (
+ "_validated"
+ if n >= args.min_to_validate and n <= args.max_to_validate
+ else ""
+ )
quizz_machine.problem.save_quizzes(
recorded[n][:72],
args.result_dir,
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:
+ if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes:
ave_seq_logproba = create_c_quizzes(
models,
quizz_machine,