parser.add_argument("--nb_gpts", type=int, default=5)
-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)
sum_logits += c_quizzes.size(0) * ave_seq_logproba
sum_nb_c_quizzes += c_quizzes.size(0)
- nb_correct = quizz_machine.comput_correctness(c_quizzes, models)
+ nb_correct = quizz_machine.compute_correctness(c_quizzes, models)
if args.dirty_debug:
nb_correct = torch.randint(
nb_validated = valid_c_quizzes(recorded, standard_validity).size(0)
- log_string(f"keep c_quizzes kept {nv} total {nb_validated} / {nb_to_create}")
+ log_string(
+ f"keep c_quizzes kept {nv} nb_accumulated {nb_validated} / {nb_to_create}"
+ )
# ------------------------------------------------------------
f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
)
+ cta = " ".join([f"{float(m.main_test_accuracy):.02f}" for m in models])
+ log_string(f"current_test_accuracies {cta}")
+
# replace a fraction of the w_quizzes with a fresh ones
quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)