parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
-parser.add_argument("--reverse_cleanup", action="store_true", default=False)
+parser.add_argument("--reverse_cleanup", action="store_true", default=True)
+
+parser.add_argument("--validation_forward_only", action="store_true", default=False)
parser.add_argument("--problem", type=str, default="sky")
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)
nb_test_samples += input.size(0)
- main_test_accuracy = quizz_machine.produce_results(
+ model.main_test_accuracy = quizz_machine.produce_results(
n_epoch=n_epoch,
model=model,
result_dir=args.result_dir,
log_string(f"test_perplexity {n_epoch} {test_perplexity}")
- model.main_test_accuracy = main_test_accuracy
-
######################################################################
):
recorded = []
- sum_logits, sum_nb_c_quizzes = 0, 0
-
nb_to_create = nb_for_train + nb_for_test
# ------------------------------------------------------------
reverse_cleanup=args.reverse_cleanup,
)
- 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, both_directions=not args.validation_forward_only
+ )
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}"
+ )
- # ------------------------------------------------------------
+ # store the new c_quizzes which have been validated
new_c_quizzes = valid_c_quizzes(recorded, standard_validity)
quizz_machine.store_c_quizzes(new_c_quizzes[:nb_for_train], for_train=True)
quizz_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"
else ""
)
- quizz_machine.problem.save_quizzes(
- valid_c_quizzes(recorded, criteria=lambda nb_correct: nb_correct == n)[:72],
- args.result_dir,
- f"culture_c_quiz_{n_epoch:04d}_N{n}{s}",
- )
+ q = valid_c_quizzes(recorded, criteria=lambda nb_correct: nb_correct == n)[:72]
- return sum_logits / sum_nb_c_quizzes
+ if q.size(0) > 0:
+ quizz_machine.problem.save_quizzes(
+ q,
+ args.result_dir,
+ f"culture_c_quiz_{n_epoch:04d}_N{n}{s}",
+ )
######################################################################
f"test_set_composition w_quizzes {quizz_machine.nb_batch_w_quizzes} c_quizzes {quizz_machine.nb_batch_c_quizzes}"
)
- # replace a fraction of the w_quizzes with a fresh ones
+ cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
+ log_string(f"current_test_accuracies {cta}")
+
+ # replace a fraction of the w_quizzes with fresh ones
quizz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
if min([m.main_test_accuracy for m in models]) >= args.accuracy_to_make_c_quizzes: