parser.add_argument("--nb_test_samples", type=int, default=None)
-parser.add_argument("--learning_rate", type=float, default=1e-4)
+parser.add_argument("--learning_rate", type=float, default=1e-3)
########################################
default_args = {
"model": "37M",
"batch_size": 100,
- "nb_train_samples": 250000,
+ "nb_train_samples": 100000,
"nb_test_samples": 10000,
}
assert args.nb_test_samples % args.batch_size == 0
quizz_machine = quizz_machine.QuizzMachine(
- sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2),
+ problem=sky.Sky(height=6, width=8, nb_birds=3, nb_iterations=2),
nb_train_samples=args.nb_train_samples,
nb_test_samples=args.nb_test_samples,
batch_size=args.physical_batch_size,
def create_c_quizzes(
- model,
- other_models,
+ models,
quizz_machine,
nb_for_train=1000,
nb_for_test=100,
min_ave_seq_logproba=None,
):
kept = []
-
+ model_indexes = []
sum_logits, sum_nb_c_quizzes = 0, 0
while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
- nb_to_generate = 4 * (nb_for_train + nb_for_test)
+ nb_to_generate = 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_generate,
+ model_for_generation=model,
+ models_for_validation=models,
+ min_ave_seq_logproba=min_ave_seq_logproba,
n_epoch=n_epoch,
result_dir=args.result_dir,
logger=log_string,
- nb=nb_to_generate,
- model=model,
- other_models=other_models,
- min_ave_seq_logproba=min_ave_seq_logproba,
)
sum_logits += new_c_quizzes.size(0) * ave_seq_logproba
sum_nb_c_quizzes += new_c_quizzes.size(0)
- to_keep = new_c_quizzes[nb_correct == len(other_models) - 1]
+ to_keep = new_c_quizzes[nb_correct == len(models) - 1]
if args.dirty_debug:
- to_keep = new_c_quizzes
+ to_keep = new_c_quizzes[
+ torch.randint(3, (new_c_quizzes.size(0),), device=new_c_quizzes.device)
+ == 0
+ ]
+
+ kept.append(to_keep)
log_string(
- f"keep {to_keep.size(0)}/{new_c_quizzes.size(0)} c_quizzes ({to_keep.size(0)*100/new_c_quizzes.size(0):.02f}%)"
+ f"keep c_quizzes {to_keep.size(0)}/{new_c_quizzes.size(0)} ({to_keep.size(0)*100/new_c_quizzes.size(0):.02f}%) total {sum([ x.size(0) for x in kept])}/{nb_to_generate}"
)
- kept.append(to_keep)
-
- new_c_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
+ new_c_quizzes = torch.cat(kept, dim=0)
+ new_c_quizzes = new_c_quizzes[
+ torch.randperm(new_c_quizzes.size(0))[: nb_for_train + nb_for_test]
+ ]
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)
)
if min([m.main_test_accuracy for m in models]) >= accuracy_to_make_c_quizzes:
- other_models = models.copy()
- other_models.remove(model)
-
ave_seq_logproba = create_c_quizzes(
- model,
- other_models,
+ models,
quizz_machine,
nb_for_train=nb_new_c_quizzes_for_train,
nb_for_test=nb_new_c_quizzes_for_test,
)
# We keep the first average logits as a reference
- if min_ave_seq_logproba is None:
- min_ave_seq_logproba = ave_seq_logproba
- else:
- log_string(
- f"min_ave_seq_logproba {min_ave_seq_logproba} ave_seq_logproba {ave_seq_logproba}"
- )
+ # if min_ave_seq_logproba is None:
+ # min_ave_seq_logproba = ave_seq_logproba
+ # else:
+ # log_string(
+ # f"min_ave_seq_logproba {min_ave_seq_logproba} ave_seq_logproba {ave_seq_logproba}"
+ # )
# We update everyone
for model in models: