######################################################################
-nb_new_c_quizzes_for_train = 1000
-nb_new_c_quizzes_for_test = 100
-
-######################################################################
-
if torch.cuda.is_available():
device = torch.device("cuda")
torch.backends.cuda.matmul.allow_tf32 = True
######################################################################
-if args.dirty_debug:
- args.accuracy_to_make_c_quizzes = 0.0
- args.nb_gpts = 2
- nb_new_c_quizzes_for_train = 100
- nb_new_c_quizzes_for_test = 10
-
-######################################################################
-
default_args = {
"model": "37M",
"batch_size": 100,
######################################################################
+nb_new_c_quizzes_for_train = args.nb_train_samples // 50
+nb_new_c_quizzes_for_test = args.nb_test_samples // 50
+
+log_string(
+ f"nb_new_c_quizzes_for_train {nb_new_c_quizzes_for_train} nb_new_c_quizzes_for_test {nb_new_c_quizzes_for_test}"
+)
+
+######################################################################
+
+if args.dirty_debug:
+ args.accuracy_to_make_c_quizzes = 0.0
+ args.nb_gpts = 2
+ nb_new_c_quizzes_for_train = 100
+ nb_new_c_quizzes_for_test = 10
+
+######################################################################
+
for n_epoch in range(args.nb_epochs):
log_string(f"--- epoch {n_epoch} ----------------------------------------")
cta = " ".join([f"{float(m.main_test_accuracy):.04f}" for m in models])
log_string(f"current_test_accuracies {cta}")
+ ##################################################
# Select, improve, and eval the worst model
weakest_model = min(models, key=lambda m: float(m.main_test_accuracy))
f"test_set_composition w_quizzes {quiz_machine.nb_batch_w_quizzes} c_quizzes {quiz_machine.nb_batch_c_quizzes}"
)
+ ##################################################
# Replace a fraction of the w_quizzes with fresh ones
quiz_machine.renew_w_quizzes(args.nb_train_samples // args.nb_gpts)
+ ##################################################
# If all the models are good enough, generate new quizzes and
# re-compute the test errors