What's "We validated our configuration on a custom split of 11 and 4 clouds" meaning？

[GSusan]

hi, sorry to disturb.
During my training on Semantic3D, I couldn't clearly understand the difference between "To train on the whole publicly available data and test on the reduced test set, run" and "We validated our configuration on a custom split of 11 and 4 clouds" in Readme.

1. what's datasets ‘test' for when training?
2. when training with train and testred, I set --db_train_name trainval, is that mean training at the same time validating with val; for the same reason, setting --db_train_name train means training withou validating?
   thanks

[Loic Landrieu]

We tuned the hyperparameters by training on 11 clouds (train) and sdelecting the best performing configuration on a custom validation set (val), see get_datasets in learning/sema3d_dataset.py. We then trained the best configuration on the entire train set (trainval = train + val).

For semantic3d the test labels are withheld so it is the only way to tune hyperparameters.

[GSusan]

Thanks for your reply.
But what maks me confused is that, when I'm training with trainval, test dataset has no labels. Though predictions_testred(labels) are withheld, the Test accuracy is just about 1.32, Test oAcc=0.0, and so on.
Thanks again for your patience.