search

ThrunGroup / implicit-hyper-opt

MIT License
1 stars 0 forks source link

1 cause of discrepancy #4

Open motiwari opened 2 years ago

motiwari commented 2 years ago

I've discovered at least one cause of discrepancy between our master and Lorraine's: https://github.com/ThrunGroup/implicit-hyper-opt/blob/master/data_loaders.py#L75-L81

Somehow, when we changed train_data to data and train_labels to targets, this results in different behavior.

Additionally, it's not clear to me why num_train is used for the val set as well

I'm also not sure why changing those field names leads to different behavior

motiwari commented 2 years ago

Oh, the num_train is used to split the train and validation set properly. Still don't know about the other questions

motiwari commented 2 years ago

I suspect that the difference in train_data, train_labels and data, targets is due to a difference in random sampling for the different attributes. It is probably not material since both