zafiirah13 / frbid Goto Github PK

This is really minor, but might become an issue if someone else tries to use this code. Currently training, validation and test examples all use only class 1 data. This is not a problem with training and validation as it is just an example. With test data however it fails with the error

raise ValueError("Only one class present in y_true. ROC AUC score "

Easiest thing to do would be to just change the label of one of the existing HDF5 files from 1 to 0 in the test_set.csv file. That way the whole training process can be run on the example data.

Currently TensorFlow tries to use all of the GPU memory. That causes problems if there are other processes running that use GPU memory (pipeline, plotting, anything else...). This causes the FRBID process to segfault and also crashes other processes.

The solution is to limit the amount of GPU memory TF can use at once, something along the line of 10-15%.