I am not sure if a pretrained machine learning model is actually protected by copyrights or not. Copyright protection exists to protect the creators of creative works from having their work "stolen", and I am not sure if training a ML model is an act of creativity.
That said, assuming that a pretrained ML model is actually protected by copyrights, then it is more likely that the model is a derived work of the data set used for training than that it is a derived work of the software that uses the model.
The software reads the model in as data, assuming that the software can be used with many differently trained models. In that case, the software and the model are considered completely independent works in the same way that MS Word and the documents you write with it are independent works for copyright.
Thus, if you want to publish the trained model with a license, I would recommend to use the BSD license that was also used for the training set.