search

bthtsang / DeepClassifierNoveltyDetection

Code accompanying the paper "Deep Neural Network Classifier for Variable Stars with Novelty Detection Capability" (submitted)
MIT License
5 stars 2 forks source link

Trained model #1

Open fluowhy opened 4 years ago

fluowhy commented 4 years ago

Hi, do you have a trained model to test the algorithm? I would appreciate your help. Kind regards, Mauricio.

bthtsang commented 4 years ago

Thanks for your query. I just pushed a trained model to the repo, under /keras_logs/asassn/. It is a model trained with an old ASAS-SN dataset and the fiducial hyper-parameters reported in the paper. Please let me know if you need help with anything.

Benny

fluowhy commented 4 years ago

Thanks a lot for the trained model! I am trying to replicate your results and I wonder about the training data. As I navigate through the repository, it is not clear which data (ASAS-SN) was used for training. The last commit from asassn.pkl is about test data and I assume sample.pkl is only a fraction of the test data. I downloaded both and the class distribution did not match with the class distribution reported in the paper. Your work is the only one related explicitly about novelty detection in full light curves, so it is a good option to use it as a baseline. For that reason I am worried about the data. Thanks again, Mauricio.

bthtsang commented 4 years ago

Thanks for your interest in our work. The training dataset used in the paper is too large to be hosted on GitHub. Therefore, both asassn.pkl and sample.pkl are example datasets containing only small fractions of the light curves, primarily for demonstration purposes.

You are welcome to email me (btsang@kitp.ucsb.edu) so I can share with you the actual dataset we used in the paper. Note that the ASAS-SN team has since put out a few major updates on their variable star database (with many more objects, refinement in object classifications, etc.). Our old dataset do not include those new updates.