search

KyleAMoore / Redshift-CNN

ResNet model for galaxy redshift estimation
MIT License
3 stars 0 forks source link

Reorganize library by type of functionality #35

Open KyleAMoore opened 4 years ago

KyleAMoore commented 4 years ago

The next phase of this project will be to develop a more general python library for redshift analysis and estimation. To facilitate this, existing functionality should be reorganized into modules based upon functionality type rather than the existing separation by phase in the process-train-test pipeline.

umesh-timalsina commented 4 years ago

Are you thinking in terms of a general CNN module/package like https://github.com/IBM/pytorch-seq2seq that builds on top of existing code?

KyleAMoore commented 4 years ago

That is the intention, more or less, as I understand it. The goal would be a simple and usable framework for training and using redshift estimation models. Additional features for understanding and analyzing the models, the input images, and the predicted redshift scores would be added after reorganization and would be given their own issues.

The biggest reorganization change that I foresee is that the creation, training, and utilization of models would be most intuitively combined into a single model module rather than separated into pipeline component modules as they are now. To accomplish this, I plan to implement a RedshiftClassifier base class that pulls in most of the functionality previously found in the train, eval-model, and model-utils modules. The ResNet and Inception classes (and any future models) would then inherit from this base class and would only need to provide a model structure, with all behaviors shared between models.