search

arfc / annsa

Artificial neural networks for spectral analysis.
BSD 3-Clause "New" or "Revised" License
12 stars 4 forks source link

Rewrite training functions to add re-binning as data augmentation #29

Closed kamuda1 closed 5 years ago

kamuda1 commented 6 years ago

Re-binning is currently implemented as fixed when the dataset is constructed. This is fine for the hyperparameter search, but it would be better to use rebinning as data augmentation when training the final networks.

We will know this new feature is an improvement when:

Compare performance of the data augmentation strategy (with similar ranges of data rebinning) to the previous method (the previous method uses different levels of courseness when sampling rebinning parameters for fixed dataset construction.) Use an F1 score as performance metric.

kamuda1 commented 5 years ago

Re-binning and other online data augmentation added with #47.