Model Training (Variational Inference)
Last updated
Last updated
Last Updated: 2023-09-27
TimberID includes many techniques for generating isoscapes from geospatial point-measurements, ranging from Kriging interpolation to our most best-performing variational inference deep learning model. Of all the alternatives we evaluated, our best results come from predicting the isotope distribution with variational inference and a deep neural network, which appears to be a novel technique for isoscape generation. We focus here on producing isoscapes with variational inference, but the same general principles apply to most of our isoscape models.
Find the Colab at: https://colab.research.google.com/github/tnc-br/ddf-isoscapes/blob/main/dnn/variational_inference.ipynb
Most users should obtain somewhat reasonable predictions out-of-the-box by running the training notebook with the default configuration options after running the data ingestion notebook with its default options and export path. However, our machine learning models are still one-off artifacts since we are rapidly iterating on model research. The TimberID processes for producing production-quality models are not automatic. For now, this notebook only attempts to provide a reasonable baseline as a place to start for running model experiments.
A few important, high-level notes:
Some overfitting is normal, but training loss should not significantly dip below validation loss. Example of a typical loss curve:
If loss drops significantly below val_loss, the model has most likely memorized its training set and is unlikely to produce a good isoscape. Common reasons for this are: not enough data, a training distribution that is not diverse enough, training for too many epohcs (overtraining). To mitigate overtraining, the notebook comes with early stoping and a default patience setting of 30 epochs.
For reproducible results, be sure to re-run the cell that contains tf.keras.utils.set_random_seed(18731) before each training run. This model is very small, which makes it highly sensitive to initialization.
We use Xavier initialization but have not performed robust experiments to compare alternatives other than no initialization at all.
The code under "Run this cell to generate an isoscape" does not appear to be fully functional at this time. For generating isoscapes, prefer the Isoscape Generation notebook instead.
