Training Deep Time Series Classification/Anomaly Models

Training time series classification models is a relatively new feature in FF. It follows the same general approach as forecasting models but with a few subtle differences. See our full tutorial notebook on Kaggle here.

File Format

For our GeneralClassificationLoader class files should be pre-formated as follows. Future data-loaders may feature additional

Datetime	MeanMilkProduction (kg/lactation)	Temperature	cow_class

Datetime	MeanMilkProduction (kg/lactation)	Temperature	cow_class
12/24/20	95	55
12/25/20	190	45
12/26/20	165	42
12/27/20		8552
	53	10529
	54	9924	2
	92	10533
	115	9167
	180	9289
	97	8343
	84	9005
	82	8867	4
	80	9352
	77	8900

Required Parameters

In general parameters are similar to the parameters you would supply models for TS forecasting

sequence_length: The length of the time series sequence you wish to classify. At the moment GeneralClassificationLoader solely supports sequences if the same length. If you have sequences of varying lengths we recommended padding the sequences.

Example Notebook

You can find an end-to-end example at the link below.

Time Series Classification with Flow Forecast (FF)

Design and Implementation Notes

Check List

Implement and test GeneralClassification Data Loader

Add calculation of additional metrics

Precision

F1

Revised plotting to Weights and Biases

Logging of ROC plots

Refactoring and cleaning up plot functions in trainer.py

End to end integration test

Example classification Kaggle notebook

Questions?

Can existing plot functions work with classification tasks?

Is a test-loader really necessary? Why would we want to plot the results in this case…

Additional Changes

There has been some odd behavior with the ROC curves so we may want to look into that 12/21/22