plasmacontrol / bes-edgeml-models Goto Github PK

@jeff-zimmerman It appears that the sample data file sample_labeled_elm_events.hdf5 is missing from the refactor branch, as best I can tell. Can you add it back, and can you investigate how the tests passed with the data file missing? Thanks.

@LakshyaMalhotra - When you get a chance, please explore the refactored multi_features_ds model in the branch drsmith. I forked the files train_ds.py and multi_features_ds_model.py from the versions in main. All FFT and DWT calculations are performed at runtime in the forward() call, so preprocessing and storing CWT is no longer necessary. The DWT is efficient like FFT and has minimal memory footprint. I commented out all logic for arguments --multi_features and --use_fft, including in trainer.py and base_arguments.py. I added arguments for the multi_features_ds model in base_arguments.py. Raw, FFT, and DWT features are on by default, and they are turned off by setting, for example, --dwt_num_filters 0. For a small problem size (5 ELM events, signal_window_size=128, and 16 features each for raw/fft/dwt), the FFT and DWT calculations had no impact on epoch elapsed time and ran without a problem on my 16 GB Mac.

The DWT calculations use pytorch_wavelets. It's not on conda-forge, so do git clone and pip install.

Is the conv kernel the full size of the input tensor for this feature model? Maybe this is set in the model instantiation.

Spatial dimensions can be 8 // maxpool_size to give flexibility for different maxpool sizes.