FFT Wavelets and more
Adds several sharing methods and their accompanying training implementations:
- FFT with frequency change based parameter selection (FrequencyAccumulator)
- Wavelet with frequency change based parameter selection (FrequencyWaveletAccumulator)
- topK with model change based parameter selection (ModelChangeAccumulator)
- TopKParams: selects the topK highest values for sharing
It also adds an example config file for each mentioned sharing method.
Additionally it adds:
- 96 nodes regular random graph with degree four
- plot.py now also json dumps the average train loss, test loss, and test loss
- changes run.sh template to store the logging data on the nfs
- adds
PyWavelets
to setup.cfg - In testing.py it will now crash if the logging directory already exists to prevent accidentally overwriting old experiments.
- converting indices to int32 before encoding
- removing not needed imports
Edited by Jeffrey Wigger