Butterfly-Net2

The code is for the paper:

Butterfly-Net2: Simplified Butterfly-Net and Fourier Transform Initialization

Environment

All code was developed on Windows 10 and tested on CentOS 7 with Python 3.6, and was implemented by Tensorflow 1.13

Experiment

all the experiments are done via code in the folder 'test', while other sources are in the folder 'src'.

4.1.1. APPROXIMATION POWER BEFORE TRAINING

testing: test_Fourier_beforetrain.py

plot: test_Fourier_beforetrain_plot.py

for other situations, just change the value of N, K, l, r.

4.1.2. APPROXIMATION POWER AFTER TRAINING

initialization: Butterfly_FTini.py

training & testing: test_Fourier_aftertrain.py

The four different networks and other hyper-parameters can be changed in paras.json.

4.1.3. TRANSFER LEARNING CAPABILITY

initialization: Butterfly_FTini.py

training: test_Fourier_trans_train.py

testing: test_Fourier_trans_eval.py

plot: test_Fourier_trans_plot.py

4.2.1. ENERGY OF LAPLACE OPERATOR

initialization: Butterfly_FTini.py

training & testing: test_Energy.py

4.2.2(3). END-TO-END ELLIPTIC PDE SOLVER

initialization: EtE_Butterfly_FTini.py & EtE_PDE_setgen.py

training & testing: test_EtE_PDE.py

plot: test_EtE_PDE_plot.py

4.3. Denoising and Deblurring of 1D Signals

initialization: EtE_Butterfly_FTini.py

training & testing: test_denoise.py & test_deblur.py

plot: test_denoise_plot.py & test_deblur_plot.py