- Using a External Model to Predict the low-pass filters and minimizing the loss function like an Autoencoder setup.
- Download the temporary dataset using wget and unzip to the
./datadirectory. - Use the visualization.py to visualize the produced filters etc..
# Initialization of Filter Prediction Model: model = FilterConv(in_channels = IN_CHANNELS, out_channels = OUT_CHANNELS) model.to(device = DEVICE) # Initialization of Autoencoder Model: data = torch.load(DATA_PATH) awt = DWT1d(filter_model = model) s # Training: awt.fit(X = data, batch_size = BATCH_SIZE, num_epochs = NUM_EPOCHS) name = f"{name_of_your_model}.pth" torch.save(awt, name)currently implemented for 1D, using transform1d.py
- Idea based on Paper Recoskie & Mann, 2018
- Implementation based on Yu-Group/adaptive-wavelets
@article{ha2021adaptive, title={Adaptive wavelet distillation from neural networks through interpretations}, author={Ha, Wooseok and Singh, Chandan and Lanusse, Francois and Upadhyayula, Srigokul and Yu, Bin}, journal={Advances in Neural Information Processing Systems}, volume={34}, year={2021} }