This project is still in its infancy, but I want to make the discrete wavelet transform accessible in keras machine learning projects. I'm a lone developer right now, and cannot commit to regular releases or implementing all wavelet functions. I intend to build out more features, examples, and tests.
You can create a model layer similar to other standard keras layers.
>>> from tensorflow.keras import layers as L
>>> from tensorflow.keras import models as M
>>> import dwt
>>>
>>> inp = L.Input((1024, 3))
>>> cA, cD = dwt.DWTPooling1D('db2')(inp)
>>> x = L.Concatenate(axis=-1)([cA, cD])
>>> x = L.Conv1D(3, 5, activation='relu')(x)
>>> model = M.Model(inp, x)
>>> model.summary()
Model: "model"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) [(None, 1024, 3)] 0
__________________________________________________________________________________________________
dwt_pooling1d (DWTPooling1D) [(None, 512, 3), (No 0 input_1[0][0]
__________________________________________________________________________________________________
concatenate (Concatenate) (None, 512, 6) 0 dwt_pooling1d[0][0]
dwt_pooling1d[0][1]
__________________________________________________________________________________________________
conv1d (Conv1D) (None, 508, 3) 93 concatenate[0][0]
==================================================================================================
Total params: 93
Trainable params: 93
Non-trainable params: 0
__________________________________________________________________________________________________Currently only the orthogonal wavelets work, but I plan on adding support for the biorthogonal wavelets.
Working wavelets include:
- Haar (haar)
- Coiflets (coif1-coif17)
- Daubechies (db1-db38)
- Symlets (sym2-sym20)
- Discrete Meyer (dmey)
- Refactor and rename to be more similar to tensorflow layers
- Organize repo / module more effectively
- Add documentatation
- Add more unit tests
- Support the other pywt wavelet families
- tensorflow
- numpy
- pywt
The pywt webpage requests citing them in scientific publications, please do so if you use this repo which relies on their wavelet generating functions.