tensorflow-wavelets is an implementation of Custom Layers for Neural Networks:

• Discrete Wavelets Transform Layer
• Duel Tree Complex Wavelets Transform Layer
• Multi Wavelets Transform Layer

git clone https://github.com/Timorleiderman/tensorflow-wavelets.git
cd tensorflow-wavelets
pip install -r requirements.txt

Installation

tested with python 3.8

pip install tensorflow-wavelets

Usage

from tensorflow import keras
import tensorflow_wavelets.Layers.DWT as DWT
import tensorflow_wavelets.Layers.DTCWT as DTCWT
import tensorflow_wavelets.Layers.DMWT as DMWT

# Custom Activation function Layer
import tensorflow_wavelets.Layers.Threshold as Threshold

Examples

DWT(name="haar", concat=0)

"name" can be found in pywt.wavelist(family)

concat = 0 means to split to 4 smaller layers

from tensorflow import keras
model = keras.Sequential()
model.add(keras.Input(shape=(28, 28, 1)))
model.add(DWT.DWT(name="haar",concat=0))
model.add(keras.layers.Flatten())
model.add(keras.layers.Dense(nb_classes, activation="softmax"))
model.summary()

_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
dwt_9_haar (DWT)             (None, 14, 14, 4)         0
_________________________________________________________________
flatten_9 (Flatten)          (None, 784)               0
_________________________________________________________________
dense_9 (Dense)              (None, 10)                7850
=================================================================
Total params: 7,850
Trainable params: 7,850
Non-trainable params: 0
_________________________________________________________________

name = "db4" concat = 1

model = keras.Sequential()
model.add(keras.layers.InputLayer(input_shape=(28, 28, 1)))
model.add(DWT.DWT(name="db4", concat=1))
model.summary()

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
dwt_db4 (DWT)                (None, 34, 34, 1)         0
=================================================================
Total params: 0
Trainable params: 0
Non-trainable params: 0
_________________________________________________________________

DMWT

functional example with Sure Threshold

x_inp = keras.layers.Input(shape=(512, 512, 1))
x = DMWT.DMWT("ghm")(x_inp)
x = Threshold.Threshold(algo='sure', mode='hard')(x) # use "soft" or "hard"
x = DMWT.IDMWT("ghm")(x)
model = keras.models.Model(x_inp, x, name="MyModel")
model.summary()

Model: "MyModel"
_________________________________________________________________
Layer (type)                 Output Shape              Param #
=================================================================
input_1 (InputLayer)         [(None, 512, 512, 1)]     0
_________________________________________________________________
dmwt (DMWT)                  (None, 1024, 1024, 1)     0
_________________________________________________________________
sure_threshold (SureThreshol (None, 1024, 1024, 1)     0
_________________________________________________________________
idmwt (IDMWT)                (None, 512, 512, 1)       0
=================================================================
Total params: 0
Trainable params: 0
Non-trainable params: 0
_________________________________________________________________

PyPi upload:

pip install --upgrade build
pip install --upgrade twine
python -m build
python -m twine upload --repository pypi dist/*

If our open source codes are helpful for your research, please cite our technical report:

@Article{e26100836,
AUTHOR = {Leiderman, Timor and Ben Ezra, Yosef},
TITLE = {Information Bottleneck Driven Deep Video Compression—IBOpenDVCW},
JOURNAL = {Entropy},
VOLUME = {26},
YEAR = {2024},
NUMBER = {10},
ARTICLE-NUMBER = {836},
URL = {https://www.mdpi.com/1099-4300/26/10/836},
ISSN = {1099-4300},
DOI = {10.3390/e26100836}
}

Free Software, Hell Yeah!