TFDWT: Fast Discrete Wavelet Transform TensorFlow Layers

Fast $1\text{D}$, $2\text{D}$ and $3\text{D}$ Discrete Wavelet Transform ($\text{DWT}$) and Inverse Discrete Wavelet Transform ($\text{IDWT}$) layers for backpropagation networks.

Available wavelet families ―

        Haar (haar)
        Daubechies (db)
        Symlets (sym)
        Coiflets (coif)
        Biorthogonal (bior)
        Reverse biorthogonal (rbio)

Note ― The permissible input shapes are $2^m$, where $m$ is a Natural number. Additionally, for 2D and 3D transforms, the shape of the input must be square and cube, i.e., equal sides.

---

Installation guide

The installation of the TFDWT package is recommended inside a virtual environment with tensorflow[and-cuda] installed at first.

Pre-installation checks (Tested in Gentoo and Debian bookworm)

• Create a new virtual enironment with a specific Python version (use the Python version supported by TensorFlow)

conda info --envs
env_name='tf219'
conda create -n $env_name python=3.12 ipykernel

# activate virtual environment
conda activate tf219

• Install TensorFlow using official guide in https://www.tensorflow.org/install/pip

# For GPU users
pip install tensorflow[and-cuda]
# Verify setup
python3 -c "import tensorflow as tf; print(f'{tf.config.list_physical_devices('GPU')}, \nTF version {tf.__version__}')"

Install TFDWT from PyPI (Option $1$)

pip install TFDWT

Install TFDWT from Github (Option $2$)

Download the package

git clone https://github.com/kkt-ee/TFDWT.git

Change directory to the downloaded TFDWT 

cd TFDWT

Run the following command to install the TFDWT package

pip install .

---

Verify installation

Compute $\text{DWT}$ $1\text{D}$ and $\text{IDWT}$ $1\text{D}$ of batched, multichannel $x$ of shape $(\text{batch, length, channels})$

"""Perfect Reconstruction 1D DWT level-1 Filter bank"""
from TFDWT.DWTIDWT1Dv1 import DWT1D, IDWT1D

LH = DWT1D(wave='bior3.1')(x)       # Analysis
x_hat = IDWT1D(wave='bior3.1')(LH)  # Synthesis

Compute $\text{DWT}$ $2\text{D}$ and $\text{IDWT}$ $2\text{D}$ of batched, multichannel $x$ of shape $(\text{batch, height, width, channels})$

"""Perfect Reconstruction 2D DWT level-1 Filter bank"""
from TFDWT.DWTIDWT2Dv1 import DWT2D, IDWT2D

LLLHHLHH = DWT2D(wave=wave)(x)      # Analysis
x_hat = IDWT2D(wave=wave)(LLLHHLHH) # Synthesis

Compute $\text{DWT}$ $3\text{D}$ and $\text{IDWT}$ $3\text{D}$ of batched, multichannel $x$ of shape $(\text{batch, height, width, depth, channels})$

"""Perfect Reconstruction 3D DWT level-1 Filter bank"""
from TFDWT.DWTIDWT3Dv1 import DWT3D, IDWT3D

LLLLLHLHLLHHHLLHLHHHLHHH = DWT3D(wave=wave)(x)      # Analysis
x_hat = IDWT3D(wave=wave)(LLLLLHLHLLHHHLLHLHHHLHHH) # Synthesis

NOTE ― Using the above forward and inverse transforms the above $\text{DWT}$ and $\text{IDWT}$ layers can be used to construct multilevel $\text{DWT}$ filter banks and $\text{Wave Packet Transform}$ filter banks.

---

Package is tested with dependency versions

        Python 3.12.7
        TensorFlow 2.15 to 2.19
        Keras 2 and 3
        Numpy 2.0.2
        CUDA 12.5.1

---

Uninstall TFDWT

pip uninstall TFDWT

---

TFDWT (C) 2025 Kishore Kumar Tarafdar, Prime Minister's Research Fellow, EE, IIT Bombay, भारत 🇮🇳