TFDST 0.0.1

Fast Discrete Shearlet Transform layers in TensorFlow

TFDST: Fast Discrete Shearlet Transform Layers in TensorFlow

TFDST provides differentiable TensorFlow/Keras layers for fast 2D and 3D discrete shearlet transforms and their inverse transforms.

Copyright 2025 Kishore Kumar Tarafdar. Filter-bank construction follows mathematical structure developed with credit to Vineet Ghule. This implementation provides differentiable TensorFlow/Keras layers with batched multichannel support and performance-oriented updates.

Capabilities

• DST2D / IDST2D through transform=None or transform="inverse"
• DST3D / IDST3D through transform=None or transform="inverse"
• Batched multichannel TensorFlow tensors
• Backpropagation through the transform layers

Limitations

• Inputs are expected to match the configured spatial size N.
• 2D input shape: [batch, N, N, channels].
• 3D input shape: [batch, N, N, N, channels].
• Wavelet filter coefficients are provided through PyWavelets.

Installation

pip install TFDST

Minimal Example

import tensorflow as tf
from TFDST.DST2DFB import DST2D
from TFDST.DST3DFB import DST3D

N = 32

# 2D
x2 = tf.random.normal([1, N, N, 1])
dst2 = DST2D(N=N, J=2, L=[1, 2], B=[4, 8], norm=True, wave="bior1.5")
idst2 = DST2D(N=N, J=2, L=[1, 2], B=[4, 8], norm=True, wave="bior1.5", transform="inverse")
y2 = dst2(x2)
r2 = idst2(y2)

# 3D
x3 = tf.random.normal([1, N, N, N, 1])
dst3 = DST3D(N=N, J=2, L=[1, 2], B=[4, 8], norm=True, wave="bior1.5")
idst3 = DST3D(N=N, J=2, L=[1, 2], B=[4, 8], norm=True, wave="bior1.5", transform="inverse")
y3 = dst3(x3)
r3 = idst3(y3)

print(x2.shape, r2.shape)
print(x3.shape, r3.shape)

Arguments

Argument	Meaning
N	Spatial size of each axis; 2D uses N x N, 3D uses N x N x N.
J	Number of multiscale shearlet levels.
L	Shear parameters per level; must have length J.
B	Bandwidth/window parameters per level; must have length J.
wave	PyWavelets wavelet name used for the radial wavelet filters, e.g. "bior1.5".
norm	If True, applies filter-bank normalization.
transform	Use None for forward DST and "inverse" for inverse DST.

License

Apache License 2.0. See LICENSE.