K3 Addons: Advanced Multi-Backend Components
Specialized neural network layers, attention mechanisms (MobileViT, CBAM, ECA), and custom loss functions engineered for Keras 3.
K3 Addons: Extensions for Keras 3
Stack: Keras 3, PyTorch, TensorFlow, JAX, Python
Code: github.com/anas-rz/k3-addons
Overview
K3 Addons extends the core Keras 3 framework with cutting-edge deep learning layers, attention blocks, and advanced training objectives that maintain full compatibility across TensorFlow, PyTorch, and JAX.
Key Modules
- Attention Blocks: MobileViT Attention, CBAM (Convolutional Block Attention Module), Efficient Channel Attention (ECA), and Squeeze-and-Excitation.
- Specialized Layers: Parity-matched adaptive pooling layers (
AdaptiveAvgPool1D/2D/3D,AdaptiveMaxPool1D/2D/3D), DropPath / Stochastic Depth, and Spectral Normalization. - Loss Functions & Metrics: Generalized IoU (GIoU), Complete IoU (CIoU), Sigmoid Focal Cross-Entropy, and Triplet Margin Loss with online hard-negative mining.
- Parity & Tests: Cross-backend numerical parity suites verified on CPU and GPU against reference PyTorch and TensorFlow implementations.