Skip to content

API Overview

This section documents the public API surface of KeyDNN.

All APIs listed here are part of KeyDNN’s presentation layer and are considered stable for user-facing development. Internal implementation details are intentionally hidden and may change without notice.

Public API Scope

KeyDNN follows a layered architecture (PIAD: Presentation → Infrastructure → Application → Domain).

  • Presentation layer
    Public, documented, and stable. Safe for direct use.

  • Domain / Infrastructure layers
    Internal. Subject to change. Not documented here.

If an object or function is documented in this section, it is safe to depend on. If it is not documented, it should be considered internal.

API Categories

The public API is organized into the following categories:

  • Tensors
    Core data structures for numerical computation and automatic differentiation.

  • Layers
    Neural network building blocks such as dense layers, convolutions, normalization, pooling, and regularization layers.

  • Activations
    Common activation functions implemented as callable layers.

  • Losses
    Standard loss functions used for training neural networks.

  • Optimizers
    Algorithms for updating model parameters based on gradients.

  • Models
    High-level abstractions for composing layers and managing training workflows.

  • Callbacks
    Hook-based utilities for extending training behavior (e.g., early stopping, checkpointing).

  • Datasets
    Convenience loaders for commonly used benchmark datasets.

  • Backend
    Utilities for querying system capabilities such as CUDA availability.

  • Utilities
    Helper functions for determinism, randomness control, and preprocessing.

Each category has its own dedicated reference page.

Import Conventions

All public APIs are re-exported at the top level of the keydnn package.

Recommended import style:

from keydnn import Tensor, Sequential, Adam