keras-team /

Keras

Live map
View repository
Snapshotmaster · afcca41

How backend-neutral layers become a model trained on TensorFlow, JAX, or PyTorch and saved portably.

Click a building or route
Choose a path

Build and compile a model, adapt input data, then execute backend-native gradient steps.

System map7 components · 7 connections
130%
Keras architectureLayers and symbolic operations build the model at the upper left; shared data and training logic sit in the center; backend execution, distribution, and saving branch to the right.symbolic tensors, vari…optimizer, losses, met…epoch iterator batchestrain, test, or predic…input batch and traini…Layers andoperationsModel graphBackendtrainerShared TrainerPortablesavingData adaptersDistribution
Connections
BuildingsEntryServiceComputeStorageRuntimeToolingExternal

Terms, in plain English

KerasTensorA symbolic tensor carrying shape and dtype through functional graph construction.

How to study this repo

  1. 1
    Build a symbolic graph

    Follow a Layer call through Operation symbolic handling into the Model's graph nodes.

  2. 2
    Compare backend steps

    Read the TensorFlow, JAX, and PyTorch trainer implementations for the same compiled model lifecycle.