Each pattern should be classified within a scope. The scopes below define how patterns apply to TensorFlow. They can be modified freely, but should be as distilled as possible to avoid overlapping concerns.

Code convention

Code convention patterns apply to the TensorFlow code is structured and written.

Related scopes include Design convention and Function. Design conventions apply at a higher level than code conventions. Functions name specific types of functions that may be implemented rather than the conventions used by those functions.

Design convention

A design convention is a high level pattern that describes an approach, methodology, or philosophy that can be applied to TensorFlow development.

If a convention can be applied to source code, it consider placing it in the Code convention scope.


A function pattern is a specific type of function used in a TensorFlow script. There may be variations within a type of function, but the pattern must nonetheless apply to the naming and purpose of a module function.


Patterns that name and describe specific tensors or operations should be placed in this scope. The intent of this scope is to drive toward consistency in naming conventions where possible. Obviously each model presents its own naming conventions, but common themes should be named as Tensor/Operation patterns.


Workflow patterns describe steps in TensorFlow work. Workflow steps may be implemented as scripts, functions, parameterized behavior (e.g. driven by flag values) or may be outside a TensorFlow coding context entirely (e.g. setting up hardware, system libraries, etc.)