tfdeploy

Allows running trained models without TensorFlow installed, relying solely on NumPy, which simplifies deployment in constrained environments like serverless or embedded systems.

Core code is under 150 lines in a single file, making it easy to copy and integrate into projects without heavy dependencies, as highlighted in the README.

Often faster than TensorFlow's native Tensor.eval for many graphs, demonstrated in performance comparisons showing significant speedups for vectorized operations.

Supports writing custom NumPy-based operations by inheriting from tfdeploy.Operation, enabling flexibility for unsupported or specialized ops.

Provides an Ensemble class to combine multiple models with methods like mean, max, min, or custom logic, useful for advanced inference scenarios.