tract-python 0.2.12

bindings of sonos/tract rust NN inference enging for python

tract-python

Project Archived: tract now maintains it's own Python package here.

tract inference engine bindings in Python via FFI. It support Neural Network inference from NNEF or ONNX.

Why

No need to compile tract or have cargo installed, fast install.

tract-cli is very feature-full but reloading a model each time you wish to do an inference is computationaly costy and slow.

Think onnxruntime except it support NNEF, and it is based on tract.

Install

Install using pip:

pip install tract_python

Usage

import tract_python

print(tract_python.TRACT_VERSION)
# "X.X.X"

tract_model = tract_python.TractModel.load_from_path(
  # This parameter can be an ONNX or NNEF filepath (in case of NNEF it can be a dir or a tgz)
  './tests/assets/test_simple_nnef/' # simple graph that mul input by 2
)
# .run take as argument names the name of input nodes in your neural network
results = tract_model.run(input_0=np.arange(6).reshape(1, 2, 3).astype(np.float32))
print(results)
#{'output_0': array([[[ 0.,  2.,  4.],
#       [ 6.,  8., 10.]]], dtype=float32)}

Status

This project is maintained with latest tract version.

Scope

Our personnal usecase is to be able to run +10M inferences with 'tract' engine. So loading/running NNEF or ONNX is sufficient.

We would be happy to support some others tract-cli features:

• computing: number of FMA operations
• computing: profiling infos

(Thought it would be better to extract from tract-cli a tract-profile crate first in original repo to avoid code duplicate) We do not have the bandwith to do more and welcome any contributor that would wish to add more features.