pyTorchAutoForge 0.5.0

PyTorchAutoForge library is based on raw PyTorch and designed to automate DNN development, model tracking and deployment, tightly integrated with MLflow and Optuna. It supports Spiking networks libraries (WIP). Deployment can be performed using ONNX, pyTorch facilities or TensorRT (WIP). The library is designed to be compatible with Jetson Orin Nano Jetpack rev6.1, with bash script to automatically configure virtualenv.

pyTorchAutoForge

WARNING: work in progress. Do not hesitate to open issues for improvements or problems!

A library based on PyTorch (https://pytorch.org/) and designed to automate ML models development, tracking and deployment, integrated with MLflow and Optuna (https://mlflow.org/, https://optuna.org/). It also supports spiking networks libraries (WIP). Model optimization and deployment can be performed using ONNx, pyTorch facilities or TensorRT (WIP). The library aims to be compatible with Jetson Orin Nano Jetpack rev6.1. Several other functionalities and utilities for sklearn and pySR (https://github.com/MilesCranmer/PySR) are included (see README and documentation).

Some brief usage guides (WIP)

TensorRT exporter quick usage

from pyTorchAutoForge.api.tensorrt import TRTengineExporter

exporter = TRTengineExporter()
engine_path = exporter.build_engine_from_onnx_path(
    onnx_model_path="/tmp/model.onnx",
    output_engine_path="/tmp/model.engine",
)
print(engine_path)

from pyTorchAutoForge.api.tensorrt import TRTengineExporter

exporter = TRTengineExporter()
engine_path = exporter.export_torch_to_trt_engine(
    torch_model=model,                    # torch.nn.Module
    input_sample=sample,                  # torch.Tensor or tuple[torch.Tensor, ...]
    onnx_export_path="/tmp/intermediate.onnx",
    output_engine_path="/tmp/final.engine",
)
print(engine_path)

Notes:

• TRTEXEC mode requires trtexec in PATH.
• PYTHON mode requires the tensorrt Python package.
• Default behavior avoids architecture-specific flags and is suitable for Jetson deployment workflows.

Installation using pip

The suggested installation method is through pip as the others are mostly intended for development and may not be completely up-to-date with the newest release versions. In whatever conda or virtual environment you like (preferably with a sufficiently new torch release, to install from pypi:

pip install pyTorchAutoForge

Or from a local copy of the repository (requires hatch module for the build):

cd pyTorchAutoforge
pip install .

An automatic installation script conda_install.sh is provided and should work in most cases. Note that it will automatically create a new environment named autoforge and makes several assumptions about your environment. Dependencies for the core modules should be installed automatically using pip. However, this is currently not fully tested. Please open related issues.