IREE TensorFlow Importers

This project contains IREE frontends for importing various forms of TensorFlow formats.

Quick Development Setup

Pip install editable (recommend to do this in a virtual environment):

# All at once:
pip install -e python_projects/*

# Or one at a time:
pip install -e python_projects/iree_tflite
pip install -e python_projects/iree_tf

Test installed:

iree-import-tflite -h
iree-import-tf -h

Run test suite

You need to make sure that the iree compiler and runtime are on your PYTHONPATH. The easiest way to do this is to install wheels with pip. For development, the following should do it:

source ~/path/to/iree-build/.env && export PYTHONPATH

Run the test suite with:

pip install lit

# Just run the default tests:
lit -v test/

# Can also run vulkan tests with:
lit -v -D FEATURES=vulkan test/

# Can disable the default LLVM CPU tests:
lit -v -D DISABLE_FEATURES=llvmcpu -D FEATURES=vulkan test/

# Individual test directories, files or globs can be run individually.
lit -v $(find test -name '*softplus*')

Updating Tensorflow Importers in CI

CI uses Tensorflow importers to run integration tests and benchmarks. They might need an update in CI if you want new features or bugfixes from the frontends.

Tensorflow importers are wrappers which call Tensorflow Python API to do conversion. CI installs a pinned version of Tensorflow in the docker images. To bump the Tensorflow version, you need to:

1. Update the pinned version of Tensorflow in integrations/tensorflow/test/requirements.txt.
2. Follow Adding or Updating an Image to rebuild the frontends docker image and its descendants.

Here is the command to rebuild and update the docker images:

python3 build_tools/docker/manage_images.py --image frontends

To modify the import tools themselves, you can directly change their code in integrations/tensorflow/python_projects without updating the dockers.