tfmiss

The missing OPs, layer & etc. for TensorFlow

Development

Environment

Install all dependencies including python headers. Do not use pyenv on MacOS X, otherwise tests mostly likely will fail.

Build PIP package manually

You can build the pip package with Bazel v0.25.3:

# GPU support
export TF_NEED_CUDA="1"

# Set these if the below defaults are different on your system
export TF_CUDA_VERSION="12.6"
export TF_CUDNN_VERSION="9"
export CUDA_TOOLKIT_PATH="/usr/local/cuda"
export CUDNN_INSTALL_PATH="/usr/lib/x86_64-linux-gnu"

./configure.py
bazel clean --expunge
bazel test --test_output=errors //tfmiss/...
bazel build build_pip_pkg
bazel-bin/build_pip_pkg wheels

Build release with Linux docker container

# Requires about 4Gb of RAM allocated to Docker
DOCKER_BUILDKIT=1 docker build -t miss --output type=local,dest=wheels --build-arg PY_VERSION=3.8 ./

Install and test PIP package

Once the pip package has been built, you can install it with:

pip install wheels/*.whl

Now you can test out the pip package:

cd /
python -c "import tensorflow as tf;import tfmiss as tfm;print(tfm.text.zero_digits('123').numpy())"

You should see the op zeroed out all nonzero digits in string "123":

000