Simple framework of typical data app actions (data readers, transforms, inference, data writers) for Python.
Install pytoch locally
pip3 install torch torchvision torchaudio
Install some extra libs
pip3 install scipy numpy pandas pyarrow
Extra libs for common data transforms
pip3 install jmespath
Install dltHub
pip3 install dlt
Install additional readers for Action development
pip3 install pipe langchain
Install Huggingface
pip3 install transformers accelerate
Go to the repository directory
cd actions
Set Huggingface token if you intend to use examples that use Huggingface API
export HF_TOKEN=<your-hf-token>
This example can be thought of as a bot proposing an answer to the thread of issue comments on GitHub.
Uses: HuggingFace model API, local Pytorch, local meta-llama/Meta-Llama-3-8B-Instruct inference, Apple MPS device (can be switched to cpu)
Input: A particular GitHub issue (#933 of the dlthub repo) will be downloaded with all its comments
Output: Files will be written to the "out" directory, containing the proposed response and the full message chat
python3 examples/github_bot.py
The following example runs every image file in data/flowers-102-100 folder through Imagenet Labeler.
Uses: local PyTorch, Apple "mps" GPU device (can run with cpu as well).
Input: image files in data/flowers-102-100 folder
Output: Standard Output will contain probabilities of image classes e.g. 'bee[0.564373791217804], daisy[0.07953538745641708]'
python3 examples/pytorch_image_labeler.py
If you get an error when executing one of the examples
ModuleNotFoundError: No module named 'core'
just add to the Python Path
export PYTHONPATH=.:$PYTHONPATH
If you are trying to do local Pytorch inference on Apple silicone and get an error when executing one of the examples
NotImplementedError: The operator 'aten::isin.Tensor_Tensor_out' is not currently implemented for the MPS device.
just run this before running the example script
export PYTORCH_ENABLE_MPS_FALLBACK=1