TileDB-ML is the repository that contains all machine learning oriented functionality TileDB supports. In this repo, we explain how someone can employ TileDB for machine learning oriented data management problems, and which are the next steps we have in mind. Here, we would firstly like to highlight our perspective on the relation of TileDB with general machine learning oriented data management problems and how TileDB engine could be the solution for efficiently storing any kind of machine learning data, i.e., from raw images, text, audio, time series and SAR to features and machine learning models. Before you proceed further, please take a quick look on our medium blog post, which targets to explain in great detail how TileDB addresses many machine learning data format requirements, overcoming the drawbacks of the other candidate formats, and take this opportunity to solicit feedback and contributions from the community.
As mentioned above, this repository contains all machine learning oriented functionality TileDB supports. Specifically, code that can (or will be able to):
-
Save machine learning models as TileDB arrays (At the moment we provide implementations for saving Tensorflow Keras, PyTorch and Scikit-Learn models.)
-
Load machine learning models from TileDB arrays.
-
Read features, in order to train machine learning models, from TileDB arrays directly to machine learning framework's data APIs. We already support the Tensorflow and PyTorch data APIs with the use of Python generators for Dense and Sparse TileDB arrays, and we are similarly working on Scikit-Learn Pipelines which will be out soon.
We provide some detailed notebook examples on how to save and load machine learning models as TileDB arrays (also on TileDB-Cloud) and explain why this is useful in order to create simple and flexible model registries with TileDB.
- Example for Tensorflow Keras Models
- Example for PyTorch Models
- Example for Scikit-Learn Models
- Example for Tensorflow Model on TileDB-Cloud
- Example for PyTorch Model on TileDB-Cloud
- Example for Scikit-Learn Model on TileDB-Cloud
We also provide detailed notebook examples on how to train Tensorflow and PyTorch models with the use of our Data APIs support for Dense TileDB arrays.
Finally, we also provide an End-To-End example on how to ingest data, train a PyTorch model and serve it with UDFs completely serverlessly on TileDB-Cloud.
TileDB-ML can be installed:
-
from source by cloning the Git repository:
git clone https://github.com/TileDB-Inc/TileDB-ML.git cd TileDB-ML # In case you want to install and check all frameworks. If you # use zsh replace .[full] with .\[full\] pip install -e .[full] # In case you want to install and check Tensorflow only. If you # use zsh replace .[tensorflow] with .\[tensorflow\] pip install -e .[tensorflow] # In case you want to install and check PyTorch only. If you # use zsh replace .[pytorch] with .\[pytorch\] pip install -e .[pytorch] # In case you want to install and check Scikit-Learn only. If you # use zsh replace .[sklearn] with .\[sklearn\] pip install -e .[sklearn] # In case you want to try any of the aforementioned machine learning framework # on TileDB-Cloud try one of the follwoing. pip install -e .[tensorflow_cloud] pip install -e .[pytorch_cloud] pip install -e .[sklearn_cloud] -
with pip from git:
pip install git+https://github.com/TileDB-Inc/TileDB-ML.git@master -
from PyPi:
pip install tiledb-ml
The above command will just install the basic dependency of tiledb-ml, hence tiledb.
In order to install the integration for a specific framework you need to use:
pip install tiledb-ml[pytorch] # e.g. For checking only the Pytorch integration
Checking all the supported frameworks you will need to use:
pip install tiledb-ml[full]
The above commands apply to bash shell in case you use zsh you will
need to escape the bracket character like the following for example:
pip install tiledb-ml\[pytorch\]
- You may run the test suite with:
python setup.py test
Here we would like to highlight that our current implementations are not optimal, and they don't support the aforementioned machine learning
frameworks 100%, e.g., serialization of model parts like numpy arrays, takes place with Pickle (which is far from optimal)
because of its Python Only nature and insecurity as described here.
We mainly show the universal data management ability of TileDB, and how elegantly applies in
machine learning data of any kind. Optimizations will follow as soon as possible.
In any case, note that the TileDB-ML repository is under development, and the API is subject to change.
We welcome all contributions! Please read the contributing guidelines before submitting pull requests.
The TileDB-ML package is Copyright 2018-2021 TileDB, Inc
MIT