This docker is intended to be used as a design/test environment for deep learning projects. It is image/graphics/text oriented but it also contains many other libraries that I find useful. The highlights are:
- TensorFlow 2.5 : Backbone for DL with tensorflow-hub for transfer learning.
- Keras : For coding, including keras-tuner package for hyper-parameter tunning.
- Scikit : Always useful.
- pandas : Essential for data handling (and uninstalling Excel).
- pydot and scikit-multilearn : Useful for multi-label problems and graph analysis. The pydot plot function is working with cairo.
- BeautifulSoup4 : Along with html5lib for data extraction.
- OpenCV : Essential for image processing.
- imgaug : For image augmentation.
- pyntcloud : For point cloud analysis.
- pytesseract : A great OCR by Google. The image includes the python wrapper and the tesseract binary.
- nltk : Natural language processing toolkit.
- textblob : Other language procesing tool.
The image is built with pipeline development support. As of writing this text TFX 1.0 is not flexible enough for me (see commits), so I decided to use a more versatile solution, Airflow alone (and a lot of custom code...). Thus, this image contains support for using such pipelines with:
- apache-airflow [2.1.1]: Pipeline...
- certifi [2019.11.28]
- Flask[1.1.4]
- SQLAlchemy[1.3.24]
- redis[3.5.3]
- tensorflow-serving-api [2.5.1]: For serving the models in production.
- Ubuntu 20.04
- CUDA 11.2.2
- CUDNN 8.1.1
- Python 3.8.10
docker build -t datawhale:tf-2.5 .
docker run -u $UID:$UID --gpus all -v <path to code folder>:/home/datawhale/code -v <path to datasets folder (optional)>:/home/datawhale/datasets -p 8888:8888 -p 6006:6006 -it datawhale:tf-2.5
The first time run:
docker-compose up --build create_user
This will create a user for you, with credentials airflow:airflow. When this is done, all you need to do is call:
docker-compose up --build
Happy codding!