Tensorflow GPU MLflow Jupyter Dockerfile

A Dockerfile defining a Docker image containing CUDA, Tensorflow 2 (GPU), Miniconda, MLflow and Jupyter Note.

About

This Dockerfile borrows from various Dockerfiles (including some of the official Tensorflow Dockerfiles). Recommended for local development but not production.

Requirements

1. Install Docker.
2. Make sure you have an NVIDIA GPU and install nvidia-docker.

Building

docker build -t <image_tag> .

Where <image_tag> is the tag you choose for this image.

Running

docker run -it -p 8888:8888 -p 5000:5000 --runtime=nvidia --name <container_name> <image_tag>

Where <container_name> is the name for your container and <image_tag> is the tag you choose for your image.

Mapping volumes

You can also map volumes to /mlflow/mlruns, /mlflow/projects and /mlflow/notebooks, e.g.

-v /your/home/mlruns:/mlflow/mlruns

Interfaces

Jupyter Notebook

When you run the container you'll see a web address in the logs similar to this:

http://127.0.0.1:8888/?token=XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX

Click it and the Jupyter Notebook interface will open in your browser.

MLflow UI

You can access the MLflow though http://localhost:5000.

Notes

In your notebooks, make sure you include this code snippet so you'll see your experiments in the MLflow UI:

from mlflow.tracking import set_tracking_uri as uri
uri('http://localhost:5000')