robolamp / KGPUComputeDocker

Docker image for Deep Learning used at @koto-bank Contains Tensorflow, Keras, PyTorch, OpenCV and a few Python ML libraries.

KGPUComputeDocker

Docker image for Deep Learning. Based on floydhub/dl-docker.

What's inside

• Ubuntu 16.04
• CUDA 8.0
• cuDNN v6
• Tensorflow
• Keras
• iPython/Jupyter Notebook with Python3 kernel by default
• Numpy
• SciPy
• Pandas
• Scikit Learn
• Matplotlib
• PyTorch
• OpenCV

TODO:

• Some Python libraries for NLP

How-to

Prerequisites

0. Some Linux distro (compatibility checked with Gentoo and Ubuntu), NVIDIA GPU
1. Docker https://docs.docker.com/engine/installation/
2. NVIDIA drivers https://www.nvidia.com/Download/index.aspx?lang=en-us
3. NVIDIA-Docker https://github.com/NVIDIA/nvidia-docker

Obtaining the Docker image

You can download the Docker image from Docker Hub, or build an image locally.

Download the Docker image from Docker Hub

You can pull an image from Docker Hub repository.

docker pull robolamp/k-gpu-compute-docker:gpu

Build the Docker image locally

1. Clone this repository

git clone https://github.com/robolamp/KGPUComputeDocker.git
cd KGPUComputeDocker

2. Check Docker's Base Device Size parameter using this guide

We need at least 16 GB of memory

3. Build the image

You can build the image which includes buidling TensorFlow and OpenCV from sources:

docker build -t robolamp/k-gpu-compute-docker:gpu -f Dockerfile.gpu .

But it will take a lot of time. Alternatively, you can build the image which is using pre-built packages:

docker build -t robolamp/k-gpu-compute-docker:gpu -f Dockerfile.bin.gpu .

Running the Docker image

Don't forget to prepare a directory to share with container using Docker volumes

You can run container with an access to container's bash:

nvidia-docker run -it -p 4444:8888 -p 3003:6006 -v /sharedfolder:/root/sharedfolder robolamp/k-gpu-compute-docker:gpu bash

Alternatively, you can run jupiter notebook in container.

1. Prepare the password for your jupiter notebook

Run python in your terminal:

python

Encrypt your password using notebook.auth.security.passwd():

>>> from notebook.auth import passwd
>>> passwd()
Enter password: 
Verify password: 

2. Launch container with notebook

Copy your password hash and insert it into the following command and run it:

nvidia-docker run -d -p 4444:8888 -p 3003:6006 -v /sharedfolder:/root/sharedfolder robolamp/k-gpu-compute-docker:gpu jupyter notebook --NotebookApp.password='your_password_hash' --allow-root