Docker-Keras

This is a shell project that builds an NVIDIA GPU-capable Python environment with Anaconda and Keras (with Tensorflow GPU) using Ubuntu Linux 18.04. One can then launch a Jupyter Notebook Server or a standard Python project using the dockerized environment.

Host Machine Setup

The host machine needs to have

• NVIDIA GPU Driver
• docker-nvidia2

Make sure to upgrade docker and docker-compose to a fairly recent version.

Install NVIDIA GPU Driver

Follow the instructions described in http://www.linuxandubuntu.com/home/how-to-install-latest-nvidia-drivers-in-linux.

Install docker-nvidia2

Again, follow the instructions from: https://github.com/NVIDIA/nvidia-docker

Check that the NVIDIA GPU drivers are working by invoking the nvidia-smi utility from the host machine.

(base) deep@deep-TUF-GAMING-FX504GM-FX80GM:~/PycharmProjects/docker-keras$ nvidia-smi
Mon Apr 22 21:54:29 2019       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 418.56       Driver Version: 418.56       CUDA Version: 10.1     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX 1060    Off  | 00000000:01:00.0 Off |                  N/A |
| N/A   57C    P0    25W /  N/A |    243MiB /  3019MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|    0      2195      G   /usr/lib/xorg/Xorg                           137MiB |
|    0      2378      G   /usr/bin/gnome-shell                          85MiB |
|    0      3536      G   ...charm-community-2019.1.1/jre64/bin/java    17MiB |
+-----------------------------------------------------------------------------+
(base) deep@deep-TUF-GAMING-FX504GM-FX80GM:~/PycharmProjects/docker-keras$ 

As a bonus, you now know which GPU laptop I'm using ;)

Next, pull an existing NVIDIA cuda containers and execute the same command. This verifies that docker-nvidia is working correctly.

 docker run --runtime=nvidia --rm nvidia/cuda nvidia-smi

You should see an output similar to above.

Install Anaconda, Keras and TensorFlow in Container

Build the keras-jupyter image using the docker-compose.yml in this project. It extends the official NVIDIA Cuda 10.1 runtime image (with CUDA and CUDNN pre-installed) by installing Anaconda3 and then adds the GPU builds of TensorFlow and Keras. It also maps the ~/notebooks directory on the host machine to /home/ubuntu/notebooks in the image so that you can save your notebooks on the host.

First launch a container based off of this image:

$ docker-compose run -p 8888:8888 keras-jupyter

Once inside the container, bring up a notebook session.

$ jupyter-notebook --ip 0.0.0.0 --port 8888 --allow-root --no-browser &

And you're ready to kick off your GPU based Machine Learning Project! Navigate to http://127.0.0.1:8888 to access the Jupyter server. The gpu_check.ipynb notebook included in this project is an example of checking TF device placement to ensure that TensorFlow is really using the GPU.