huleg / ml-notebook

Dockerfile for multiple machine learning tools.

ml-notebook

This project is aimed at providing an accessible and reproducible environment for a variety of machine learning toolkits, with a focus on deep learning toolkits. Instead of asking you to follow a set of complex setup instructions, ml-notebook asks you to wait while a tested, pre-built image is installed.

The following tools are available inside the Ubuntu 14.04 image, with Jupyter as an interface:

General:

• matplotlib*
• scipy*
• numpy*
• pyOSC
• Jupyter

Machine Learning:

• gensim
• scikit-learn
• scipy*
• pandas*

Deep Learning:

• Caffe (with pycaffe)
• Chainer
• TensorFlow
• Theano (with Keras and Lasagne)
• Torch

*These are requirements of other libraries, but also interesting in their own right.

Usage

This is only tested this on OSX. Something similar should work on Linux, and possibly Windows.

1. Install Docker.
2. Clone this repository.
3. (Optional) Run ./update.sh Notice: this downloads almost 2GB of data and examples. If you just want to look around, browse ml-examples.
4. Run ./run.sh Notice: this downloads another 2GB of data, a pre-built image from Docker.

After installing Docker, steps 2-4 look like this:

$ git clone https://github.com/kylemcdonald/ml-notebook.git && cd ml-notebook
$ ./update.sh
$ ./run.sh

Type ctrl+d to exit ml-notebook. If you accidentally close the Terminal, the Jupyter notebook will keep running in the background. Whenever you want to run the environment again, just call ./run.sh. Calling ./run.sh when ml-notebook is running in the background will restart ml-notebook.

Adding more memory or CPUs

1. Stop the Docker VM by running docker-machine stop default.
2. Open up VirtualBox and click on the "default" machine. Under "System" make the changes you want to see. E.g., change the memory to 8192MB and CPUs to 4.
3. Run docker-machine start default to restart the VM.
4. Run run.sh and check the free memory and number of processors:

$ root@a86d27857c85:~/shared# nproc 
4
$ root@a86d27857c85:~/shared# free -h
             total       used       free     shared    buffers     cached
Mem:          7.8G       295M       7.5G       124M        10M       150M
...

Build notes

The Chainer portion was based on this process:

$ git clone https://github.com/pfnet/chainer-test.git && cd chainer-test
$ ./make_docker.py --base ubuntu14_py2 --numpy numpy110 --cuda none --cudnn none
$ open -t Dockerfile

And removing RUN pip install numpy==1.10.2.

The TensorFlow portion was based on the official TensorFlow Dockerfile.

Acknowledgements

Most of the deep learning toolkits are installed based on Dockerfiles from Kaixhin.