Hands-on: Deep Learning and CNN

Taiwan R User Group, 2015-05-18

Meetup link: http://www.meetup.com/Taiwan-R/events/220799374/ Youtube screencast: http://youtu.be/YI5CLMcvnpY

Slides and Source Code

Nearly everything is originally from Stanford CS231n under MIT License. Especially I didn't make most of the slide content, they are from excellent talks / papers / study notes and I carefully cite their works.

I modified all code to be PEP8 compliant and Py3k compatible. In addition, most of the practices are shorter to fit in an one-hour talk.

Envorinment Setup

Root Python version

Recommend to use miniconda3 or Anaconda3 (they are same in essence). This should work on Windows, Linux, and OSX.

I target for Python 3.4+ but should be fine on Pyhton 3.3. Never ask me how to run on Python 2.7. ...okay, the original source runs on Python 2.7 and I simply change some library import path and classic 2vs3 difference.

NOTE for Windows user: VS 2010 Professional+ is needed to have a 64bit C compiler. DONT use Cgywin or MinGW if you simply just googled and found them.

(Skippable) Setup using pyenv

You can use pyenv to manage multiple Python versions without breaking the system-wide setting. Follow pyenv's readme to set up, which has been tested to work on Debian, Ubuntu, and OSX (but not on Windows).

# under this repo root
pyenv install miniconda3-3.8.3
pyenv local miniconda3-3.8.3

Now all python command under this repo root use miniconda's python, which can be checked by

pyenv which python
# ~/.pyenv/versions/miniconda3-3.8.3/bin/python
python
# Python 3.4.3 |Continuum Analytics, Inc.| (default, Mar  6 2015, 12:07:41)
# ...
# >>>

Python virtual environment

(Mini)conda handles the virtual environment itself. It is powerful and makes everything simple for numerical computing packages.

conda create -n dnn python=3.4 \
	numpy cython            \
	matplotlib              \
	ipython-notebook

Activate and deactivate the envrionment is easy,

source activate dnn
deactivate

Init the dataset

Go to code/cs231n/datasets and run the script get_datasets.sh.

NOTE for Windows user: In addition, by default tools like wget and tar do not exist so get_dataset.sh will fail, but it can be done using pure Python through urllib.request.urlretrieve(url, local_pth) and python -m tarfile -e xxxx.tar.gx.

Run the notebook

Go to code and run

source activate dnn
ipython notebook

A web browser will pop up on http://localhost:8888 by default. If the code runs on a server, try bind to the correct ip address by ipython notebook --ip="<your-ip>" and make sure the port is permitted by the firewall.

License

The slide is powered by

• shower.js: HTML5 slideshow framework by Vadim Makeev et al., under MIT license
• highlight.js: Syntax highlight library by Ivan Sagalaev et al., under MIT license

Unless explicitly stated,

• the content of the slides (at slides folder) is shared under Creative Commons 4.0 BY licesne.
• the code is shared under MIT license.

More information about license at CC Attribution 4.0 and LICENSE_MIT.