2015 PyCon (SG) Presentation :: 2016 version is 'deep-learning-workshop'
Jupyter Notebook Python HTML
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
.geany
2015-06-19_Presentation-PyConSG
bin
data
ipynb
models
.gitattributes
.gitignore
LICENSE
README.md
gpu_check.py
requirements.txt
theano.cuda-gpuarray.rc

README.md

Machine Learning : Going Deeper with Python and Theano

This repo contains the materials for my presentation to PyCon (SG), on 19 June 2015.

If you are interested in the wonderful Virtual Machine goodness of my 2016 PyCon presentation, the repo you should be looking at is this one.

Installation of requirements

Even if you like to virtualenv, it may make sense to use system-wide installations of some of the basic Python numerical packages, since they're likely to be ready-optimized (even if slightly older) than ones installed/compiled by pip (which will willingly install without OpenBLAS, for instance) :

dnf install scipy numpy python-pandas Cython 

(and, since it's so handy):

dnf install pydot

Note that installing python-ipython-notebook system-wide doesn't seem to work well, because there is a version conflict involving tornado.

Then, as usual (but making use of these system-site-packages) :

virtualenv --system-site-packages env
. env/bin/activate
pip install --upgrade pip
pip install -r requirements.txt 

# Wait 10mins (107Mb of stuff in env/lib/python2.7/site-packages/)

Running the Presentation

Starting at the live-plotting example (which will also want bokeh-server running, see below) :

. env/bin/activate
ipython notebook ipynb/1-LivePlotting.ipynb

# Then open a browser at : http://localhost:8888/
# or, more specifically  : http://localhost:8888/ipynb/1-LivePlotting.ipynb

If you have a browser already running, it may be best to use the --browser option to prevent the distracting launch of an additional browser window:

ipython notebook --port=8888 --browser=none

To run the live-plotting example, you'll also need to start the bokeh-server in another process :

. env/bin/activate
bokeh-server

Running the Presentation (across a network)

. env/bin/activate
ipython notebook --ip=0.0.0.0 --port=8888 --browser=none &
bokeh-server --ip=0.0.0.0 --port=8889

The iPython notebook call to bokeh.io.output_notebook() appears to find the correct port for the bokeh-server automagically.

Remember to adjust the firewall to allow these two open ports...

GPU-aware iPython

Laptop

Making use of an Nvidia card in a notebook (where COMMAND-LINE is the python command that one would ordinarily run) is usually as follows ::

THEANORC=theano.cuda-gpuarray.rc optirun {COMMAND-LINE}

However, because IPython spawns sub-processes to handle each kernel/notebook, the optirun invocation isn't made for the child processes that should actually perform the work on the GPU.

So far, the only route to making this work has been to replace the env/bin/python2 with a script that runs optirun python2-bin where python2-bin is a copy of the previously existing python2. But doing this the causes all python (in that virtualenv) to switch the GPU on, which wasn't really the plan.

Notes : Git-friendly iPython Notebooks

Using the code from : http://pascalbugnion.net/blog/ipython-notebooks-and-git.html (and https://gist.github.com/pbugnion/ea2797393033b54674af ), you can enable this kind of feature just on one repository, rather than installing it globally, as follows...

Within the repository, run :

# Set the permissions for execution :
chmod 754 ./bin/ipynb_optional_output_filter.py

git config filter.dropoutput_ipynb.smudge cat
git config filter.dropoutput_ipynb.clean ./bin/ipynb_optional_output_filter.py

this will add suitable entries to ./.git/config.

or, alternatively, create the entries manually by ensuring that your .git/config includes the lines :

[filter "dropoutput_ipynb"]
	smudge = cat
	clean = ./bin/ipynb_output_filter.py

(where REPO is the absolute path to the root of the checked out repository).

Note also that there's a <REPO>/.gitattributes file here containing the following:

*.ipynb    filter=dropoutput_ipynb

There are two different approaches to doing the cleansing in the REPO/bin directory :

  • ipynb_output_filter.py : which is probably more comprehensive, since it uses iPython itself to parse and output the notebooks - but care must be taken to ensure that it is run within the current env

  • ipynb_optional_output_filter.py : This is my current chosen approach, which only uses import json to parse the notebook files (and so can be executed as a plain script). It also includes the git:suppress_outputs=false option that might be useful...

To include disable the output-cleansing feature in a notebook in the latter case, simply add to its metadata (Edit-Metadata) as a first-level entry (true is the default):

  "git" : { "suppress_outputs" : false },

Notes : Building the Presentation

For 'blocks-introduction-mnist.ipynb' I used the tutorial from the blocks documentation as a starter :

  • wget https://raw.githubusercontent.com/mila-udem/blocks/master/docs/tutorial.rst
  • pandoc --mathjax --from=rst --to=markdown_mmd tutorial.rst > tutorial.md

Also useful :

Notes : Installing PyGPU

On the date of the PyCon, building the libgpuarray library from github FAILS :

NOW DONE - submitted a PR for : 
           https://github.com/Theano/libgpuarray/issues/55
           """ gcc 5.1.1 : max_align_t also defined in stddef.h """

There is a full write-up on how to install an Nvidia GPU under Fedora 22 as a blog posting

And an additional write-up for the case that you're installing to a laptop with 'dual graphics cards' in this blog post.