The tensorflow-vgg repo is contained in this repo as a "fake submodule"

1. Clone this repository

git clone this_repository
cd this_repository

2. Download yearbook dataset

wget http://www.cs.utexas.edu/~philkr/cs395t/yearbook_trainval.tar.gz
tar -xzf yearbook_trainval.tar.gz 

3. create a small subset of yearbook dataset

ipython subset_dataset.py

4. test if input_pipeline.py works correctly

ipython input_pipeline.py

input_pipeline.py is a file used by almost all other files to create an "input pipeline" in the tensorflow graph. (Graph is a concept in tensorflow. See https://www.tensorflow.org/versions/r0.10/api_docs/python/framework.html)

5. choose whatever model you want to work on:

ipython runvgg.py: train VGG-19.
You may want to tune the parameter by running python runvgg.py -learning_rate 1e-4 -eps 1e-8 to get good results. Note: if you want to specify parameters, you cannot use ipython runvgg.py. You have to use python runvgg.py. I don't know why this won't work for ipython...

ipython fully_connected.py: train a 2 hidden layer network.

ipython tinyfull.py: train a softmax classifier.
i.e. a tiny fully connected network with no hidden layers.
tinyfull.py is designed to be as short and simple as possible so that I can test out new code / debugging a Tensorflow problem. So if you want to know the structure of the other files, I recommend reading the code in tinyfull.py first.

export lr=1e-6
export eps=1e-3
sbatch cmd.sh

And then, use showq -u to see your jobs. use scancel 293409(job id) to cancel your jobs. The output of your jobs will be at the filename specified in cmd.sh.