Ayan Chakrabarti <ayanc@ttic.edu>
This repository contains boiler-plate code that you can use as a starting point to write your tensorflow code for neural-network training. This isn't a library of classes for general purpose use: modify the code to adapt it to your task. I have created this repo mainly as a place to put reference code for students I work with. But all code is released for public use under the MIT license.
There are currently two directories:
-
cifar100
: Somewhat simpler of the two setups. Seetrain_val.py
andtest.py
and follow pointers from there. If you want to run a quick test, remember to set the environment variableCIFAR100
to the directory where you extracted the cifar100 dataset, before you runtrain_val.py
. -
imagenet
: Imagenet training and validation with a vgg-16 architecture network. Start by looking attrain.py
/train_avg.py
andval.py
. You'll need to create thetrain.txt
andval.txt
with lists of JPEG file names and labels: you can usedata/mk_data.sh
to do this.