Tensorflow training framework

This repository includes an extendable framework for training neural network models using TensorFlow. The framework handles things like parallel data reading, batching, shuffling, saving/loading checkpoints, and logging. It also supports training on multiple GPUs. You can simply add your own dataset by following the example in dataset/mnist.py, and write your own classifier similar to the bundled CNN classifier in model/alexnet.py. The framework does most of the boilerplate code for you, letting you focus on developing the actual neural net model.

Also, make sure to checkout Effective Tensorflow, which to some extents explains the code in this framework.

Pull requests with new datasets or models are welcome!

Install dependencies

pip install tensorflow numpy pillow matplotlib six

Training

To train an mnist classification model run:

python -m trainer --config=configs/mnist.json

To visualize the training logs on Tensorboard run:

tensorboard --logdir=output

Current included datasets are:

• mnist
• cifar10
• cifar100

Bundled models include:

• alexnet