tensorpack

Why we included this project

Tensorpack exists to make graph-mode TensorFlow train faster. It is a training interface rather than another layer of model wrappers, and it ships off-the-shelf multi-GPU and distributed data-parallel strategies. The dataflow system parallelizes pure-Python preprocessing so your data pipeline does not become the bottleneck, a trade-off the project defends explicitly against tf.data. The repo also includes faithful re-implementations of published papers instead of toy MNIST demos, which gives researchers a solid starting point for baselines. If you work in TensorFlow and need both speed and flexibility, this is a mature option.