README.md

AWS Seoul Summit 2017 Demos

Demo codes in our presentation about MXNet in AWS Seoul Summit 2017

File Name	Description
mxnet-logistic_regression_diabetes.ipynb	The logistic regression example
mxnet-2hidden_fnn_diabetes.ipynb	The classification example using FNN with 2 hidden layers
mxnet-mnist_deep_cnn.ipynb	Example of classifying MNIST digits with a CNN
mxnet-seq2seq.ipynb	The sequence-to-sequence learning example
mxnet_seq2seq_cudnn_speed.py	The MXNet side script that uses cudnn accelerated LSTM for seq2seq model
mxnet_seq2seq_native_speed.py	The MXNet side script that uses the native implemented LSTM for seq2seq model
keras_seq2seq_speed.py	The Keras side script for seq2seq model

You can preview all the notebooks here or using in https://nbviewer.jupyter.org/github/sxjscience/aws-summit-2017-seoul/tree/master/. Also, you can refer to the code and youtube tutorials in https://github.com/hunkim/DeepLearningZeroToAll for more explanation.

For the speed comparison, we use these commands:

MXNet with CUDNN accelerated LSTM

python3 mxnet_seq2seq_cudnn_speed.py

MXNet with native LSTM

python3 mxnet_seq2seq_native_speed.py

Keras 2.0.3 with TensorFlow Backend (TensorFlow version 1.0.1)

KERAS_BACKEND=tensorflow python3 keras_seq2seq_speed.py

Keras 2.0.3 with the Theano backend (Theano version 0.9)

KERAS_BACKEND=theano python3 keras_seq2seq_speed.py

We use a single GeForce GTX 1080 GPU + CUDNN V5.1

Implementation	Time spent
MXNet with CUDNN LSTM	5.34s
MXNet with native LSTM	16.96s
Keras with TF backend	125.59s
Keras with Theano backend	66.24s

Also for the Theano speed test, I've run the script twice to eliminate the compilation time.