AC-BLSTM

MXNet Scala module implementation of my work AC-BLSTM[1].

Setup

Environment

Tested on Ubuntu 14.04, using CUDA 8.0.61.

Build Steps

Build MXNet

make -j4 USE_MKLDNN=0 USE_CUDA=1 USE_CUDA_PATH=/usr/local/cuda USE_CUDNN=1

For more details how to build MXNet from source pls refer to: http://mxnet.io/get_started/ubuntu_setup.html.

Requirements to Build MXNet-Scala-Package

• Java 8
• maven

make scalapkg

For more details how to build MXNet-Scala-Package pls refer to: http://mxnet.io/get_started/ubuntu_setup.html#install-the-mxnet-package-for-scala.

Build AC-BLSTM Project

Requirements

• sbt 0.13

under the AC-BLSTM folder:

mkdir lib
cp mxnet/scala-package/assembly/linux-x86_64-gpu/target/mxnet-full_2.11-linux-x86_64-gpu-0.1.2-SNAPSHOT.jar lib

Then run sbt and compile the project

Run Experiments

Download Word2Vec Model

You can download the pretrained Word2Vec Model in this url: https://code.google.com/archive/p/word2vec/, then put the GoogleNews-vectors-negative300.bin file to the datas path.

Run Experiments

AC-BLSTM on MR Dataset

cd run_scripts
bash train_ac_blstm.sh

G-AC-BLSTM on MR Dataset

cd run_scripts
bash train_g_ac_blstm.sh

Because I was doing the 10-fold cross-validation on MR dataset, so you can modify the CROSS_VALIDATION_ID= flag from 0 to 9 for the cross-validation expriements.

By the way, If you can successfully reproduce the result reported in the paper, congratulations :) .

If not, God knows what happen :( .

May the force be with you :) .....

References

[1] Liang, Depeng, and Yongdong Zhang. "AC-BLSTM: Asymmetric Convolutional Bidirectional LSTM Networks for Text Classification." arXiv preprint arXiv:1611.01884 (2016).