aonotas / adversarial_text

Code for Adversarial Training Methods for Semi-Supervised Text Classification

Adversarial Training Methods for Semi-Supervised Text Classification

Code for Adversarial Training Methods for Semi-Supervised Text Classification

This code reproduce the [Miyato et al., 2017] with Chainer.

Setup envirment

Please install Chainer and Cupy.

You can set up the environment easily with this Setup.md.

Download Pretrain Model

Please download pre-trained model.

$ wget http://sato-motoki.com/research/vat/imdb_pretrained_lm.model

Result

Model	Error Rate
Baseline [Miyato et al., 2017]	7.39
Baseline (Our code)	6.62
Adversarial [Miyato et al., 2017]	6.21
Adversarial Training (Our code)	6.35
Virtual Adversarial Training [Tensorflow code]	6.40
Virtual Adversarial Training [Miyato et al., 2017]	5.91
Virtual Adversarial Training (Our code)	5.82

Run

Pretrain

$ python -u pretrain.py -g 0 --layer 1 --dataset imdb --bproplen 100 --batchsize 32 --out results_imdb_adaptive --adaptive-softmax

Note that this command takes about 30 hours with single GPU.

Train (VAT: Semi-supervised setting)

$ python train.py --gpu=0 --n_epoch=30 --batchsize 32 --save_name=imdb_model_vat --lower=0 --use_adv=0 --xi_var=5.0  --use_unlabled=1 --alpha=0.001 --alpha_decay=0.9998 --min_count=1 --ignore_unk=1 --pretrained_model imdb_pretrained_lm.model --use_exp_decay=1 --clip=5.0 --batchsize_semi 96 --use_semi_data 1

Note that this command takes about 8 hours with single GPU.

Train (Adversarial Training: Supervised setting)

$ python train.py --gpu=0 --n_epoch=30 --batchsize 32 --save_name=imdb_model_adv --lower=0 --use_adv=1 --xi_var=5.0  --use_unlabled=1 --alpha=0.001 --alpha_decay=0.9998 --min_count=1 --ignore_unk=1 --pretrained_model imdb_pretrained_lm.model --use_exp_decay=1 --clip=5.0

Note that this command takes about 6 hours with single GPU.

Authors

We thank Takeru Miyato (@takerum) who suggested that we reproduce the result of a [Miyato et al., 2017].

• Code author: @aonotas
• Thanks for Adaptive Softmax implementation: @soskek Adaptive Softmax: https://github.com/soskek/efficient_softmax

Reference

[Miyato et al., 2017]: Takeru Miyato, Andrew M. Dai and Ian Goodfellow
Adversarial Training Methods for Semi-Supervised Text Classification.
International Conference on Learning Representation (ICLR), 2017