ran

ran copied to clipboard

Introduction

This codebase supports replication of the language modeling results in Recurrent Additive Networks (Kenton Lee, Omer Levy, and Luke Zettlemoyer).

Recurrent Additive Networks

The TensorFlow implementation of Recurrent Additive Networks (RAN) is found in ran.py and is used by the experiments in the subdirectories.

Experiments

Penn Treebank:

The word-level language modeling for Penn Treebank is found under the ptb directory. This code is derived from https://github.com/tensorflow/models/tree/master/tutorials/rnn/ptb.

Data preparation

• curl -O http://www.fit.vutbr.cz/~imikolov/rnnlm/simple-examples.tgz
• mkdir data
• tar -xzvf simple-examples.tgz -C data

Train and Evaluate

• python -m ptb.ptb_word_lm --data_path=data/simple-examples/data --model=tanh_medium

Replace tanh_medium with the desired setting.

Billion-word Benchmark:

The word-level language modeling for the billion-word benchmark is found under the bwb directory. This code is derived from https://github.com/rafaljozefowicz/lm.

Data preparation

• curl -O http://www.statmt.org/lm-benchmark/1-billion-word-language-modeling-benchmark-r13output.tar.gz
• mkdir data
• tar -xzvf 1-billion-word-language-modeling-benchmark-r13output.tar.gz -C data
• curl -o data/1-billion-word-language-modeling-benchmark-r13output/1b_word_vocab.txt https://raw.githubusercontent.com/rafaljozefowicz/lm/master/1b_word_vocab.txt

Train

• CUDA_VISIBLE_DEVICES=0,1 python -m bwb.single_lm_train --logdir logs --num_gpus 2 --hpconfig num_shards=2 --datadir data/1-billion-word-language-modeling-benchmark-r13output

Evaluate

• CUDA_VISIBLE_DEVICES= python -m bwb.single_lm_train --logdir logs --mode eval_test_ave --hpconfig num_shards=2 --datadir data/1-billion-word-language-modeling-benchmark-r13output

Text8:

The character-level language modeling for Text8 is found under the text directory. This code is derived from https://github.com/julian121266/RecurrentHighwayNetworks

Data Preparation

• curl -O http://mattmahoney.net/dc/text8.zip
• mkdir data
• unzip text8.zip -d data

Train and Evaluate

• python -m text8.char_train