DAG Search

This repositary implements the beamsearch algorithm for DA-Transformer, please see it for usage. It is modified based on the kenlm.

Due to the original license of kenlm, the modified part are licensed under GNU Lesser General Public License (LGPL) 2.1. You can compile a dynamic linking library and then import it in your python project, so your project does not neccesarily need to use LGPL.

Build a n-gram LM

To build a LM for beamsearch, first compile the codes following Compiling.

Then, train the LM:

./build/bin/lmplz -o 5 -S 30% -T /tmp --prune 0 1 5 </path/to/training.target >lm.arpa

(See this for explanations of parameters.) You can modify prune to adjust the size of LM (It also speeds up the decoding). In DA-Transformer, we use a model with ~200M for En-De and ~300M for Zh-En

Modified File List

The implementation are detailed in ./python/Readme.md

python
├── _kenlm.pxd           # Cython header for ngram LM
├── algorith.pxd         # Cython header for CPP STL algorithm lib
├── atomic.pxd           # Cython header for CPP STL atomic lib
├── memviewslice.h       # Cpp header for numpy memviewslice
├── SearchBeam.pxd       # Cython header for SearchBeam (main files)
├── SearchBeam.h         # Cpp header for SearchBeam (main files)
├── SearchBeam.cpp       # Cpp file for SearchBeam (main files)
├── dag_search.cpp       # Automatic Cython generated cpp file
├── dag_search.pyx       # Cython file for dag_search (main files)
└── Readme.md            # Algorithm description

Citing

Please kindly cite us if you find the codes useful.

@inproceedings{huang2022DATransformer,
  author = {Fei Huang and Hao Zhou and Yang Liu and Hang Li and Minlie Huang},
  title = {Directed Acyclic Transformer for Non-Autoregressive Machine Translation},
  booktitle = {Proceedings of the 39th International Conference on Machine Learning, {ICML} 2022},
  year = {2022}
}

kenlm (The original README)

Language model inference code by Kenneth Heafield (kenlm at kheafield.com)

The website https://kheafield.com/code/kenlm/ has more documentation. If you're a decoder developer, please download the latest version from there instead of copying from another decoder.

Compiling

Use cmake, see BUILDING for build dependencies and more detail.

mkdir -p build
cd build
cmake ..
make -j 4

Compiling with your own build system

If you want to compile with your own build system (Makefile etc) or to use as a library, there are a number of macros you can set on the g++ command line or in util/have.hh .

• KENLM_MAX_ORDER is the maximum order that can be loaded. This is done to make state an efficient POD rather than a vector.
• HAVE_ICU If your code links against ICU, define this to disable the internal StringPiece and replace it with ICU's copy of StringPiece, avoiding naming conflicts.

ARPA files can be read in compressed format with these options:

• HAVE_ZLIB Supports gzip. Link with -lz.
• HAVE_BZLIB Supports bzip2. Link with -lbz2.
• HAVE_XZLIB Supports xz. Link with -llzma.

Note that these macros impact only read_compressed.cc and read_compressed_test.cc. The bjam build system will auto-detect bzip2 and xz support.

Estimation

lmplz estimates unpruned language models with modified Kneser-Ney smoothing. After compiling with bjam, run

bin/lmplz -o 5 <text >text.arpa

The algorithm is on-disk, using an amount of memory that you specify. See https://kheafield.com/code/kenlm/estimation/ for more.

MT Marathon 2012 team members Ivan Pouzyrevsky and Mohammed Mediani contributed to the computation design and early implementation. Jon Clark contributed to the design, clarified points about smoothing, and added logging.

Filtering

filter takes an ARPA or count file and removes entries that will never be queried. The filter criterion can be corpus-level vocabulary, sentence-level vocabulary, or sentence-level phrases. Run

bin/filter

and see https://kheafield.com/code/kenlm/filter/ for more documentation.

Querying

Two data structures are supported: probing and trie. Probing is a probing hash table with keys that are 64-bit hashes of n-grams and floats as values. Trie is a fairly standard trie but with bit-level packing so it uses the minimum number of bits to store word indices and pointers. The trie node entries are sorted by word index. Probing is the fastest and uses the most memory. Trie uses the least memory and is a bit slower.

As is the custom in language modeling, all probabilities are log base 10.

With trie, resident memory is 58% of IRST's smallest version and 21% of SRI's compact version. Simultaneously, trie CPU's use is 81% of IRST's fastest version and 84% of SRI's fast version. KenLM's probing hash table implementation goes even faster at the expense of using more memory. See https://kheafield.com/code/kenlm/benchmark/.

Binary format via mmap is supported. Run ./build_binary to make one then pass the binary file name to the appropriate Model constructor.

Platforms

murmur_hash.cc and bit_packing.hh perform unaligned reads and writes that make the code architecture-dependent.
It has been sucessfully tested on x86_64, x86, and PPC64.
ARM support is reportedly working, at least on the iphone.

Runs on Linux, OS X, Cygwin, and MinGW.

Hideo Okuma and Tomoyuki Yoshimura from NICT contributed ports to ARM and MinGW.

Decoder developers

• I recommend copying the code and distributing it with your decoder. However, please send improvements upstream.

• It's possible to compile the query-only code without Boost, but useful things like estimating models require Boost.

• Select the macros you want, listed in the previous section.

• There are two build systems: compile.sh and cmake. They're pretty simple and are intended to be reimplemented in your build system.

• Use either the interface in lm/model.hh or lm/virtual_interface.hh. Interface documentation is in comments of lm/virtual_interface.hh and lm/model.hh.

• There are several possible data structures in model.hh. Use RecognizeBinary in binary_format.hh to determine which one a user has provided. You probably already implement feature functions as an abstract virtual base class with several children. I suggest you co-opt this existing virtual dispatch by templatizing the language model feature implementation on the KenLM model identified by RecognizeBinary. This is the strategy used in Moses and cdec.

• See lm/config.hh for run-time tuning options.

Contributors

Contributions to KenLM are welcome. Please base your contributions on https://github.com/kpu/kenlm and send pull requests (or I might give you commit access). Downstream copies in Moses and cdec are maintained by overwriting them so do not make changes there.

Python module

Contributed by Victor Chahuneau.

Installation

pip install https://github.com/kpu/kenlm/archive/master.zip

Basic Usage

import kenlm
model = kenlm.Model('lm/test.arpa')
print(model.score('this is a sentence .', bos = True, eos = True))

See python/example.py and python/kenlm.pyx for more, including stateful APIs.

---

The name was Hieu Hoang's idea, not mine.