Code for our EMNLP 2021 paper "Learning Kernel-Smoothed Machine Translation with Retrieved Examples"

Last update: Nov 24, 2022

Related tags

Overview

KSTER

Code for our EMNLP 2021 paper "Learning Kernel-Smoothed Machine Translation with Retrieved Examples" [paper].

Usage

Download the processed datasets from this site. You can also download the built databases from this site and download the model checkpoints from this site.

Train a general-domain base model

Take English -> Germain translation for example.

export CUDA_VISIBLE_DEVICES=0,1,2,3
python3 -m joeynmt train configs/transformer_base_wmt14_en2de.yaml

Finetuning trained base model on domain-specific datasets

Take English -> Germain translation in Koran domain for example.

export CUDA_VISIBLE_DEVICES=0,1,2,3
python3 -m joeynmt train configs/transformer_base_koran_en2de.yaml

Build database

Take English -> Germain translation in Koran domain for example, wmt14_en_de.transformer.ckpt is the path of trained general-domain base model checkpoint.

mkdir database/koran_en_de_base
export CUDA_VISIBLE_DEVICES=0
python3 -m joeynmt build_database configs/transformer_base_koran_en2de.yaml \
        --ckpt wmt14_en_de.transformer.ckpt \
        --division train \
        --index_path database/koran_en_de_base/trained.index \
        --token_map_path database/koran_en_de_base/token_map \
        --embedding_path database/koran_en_de_base/embeddings.npy

Train the bandwidth estimator and weight estimator in KSTER

Take English -> Germain translation in Koran domain for example.

export CUDA_VISIBLE_DEVICES=0,1,2,3
python3 -m joeynmt combiner_train configs/transformer_base_koran_en2de.yaml \
        --ckpt wmt14_en_de.transformer.ckpt \
        --combiner dynamic_combiner \
        --top_k 16 \
        --kernel laplacian \
        --index_path database/koran_en_de_base/trained.index \
        --token_map_path database/koran_en_de_base/token_map \
        --embedding_path database/koran_en_de_base/embeddings.npy \
        --in_memory True

Inference

We unify the inference of base model, finetuned or joint-trained model, kNN-MT and KSTER with a concept of combiner (see joeynmt/combiners.py).

Combiner type	Methods	Description
NoCombiner	Base, Finetuning, Joint-training	Directly inference without retrieval.
StaticCombiner	kNN-MT	Retrieve similar examples during inference. mixing_weight and bandwidth are pre-specified.
DynamicCombiner	KSTER	Retrieve similar examples during inference. mixing_weight and bandwidth are dynamically estimated.

Inference with NoCombiner for Base model

Take English -> Germain translation in Koran domain for example.

export CUDA_VISIBLE_DEVICES=0
python3 -m joeynmt test configs/transformer_base_koran_en2de.yaml \
        --ckpt wmt14_en_de.transformer.ckpt \
        --combiner no_combiner

Inference with StaticCombiner for kNN-MT

Take English -> Germain translation in Koran domain for example.

export CUDA_VISIBLE_DEVICES=0
python3 -m joeynmt test configs/transformer_base_koran_en2de.yaml \
        --ckpt wmt14_en_de.transformer.ckpt \
        --combiner static_combiner \
        --top_k 16 \
        --mixing_weight 0.7 \
        --bandwidth 10 \
        --kernel gaussian \
        --index_path database/koran_en_de_base/trained.index \
        --token_map_path database/koran_en_de_base/token_map

Inference with DynamicCombiner for KSTER

Take English -> Germain translation in Koran domain for example, koran_en_de.laplacian.combiner.ckpt is the path of trained bandwidth estimator and weight estimator for Koran domain.
--in_memory option specifies whether to load the example embeddings to memory. Set in_memory == True for faster inference, set in_memory == False for lower memory demand.

export CUDA_VISIBLE_DEVICES=0
python3 -m joeynmt test configs/transformer_base_koran_en2de.yaml \
        --ckpt wmt14_en_de.transformer.ckpt \
        --combiner dynamic_combiner \
        --combiner_path koran_en_de.laplacian.combiner.ckpt \
        --top_k 16 \
        --kernel laplacian \
        --index_path database/koran_en_de_base/trained.index \
        --token_map_path database/koran_en_de_base/token_map \
        --embedding_path database/koran_en_de_base/embeddings.npy \
        --in_memory True

See bash_scripts/test_*.sh for reproducing our results.
See logs/*.log for the logs of our results.

Acknowledgements

We build the models based on the joeynmt codebase.

Code for our EMNLP 2021 paper "Learning Kernel-Smoothed Machine Translation with Retrieved Examples"

Related tags

Overview

KSTER

Usage

Train a general-domain base model

Finetuning trained base model on domain-specific datasets

Build database

Train the bandwidth estimator and weight estimator in KSTER

Inference

Inference with NoCombiner for Base model

Inference with StaticCombiner for kNN-MT

Inference with DynamicCombiner for KSTER

Acknowledgements

Owner

jiangqn

Official Repository of NeurIPS2021 paper: PTR

PatchMatch-RL: Deep MVS with Pixelwise Depth, Normal, and Visibility

AgeGuesser: deep learning based age estimation system. Powered by EfficientNet and Yolov5

Automatic Data-Regularized Actor-Critic (Auto-DrAC)

Snscrape-jsonl-urls-extractor - Extracts urls from jsonl produced by snscrape

Project page for the paper Semi-Supervised Raw-to-Raw Mapping 2021.

The official homepage of the COCO-Stuff dataset.

HCQ: Hybrid Contrastive Quantization for Efficient Cross-View Video Retrieval

PyTorch implementation of Algorithm 1 of "On the Anatomy of MCMC-Based Maximum Likelihood Learning of Energy-Based Models"

A library for efficient similarity search and clustering of dense vectors.

FMA: A Dataset For Music Analysis

SymPy-powered, Wolfram|Alpha-like answer engine totally in your browser, without backend computation

Learning based AI for playing multi-round Koi-Koi hanafuda card games. Have fun.

Pervasive Attention: 2D Convolutional Networks for Sequence-to-Sequence Prediction

Parasite: a tool allowing you to compress and decompress files, to reduce their size

Official Pytorch implementation of Scene Representation Networks: Continuous 3D-Structure-Aware Neural Scene Representations

A state of the art of new lightweight YOLO model implemented by TensorFlow 2.

MarcoPolo is a clustering-free approach to the exploration of bimodally expressed genes along with group information in single-cell RNA-seq data

Implementation of the paper Recurrent Glimpse-based Decoder for Detection with Transformer.

Citation Intent Classification in scientific papers using the Scicite dataset an Pytorch