lhz1029 / dlcl

The implementation of "Learning Deep Transformer Models for Machine Translation"

Learning Deep Transformer Models for Machine Translation on Fairseq

The implementation of Learning Deep Transformer Models for Machine Translation [ACL 2019] (Qiang Wang, Bei Li, Tong Xiao, Jingbo Zhu, Changliang Li, Derek F. Wong, Lidia S. Chao)

This code is based on Fairseq v0.5.0

Installation

1. pip install -r requirements.txt
2. python setup.py develop
3. python setup.py install

NOTE: test in torch==0.4.1

Prepare Training Data

1. Download the preprocessed WMT'16 En-De dataset provided by Google to project root dir

2. Generate binary dataset at data-bin/wmt16_en_de_google

bash runs/prepare-wmt-en2de.sh

Train

Train deep pre-norm baseline (20-layer encoder)

bash runs/train-wmt-en2de-deep-prenorm-baseline.sh

Train deep post-norm DLCL (25-layer encoder)

bash runs/train-wmt-en2de-deep-postnorm-dlcl.sh

Train deep pre-norm DLCL (30-layer encoder)

bash runs/train-wmt-en2de-deep-prenorm-dlcl.sh

NOTE: BLEU will be calculated automatically when finishing training

Results

Model	#Param.	Epoch*	BLEU
Transformer (base)	65M	20	27.3
Transparent Attention (base, 16L)	137M	-	28.0
Transformer (big)	213M	60	28.4
RNMT+ (big)	379M	25	28.5
Layer-wise Coordination (big)	210M*	-	29.0
Relative Position Representations (big)	210M	60	29.2
Deep Representation (big)	356M	-	29.2
Scailing NMT (big)	210M	70	29.3
Our deep pre-norm Transformer (base, 20L)	106M	20	28.9
Our deep post-norm DLCL (base, 25L)	121M	20	29.2
Our deep pre-norm DLCL (base, 30L)	137M	20	29.3

NOTE: * denotes approximate values.