conda environment
cudatoolkit               10.0.130
cudnn                     7.6.5
cython                    0.29.15
numpy                     1.18.1
apex                      0.1
pytorch                   1.4.0 

Install

python setup.py develop

Evaluate of Model
Correpondence between index and dimension
256    320    384    448    512    576    640    704    768    832    896    960   1024
 0      1      2      3      4      5      6      7      8      9      10     11    12

to test the specific model, modify index on the following file.

Line 15 in fairseq/sequence_generator.py

stage 1
fairseq-generate data-bin/wmt16_en_de_bpe32k --path ./en_de_setup1/stage1.pt --beam 4 --lenpen 0.6 --remove-bpe > wmt16_gen.txt 
bash scripts/compound_split_bleu.sh wmt16_gen.txt
Table 2 Stage 1 in main paper

stage 2
fairseq-generate data-bin/wmt16_en_de_bpe32k --path ./en_de_setup1/stage2.pt --beam 4 --lenpen 0.6 --remove-bpe > wmt16_gen.txt 
bash scripts/compound_split_bleu.sh wmt16_gen.txt
Table 2 Stage 1+2 in main paper

stage 3
fairseq-generate data-bin/wmt16_en_de_bpe32k --path ./en_de_setup1/stage3.pt --beam 4 --lenpen 0.6 --remove-bpe > wmt16_gen.txt 
bash scripts/compound_split_bleu.sh wmt16_gen.txt
Table 2 Stage 1+2+3 in main paper

Anoymous Google Drive Link
https://drive.google.com/file/d/1_dYo5gd3LPvXKMwSDWMlT6lZnD-u-xkO/view?usp=sharing
