This code repo is built on a frozen version of fairseq. We run our experiment on 2 RTX 6000 GPU (24GB). If your GPUs have smaller memory, you may need to adjust batch size or the maximum length.
For model training, you can refer to the script in train script.
./finetune_sinkhorn_hepos.sh
For model decoding, you can use the command
python longbart_decode.py --path [/path/to/save/summary] --model_dir [/path/to/model] --data_dir [/path/to/data]
For ROUGE evaluation, you can run
python eval_mode.py --decode_dir [/path/to/save/summary] --ref_dir [/path/to/ref] --rouge
Processed GovReport Data can be found here: https://drive.google.com/drive/folders/128KyqPTwZ0Si9RV_IX-md2dcHeRTUHkr