Skip to content

Latest commit

 

History

History
57 lines (47 loc) · 1.88 KB

README-gated.md

File metadata and controls

57 lines (47 loc) · 1.88 KB
Raw

Reproduce results for Gated Fusion

1. Preprocess

tgt=de
TEXT=data/multi30k-en-$tgt/
python preprocess.py --source-lang en --target-lang $tgt \
  --trainpref $TEXT/train \
  --validpref $TEXT/valid \
  --testpref $TEXT/test.2016,$TEXT/test.2017,$TEXT/test.coco \
  --destdir data-bin/multi30k.en-$tgt \
  --workers 8 --joined-dictionary

2. Training(en->de as an example)

DATA='data-bin/multi30k.en-de/'  # input data
ARCH='gated_tiny'  # model structure
SAVE='checkpoints/gated.en-de.tiny'  # save dir
FEATURE=xxxxx  # path to visual features you downloaded
tgt='de'

CUDA_VISIBLE_DEVICES=0,1 python train.py $DATA --task mmt \
      --arch $ARCH --share-all-embeddings --dropout 0.3 \
      --warmup-updates 2000 --lr 0.005 \
      --max-tokens 4096 \
      --max-update 8000 --target-lang $tgt \
      --save-dir $SAVE \
      --visual_feature_file $FEATURE \
      --find-unused-parameters --patience 10

3. Evaluate

bash evaluate.sh -g 0 -d $DATA -s test -t mmt -p $SAVE

evaluation script parameters:

  • -s test subset {chose from test(2016)/test1(2017)/test2(coco)}
  • -g gpu id you want to use
  • -d input data
  • -p checkpoint path (note: just path to checkpoint dir, not to the file)
  • -b beam size, default to 5
  • -t task name, {translation/mmt/retrieval_translation}

Run the evaluation commanda above, you are supposed to see:

Generate test with beam=5: BLEU4 = 41.96, 71.4/48.1/35.0/25.7 (BP=1.000, ratio=1.001, syslen=12121, reflen=12103)

4. Model analysis

If you would like to analysis the gate value, uncomment L313-314 in gated.py

And then run the evaluation script as above to write gating matrix to a local file, then you can compute the averaged gate value using:

python scripts/visual_awareness.py --input checkpoints/gated.en-de.tiny/gated.txt