Explore regularizers for glove
kudkudak opened this issue · comments
Stanislaw Jastrzebski commented
Assume that performance on farthest bucket is a good predictor of test performance. Performance of farthest bucket is not ATM computed within script (will add before starting this task)
Try to see if any regularizer for glove improves scores. Idea:
- Embedding dropout
- Learning rate schedule with very large LR at the beginning
Compare to numbers in https://www.overleaf.com/12326026ybwcgtxbxpkr#/46877048/. Generate your own data with python scripts/data/split_intrinsic_ACL.py
Example command (glove embedding, l2=1e-5, random split)
python scripts/train_factorized.py root $SCRATCH/l2lwe/results/factorized/3_01_root_conceptnet_my_random_100k_l2a=1e-5_glove --data_dir=$DATA_DIR/LiACL/conceptnet_my_random_100k --l2_a=1e-5 --embedding_file=embeddings/LiACL/embeddings_glove200_norm.txt
Stanislaw Jastrzebski commented
Simple version of this is pushed here: https://github.com/kudkudak/common-sense-prediction/tree/master/scripts/experiments/glove_vs_OMCS