fenchri / dsre-vae Goto Github PK

Hello. I'm running this code and when I try to make the data:
cd data/KB/
python make_data.py --data Freebase
--train_file ../LANG/nyt10/nyt10_train.txt
--val_file ../LANG/nyt10/nyt10_val.txt
--test_file ../LANG/nyt10/nyt10_test.txt

It could not get access to "https://docs.google.com/uc?export=download".
Can you share a link to directly download the data?

Hi, thanks for your great work! I followed the README to run codes step by step, but I cannot replicate the results reported in the paper.

1. I have tried to use your provided pairs_mu_64d.json to run python main.py --config ../configs/nyt10_prior.yaml --mode train, and got the best_epoch = 7. Then I ran python main.py --config ../configs/nyt10_prior.yaml --mode test, and got the AUC of 0.365.
2. I have also tried to train the KB embeddings from scratch. Following the README, I ran train_embeds.sh and calculate_priors.py, and then used the generated pairs_mu_64d.json to run python main.py --config ../configs/nyt10_prior.yaml --mode train. In this way, I got the best_epoch = 15 and the final AUC of 0.391.

I would like to know why these results are far from that reported in the paper. Is there any operating step I missed?

If it cannot be figured out, could you please provide your ckpt files to me? I need the ckpts of N(0, 1) and N(mu, 1) for nyt10 and nyt10_570k for compared baselines. Of course I'll cite your paper. Thank you 😊