Closed neubig closed 4 years ago
I think this should basically be complete (including upgrade of the environment, and conversion of all deprecated PyTorch functions). I've confirmed that the accuracy seems to be fine on geoquery, getting 89.6 for a single run, which is higher than the TranX paper:
{'accuracy': 0.8964285714285715, 'oracle_accuracy': 0.9214285714285714}
I'm running CoNaLa now and will report, but I think this is ready for a check if you have time @pcyin
It seems to also be working on CoNaLa as well, where accuracy is within error of the reported results in the ACL 2019 paper (https://www.aclweb.org/anthology/P19-1447):
{'corpus_bleu': 0.24097779142849596, 'oracle_corpus_bleu': 0.3697198327631323, 'avg_sent_bleu': 0.20894283286460824, 'oracle_avg_sent_bleu': 0.34892593973337316, 'exact_ match': 0.02, 'oracle_exact_match': 0.052}
This is a first draft of an upgrade that doesn't specify version numbers in the conda file, allowing users to use the most recent version of the libraries. Changes in PyTorch broke some things, but I think I fixed them.
This should help address this: https://github.com/pcyin/tranX/issues/10, and should be checked after this: https://github.com/pcyin/tranX/pull/12