Open anshumitts opened 5 years ago
@anshumitts You would need to see the corresponding datasets as listed on paper github repo. You will find that this in turn points to SentEval toolkit
@vineetm This uses pytorch, are we supposed to write codes for evaluation tasks as well?
@anshumitts I don't expect you to write code for evaluation tasks. See this example on SentEval repo page. It shows how tensorflow code can be called.. https://github.com/facebookresearch/SentEval/blob/master/examples/googleuse.py
@vineetm http://ixa2.si.ehu.es/stswiki/index.php/STSbenchmark none of the links are working here please help
@vineetm TREC data is also not available on the given site.
@anshumitts Given the short time to go before the deadline, I would advise you to work with data that is available. You can skip the data you are not able to access
@vineetm It's not clear whether we should use encoder RNN to create sentence embeddings or Bag of words. Can you please suggest me the same?
@anshumitts You should use encoder RNN to create sentence embeddings. If you have time, compare the results with bag of words (Optional).
@vineetm This is causing an issue Tensorflow in eager mode is conflicting with pytorch causing.
buss error
Please suggest a solution
@anshumitts I can suggest you an alternate solution. You can pre-compute the sentence vectors for the evaluation tasks, and store them in a numpy array or equivalent. When working with pytorch you can use this numpy array... This might work as you don't really care about Tensorflow model, you only care about the sentence vectors.
Can you please point me towards the standard datasets being used in 10.1, 10.2, 10.3 and 10.4.