Seq2Vec for all Seq2Seq

[rangwani-harsh]

System (please complete the following information):

• OS: Ubuntu 18.0
• Python version: 3.6
• AllenNLP version: 0.7.1

Question Hi team, Is there a way we could probably use this https://github.com/allenai/allennlp/blob/master/allennlp/modules/seq2vec_encoders/pytorch_seq2vec_wrapper.py to generate a vector for all seq2seq encoders available in allennlp?

Thanks

[matt-gardner]

What's the motivation for this? You can already use something like a bag of words encoder, or some other kind of pooling, on top of any seq2seq encoder. I'm not sure what you're hoping to accomplish with what you're suggesting.

[rangwani-harsh]

As allennlp has implementations for stacked LSTM with highway etc. which in a way are similar to LSTM. I want to use them as a Seq2Vec encoder opposed to using them as a Seq2Seq Encoder. As this code does the same thing i.e. to generate a vector from the hidden state and the time distributed state and was implemented for the pytorch RNN and LSTM units. I thought it shouldn't be hard to modify and make the code worked for stacked versions too.

There is a paper which uses this kind of seq2vec encoding http://aclweb.org/anthology/S18-1006 and achieves good results.

[HarshTrivedi]

I am not sure if this is what you want but in case you are looking for a way to extract final hidden states of a Seq2SeqEncoder output, get_final_encoder_states function is what you need.

[rangwani-harsh]

Thanks. This was what I was looking for. Although I do think that this information must be present on the seq2vec and seq2seq documentation page.