maluuba / gensen Goto Github PK

Thank you for this great contribution.

I am currently performing some experiments where it would be very helpful if I had a network that can encode sentences, decode them, and calculate entailment with regard to that same encoding.

I think it would be very useful to the community if you provided a link to your pretrained model in the same checkpointing format as that of your train.py file. I hope you still have it somewhere ..

Thanks a lot !

Hi,

thanks for the code!
The provided glove2h5.py does not work on my machine as there are some words in the gloVe file which contain spaces and thus the code crashes when trying to convert the splitted lines to float.
The following lines should be changed:
vocab = [line[0] for line in glove_vectors]
into
vocab = [' '.join(line[0:-300]) for line in glove_vectors]

and
vectors = np.array( [[float(val) for val in line[1:]] for line in glove_vectors] ).astype(np.float32)
into
vectors = np.array( [[float(val) for val in line[-300:]] for line in glove_vectors] ).astype(np.float32)

some cudnn version error

Hi Sandeep,
I have a question about the input word embeddings. As you described in READEME, the pre-trained GLOVE is used. But in the paper, the word embeddings were learned. If I understand correctly, GLOVE is only used in the case if we want to expand the Vocabulary. When generating sentence representations, the model still uses the learned word embeddings. Is this right?

Hi Sandeep,
I would like to encoding tweets to vectors. I can use your pre-trained model. But it is possible to train (warm-start with your pre-trained parameters) your model with tweets data?

Cheers,
Shun

I read the peephole GRU implementation in models.py :

newgate = F.tanh(i_n + resetgate * h_n + p_n) （line54）
hy = newgate + inputgate * (hidden - newgate) . (line 55 )

Are they in consistent with the (3) and (4) equations in the paper?
I think the line 54 missed the “entrywise product of r_t and h_t-1” and the line 55 also not looks like the equation (4).

It would be great if you can release python 3 compatible code, thanks