How to predict with a lda2vec model

[alexwongdl]

When I have a lda2vec model, how can I predit the vector of a document with this model. I couldn't find the prediction method in any of following files : corpus.py dirichlet_likelihood.py embed_mixture.py fake_data.py init.py negative_sampling.py preprocess.py topics.py tracking.py utils.py.

[yg37]

It doesn't have a prediction function yet

[ali3assi]

@AlexWang90 @yg37 did you have a solution for predicting the embedding for a new document? can you please share your slolution

[yg37]

@TamouzeAssi I wrote a prediction function and tested it on some new documents. The results are not very good (I think it's because I did not run enough epochs). I can send you the function if you want to experiment with it more.

[ali3assi]

@yg37 yes please send it, i wanna try it and check what can i do and improve it. thank you

[yg37]

@TamouzeAssi What's your email address?

[nateraw]

@yg37 can you email me? I tried to reach out via the email on your bio, but it seems it may be dead! My email is [email protected] . Thank you!

[nateraw]

@yg37 Thank you for your quick reply!! I am using Tensorflow, as I don't know Chainer. Can you give me a theoretical understanding of what you are doing, please?

This is my current thought process:

1. Append a new random document weight to document matrix (Within the range of current weights)
2. Collect skipgram pairs for this document (~30-100 pairs)
3. Collect random sampling of previously trained skipgram pairs (~10000 pairs)
4. Shuffle the two lists of pairs together
5. Train using the combined list of skipgram pairs, locking word vectors
6. Evaluate the new document/topic embeddings

Am I close to what you are doing? Again, thank you!!

[arhelgens]

@yg37 I am I would also be interested in seeing what you have implemented for a prediction function.

Any way that you could either publish your code, or send me what you have.

Thanks

[simon19891101]

Hi @yg37 just wondering if I could have the implementation of your prediction function? My email is [email protected]. Much appreaciated :).

[raziel130889]

Hi @yg37 I would like to use your function to predict, please could you send it to me, this is my email [email protected], in advance thank you

[sudheernaidu53]

hi, @yg37 ,I am working with lda2vec as well, I will be thankful if you can send the predict function across to me.

[Shivanshu156]

Please send the predict function here as well [email protected]

[kandakji]

Hi @yg37 I would much appreciate it if you would send me your predict function implementation :) [email protected] Thank you in advance. :)

[leewide]

Hi @yg37 I would much appreciate it if you would send me your predict function implementation, this is my email [email protected], in advance thank you