search

GriffinLiang / vrd-dsr

Code for Visual Relationship Detection with Deep Structural Ranking (AAAI2018)
122 stars 32 forks source link

About Using Semantic Embedding Cue #15

Closed coldmanck closed 6 years ago

coldmanck commented 6 years ago

Hi @GriffinLiang

Thank you for you fantastic work again. I would like to utilize the fused feature vector to apply it to my dataset (in specifc, COCO). However it seems the model is dataset specific -- only apply to VRD which the number of classes is 100 when I want to retrieve the feature, the error mainly results from semantic embedding part. As COCO does not have relationship labels, it's currently impossible to re-train the model on it.

You indicated in the paper that you used off-the-shelf language model to acquire the word representations, so I am wondering how can I get the semantic embedding directly?

GriffinLiang commented 6 years ago

I use word2vec model. You can refer to the following link https://radimrehurek.com/gensim/models/word2vec.html.

coldmanck commented 6 years ago

Hi @GriffinLiang Thank you for your response. Could you please share your snippet of creating the features for reference purpose? :)

GriffinLiang commented 6 years ago

You should load a pretrained model, and get the representation of an input word. model = Word2Vec.load(fname) model.wv['computer']

You can get the pretrained model following this blog.

coldmanck commented 6 years ago

Thanks for the explanation!