language model is not used in this version

[quan1e]

Thanks for the code. I cannot find the language model in this version. Have u used it?

[hwfan]

Thanks for using Activity2Vec!

In L49 of activity2vec/networks/pasta_net.py, we use BERT to construct a C x 768 language feature matrix, where C is the class number of part states. Then, in L295 we multiply the feature matrix with the PaSta scores, and obtain the final language features.

In the TensorFlow-based implementation HAKE-Action, these features are further used for HOI classification, while Activity2Vec mainly focuses on verb classification and only use these features as general activity embeddings.

[hwfan]

It should be noted that our language model is pretrained BERT and will not be further finetuned by Activity2Vec, and thus we use the generated features for the downstream classification model.

[quan1e]

Thank you! Your code helps me a lot.