Inconsistency between the paper and the code on getting sequence embedding #2
Description
In the Materials and Methods section of your paper, it is mentioned that you used mean pooling of amino-acid embeddings along the length of the sequence to get sequence embedding. However, in the code ppi/modeling.py and symmetry/modeling.py , you used the first token embedding (pooler_output) as the sequence embedding.
p1_embedding = self.language_model(p1_tokens, p1_attention)[1]
p2_embedding = self.language_model(p2_tokens, p2_attention)[1]
mean_embedding = torch.div(torch.add(p1_embedding, p2_embedding), 2)
Which method did you actually use for the results you showed on the paper?