I’m working on a siamese-like architecture with triplet loss where the network inputs are a mix of numerical, categorical and textual features. Some of the features like “alternate names” come with multiple values. I’d like to
-
share my encoder weights so that the same (text) encoder used for single-value fields can be reused for multi-value fields too.
-
apply the text encoder on each value of a multivalue feature, followed by max pooling for extracting the most relevant signal out of the available values.
what’s the most efficient way to achieve the above?
Thanks