Initializing modelingBert as an identity transformation

given a batch of inputs_embeds, i want to propagate it through BERTModel such that without any training the output of the BERTModel will be identical to the inputs_embeds.

let me clarify that it’s important that the actual operations take place and so what is needed is a certain weight initialization that i’m not sure is very trivial…

any ideas on how to do that?