Context: instead of a text-to-condition, I have directly a vector, which is an output from some other model (which is also learnable ).
In this context, how to do conditional fine-tuning of the GPT-2 model. From the GPT-2 model’s perspective, we can think that the dataset to train is a list of tuples of the form (condition_vec,text_sample). But note that condition_vec is an output from some other model that is also involved in the optimisation.
One naive approach is to feed this vector directly as the context vector (Assume dimensions match).
Will it work? (I could test it soon)
Or is there any better solution?