ENH: Enable Global Similarity Comparison Between Text and Image Inputs

[ozan-oktay]

#604
Proposed changes:

• Image inference engine returns global image embeddings if user needs to compute image and sentence level alignment.
• Similarly, VLP engine can use this method to return a single cosine similarity score as in CLIP method.

[fepegar]

Thanks, @ozan-oktay. It would be best to make the methods slightly more modular: one function to load, one function to transform, one function to compute embeddings. This will maximise flexibility and simplify testing. I'm happy to work on that if you like.

[ozan-oktay]

Thanks, @ozan-oktay. It would be best to make the methods slightly more modular: one function to load, one function to transform, one function to compute embeddings. This will maximise flexibility and simplify testing. I'm happy to work on that if you like.

@fepegar The PR is not introducing any new loading or transformation functions -- It's using the existing class methods to return a global embedding that is similar to patch embedding computation. It also adds a separate method in VLP inference engine to return a similarity score. Please feel free to create a separate issue for the idea you have in mind, and if there is something particularly applicable within this context, please let me know. Thanks.