Support multi-modal items

[AlekseySh]

Currently we consider items as texts or images. But in some real world applications an item may be presented as a set of arbitrary number of images, texts and even categorical or continuous features. For example, in e-commerce an item in catalog may include a few images, a few text fields and some other variables. The idea is to implement Dataset and Model that support may support such kind of data, the rest of pipeline is modality agnostic, so it should not be changed.

Since such implementation has a lot of moving parts and design decisions, let's first discuss how the potential result may look like.

[AlekseySh]

A few links that may be useful:

• Watch and Buy: Multimodal Product Identification Challenge 2021
• Workshop on Open-Domain Reasoning Under Multi-Modal Settings 2023
• Shopee - Price Match Guarantee