Implement entire database retrieval metrics (after getting checkpoints)

[AdrianM0]

• check the retrieval between non-paired modalities
• embedding arithmetic
• check if embeddings for the same SMILES but rearranged are close to each other
• retrieval top k between central modality and the other modalities

[kjappelbaum]

I feel for this, at least for the subset where we can cleanly convert between modalities, it would be great to have the same dataset size from which we retrieve.

In other cases, I would sample to a fixed size

[kjappelbaum]

additional metrics to consider:

• mean reciprocal rank
• Discounted cumulative gain (DCG)

[AdrianM0]

additional metrics to consider:

• mean reciprocal rank
• Discounted cumulative gain (DCG)

I am doing now MRR on the batch only.