[integration] Work towards full model2vec integration

[tomaarsen]

Hello!

Pull Request overview

• Be able to load Model2Vec weight files as long as the model itself has a modules.json pointing to StaticEmbedding.

Details

This PR is to work towards compatibility of:

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("minishlab/potion-base-8M")
embeddings = model.encode(["It's wonderful weather!", "I love the weather today.", "It's raining cats and dogs."])
print(embeddings.shape)

similarity = model.similarity(embeddings, embeddings)
print(similarity)

(3, 256)
tensor([[1.0000, 0.7467, 0.3813],
        [0.7467, 1.0000, 0.4091],
        [0.3813, 0.4091, 1.0000]])

Beyond the changes in this PR, the Model2Vec models must be updated with a modules.json like:

[
    {
        "idx": 0,
        "name": "0",
        "path": ".",
        "type": "sentence_transformers.models.StaticEmbedding"
    }
]

or

[
    {
        "idx": 0,
        "name": "0",
        "path": ".",
        "type": "sentence_transformers.models.StaticEmbedding"
    },
    {
        "idx": 1,
        "name": "1",
        "path": "1_Normalize",
        "type": "sentence_transformers.models.Normalize"
    }
]

@Pringled and @stephantul are working towards the changes on the model2vec side.

• Tom Aarsen