Fix paraphrase minilm

[hh-space-invader]

import torch
import numpy as np
from transformers import AutoTokenizer, AutoModel
from sentence_transformers import SentenceTransformer

from fastembed import TextEmbedding


model_name = 'sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2'
docs = ["hello world", "flag embedding"]

model = SentenceTransformer(model_name)
st_embeddings = model.encode(docs)
rounded_st_embeddings = np.array([[round(value, 4) for value in embedding[:5]] for embedding in st_embeddings])
print(f"st_embeddings: {rounded_st_embeddings}")

# Mean Pooling - Take attention mask into account for correct averaging
def mean_pooling(model_output, attention_mask):
    token_embeddings = model_output[0] #First element of model_output contains all token embeddings
    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)

# Load model from HuggingFace Hub
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModel.from_pretrained(model_name)

# Tokenize sentences
encoded_input = tokenizer(docs, padding=True, truncation=True, return_tensors='pt')

# Compute token embeddings
with torch.no_grad():
    model_output = model(**encoded_input)

# Perform pooling. In this case, max pooling.
hf_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])
rounded_hf_embeddings = np.array([[round(value, 4) for value in embedding[:5]] for embedding in np.array(hf_embeddings)])
print(f"hf_embeddings: {rounded_hf_embeddings}")

model = TextEmbedding(model_name, cache_dir="models")
fe_embeddings = np.array(list(model.embed(documents=docs)))
rounded_fe_embeddings = np.array([[round(value, 4) for value in embedding[:5]] for embedding in np.array(fe_embeddings)])
print(f"fe_embeddings: {rounded_fe_embeddings}")

print(np.allclose(
    a=rounded_st_embeddings,
    b=rounded_hf_embeddings,
    atol=1e-3 # taken from test file
))

print(np.allclose(
    a=rounded_st_embeddings,
    b=rounded_fe_embeddings,
    atol=1e-3 # taken from test file
))

All Submissions:

ref: #368

• Have you followed the guidelines in our Contributing document?
• Have you checked to ensure there aren't other open Pull Requests for the same update/change?

New Feature Submissions:

• Does your submission pass the existing tests?
• Have you added tests for your feature?
• Have you installed pre-commit with pip3 install pre-commit and set up hooks with pre-commit install?

New models submission:

• Have you added an explanation of why it's important to include this model?
• Have you added tests for the new model? Were canonical values for tests computed via the original model?
• Have you added the code snippet for how canonical values were computed?
• Have you successfully ran tests with your changes locally?

[joein]

We need to add a warning that the model has been updated, something similar to the one we're showing when users use splade with incorrect spelling

[joein]

could you please upload notebooks to reproduce canonical vectors to the dedicated repo?

[joein]

should not be in this PR

[joein]

could you please add fastembed version in which it's gonna changed? (0.5.2)