# Fast-Inference with Ctranslate2

Speedup inference while reducing memory by 2x-4x using int8 inference in C++ on CPU or GPU.

quantized version of intfloat/e5-small-v2

pip install hf-hub-ctranslate2>=2.12.0 ctranslate2>=3.17.1

# from transformers import AutoTokenizer
model_name = "michaelfeil/ct2fast-e5-small-v2"
model_name_orig="intfloat/e5-small-v2"

from hf_hub_ctranslate2 import EncoderCT2fromHfHub
model = EncoderCT2fromHfHub(
        # load in int8 on CUDA
        model_name_or_path=model_name,
        device="cuda",
        compute_type="int8_float16"
)
outputs = model.generate(
    text=["I like soccer", "I like tennis", "The eiffel tower is in Paris"],
    max_length=64,
) # perform downstream tasks on outputs
outputs["pooler_output"]
outputs["last_hidden_state"]
outputs["attention_mask"]

# alternative, use SentenceTransformer Mix-In
# for end-to-end Sentence embeddings generation
# (not pulling from this CT2fast-HF repo)

from hf_hub_ctranslate2 import CT2SentenceTransformer
model = CT2SentenceTransformer(
    model_name_orig, compute_type="int8_float16", device="cuda"
)
embeddings = model.encode(
    ["I like soccer", "I like tennis", "The eiffel tower is in Paris"],
    batch_size=32,
    convert_to_numpy=True,
    normalize_embeddings=True,
)
print(embeddings.shape, embeddings)
scores = (embeddings @ embeddings.T) * 100

# Hint: you can also host this code via REST API and
# via github.com/michaelfeil/infinity  

Checkpoint compatible to ctranslate2>=3.17.1 and hf-hub-ctranslate2>=2.12.0

• compute_type=int8_float16 for device="cuda"
• compute_type=int8 for device="cpu"

Converted on 2023-10-13 using

LLama-2 -> removed <pad> token.

Licence and other remarks:

This is just a quantized version. Licence conditions are intended to be idential to original huggingface repo.

Original description

E5-small-v2

Text Embeddings by Weakly-Supervised Contrastive Pre-training. Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, Furu Wei, arXiv 2022

This model has 12 layers and the embedding size is 384.

Usage

Below is an example to encode queries and passages from the MS-MARCO passage ranking dataset.

import torch.nn.functional as F

from torch import Tensor
from transformers import AutoTokenizer, AutoModel


def average_pool(last_hidden_states: Tensor,
                 attention_mask: Tensor) -> Tensor:
    last_hidden = last_hidden_states.masked_fill(~attention_mask[..., None].bool(), 0.0)
    return last_hidden.sum(dim=1) / attention_mask.sum(dim=1)[..., None]


# Each input text should start with "query: " or "passage: ".
# For tasks other than retrieval, you can simply use the "query: " prefix.
input_texts = ['query: how much protein should a female eat',
               'query: summit define',
               "passage: As a general guideline, the CDC's average requirement of protein for women ages 19 to 70 is 46 grams per day. But, as you can see from this chart, you'll need to increase that if you're expecting or training for a marathon. Check out the chart below to see how much protein you should be eating each day.",
               "passage: Definition of summit for English Language Learners. : 1  the highest point of a mountain : the top of a mountain. : 2  the highest level. : 3  a meeting or series of meetings between the leaders of two or more governments."]

tokenizer = AutoTokenizer.from_pretrained('intfloat/e5-small-v2')
model = AutoModel.from_pretrained('intfloat/e5-small-v2')

# Tokenize the input texts
batch_dict = tokenizer(input_texts, max_length=512, padding=True, truncation=True, return_tensors='pt')

outputs = model(**batch_dict)
embeddings = average_pool(outputs.last_hidden_state, batch_dict['attention_mask'])

# normalize embeddings
embeddings = F.normalize(embeddings, p=2, dim=1)
scores = (embeddings[:2] @ embeddings[2:].T) * 100
print(scores.tolist())

Training Details

Please refer to our paper at https://arxiv.org/pdf/2212.03533.pdf.

Benchmark Evaluation

Check out unilm/e5 to reproduce evaluation results on the BEIR and MTEB benchmark.

Support for Sentence Transformers

Below is an example for usage with sentence_transformers.

from sentence_transformers import SentenceTransformer
model = SentenceTransformer('intfloat/e5-small-v2')
input_texts = [
    'query: how much protein should a female eat',
    'query: summit define',
    "passage: As a general guideline, the CDC's average requirement of protein for women ages 19 to 70 is 46 grams per day. But, as you can see from this chart, you'll need to increase that if you're expecting or training for a marathon. Check out the chart below to see how much protein you should be eating each day.",
    "passage: Definition of summit for English Language Learners. : 1  the highest point of a mountain : the top of a mountain. : 2  the highest level. : 3  a meeting or series of meetings between the leaders of two or more governments."
]
embeddings = model.encode(input_texts, normalize_embeddings=True)

Package requirements

pip install sentence_transformers~=2.2.2

Contributors: michaelfeil

FAQ

1. Do I need to add the prefix "query: " and "passage: " to input texts?

Yes, this is how the model is trained, otherwise you will see a performance degradation.

Here are some rules of thumb:

• Use "query: " and "passage: " correspondingly for asymmetric tasks such as passage retrieval in open QA, ad-hoc information retrieval.

• Use "query: " prefix for symmetric tasks such as semantic similarity, paraphrase retrieval.

• Use "query: " prefix if you want to use embeddings as features, such as linear probing classification, clustering.

2. Why are my reproduced results slightly different from reported in the model card?

Different versions of transformers and pytorch could cause negligible but non-zero performance differences.

3. Why does the cosine similarity scores distribute around 0.7 to 1.0?

This is a known and expected behavior as we use a low temperature 0.01 for InfoNCE contrastive loss.

For text embedding tasks like text retrieval or semantic similarity, what matters is the relative order of the scores instead of the absolute values, so this should not be an issue.

Citation

If you find our paper or models helpful, please consider cite as follows:

@article{wang2022text,
  title={Text Embeddings by Weakly-Supervised Contrastive Pre-training},
  author={Wang, Liang and Yang, Nan and Huang, Xiaolong and Jiao, Binxing and Yang, Linjun and Jiang, Daxin and Majumder, Rangan and Wei, Furu},
  journal={arXiv preprint arXiv:2212.03533},
  year={2022}
}

Limitations

This model only works for English texts. Long texts will be truncated to at most 512 tokens.