import torch
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]

def get_position_ids(input_ids: Tensor, max_original_positions: int=512, encode_max_length: int=4096) -> Tensor:

    position_ids = list(range(input_ids.size(1)))
    factor = max(encode_max_length // max_original_positions, 1)
    if input_ids.size(1) <= max_original_positions:
        position_ids = [(pid * factor) for pid in position_ids]
        
    position_ids = torch.tensor(position_ids, dtype=torch.long)
    position_ids = position_ids.unsqueeze(0).expand_as(input_ids)
    
    return position_ids

# 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('dwzhu/e5-base-4k')
model = AutoModel.from_pretrained('dwzhu/e5-base-4k')

# Tokenize the input texts
batch_dict = tokenizer(input_texts, max_length=4096, padding=True, truncation=True, return_tensors='pt')
batch_dict['position_ids'] = get_position_ids(batch_dict['input_ids'], max_original_positions=512, encode_max_length=4096)

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/abs/2404.12096.pdf. Note that E5-Base-4k simply expands the position embedding matrix to allow for 4,096 position ids. The embedding vectors for the original pids {0,1,2,...,511} is mapped to represent {0,8,16,...,4088}. Embedding vectors for other pids are trained. So for inputs not exceeding 512 tokens, please multiply the position ids by 8 to maintain the original behavior, as shown in the code above.

Benchmark Evaluation

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

Citation

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

@article{zhu2024longembed,
  title={LongEmbed: Extending Embedding Models for Long Context Retrieval},
  author={Zhu, Dawei and Wang, Liang and Yang, Nan and Song, Yifan and Wu, Wenhao and Wei, Furu and Li, Sujian},
  journal={arXiv preprint arXiv:2404.12096},
  year={2024}
}