flan-ul2-text-encoder

The encoder model extracted from flan-ul2 via a new class add in a recent release.

⚠️ This model is 17.44 GB in bfloat16 precision ⚠️

basic usage

from transformers import AutoTokenizer, AutoModelForTextEncoding

tokenizer = AutoTokenizer.from_pretrained("pszemraj/flan-ul2-text-encoder")
model = AutoModelForTextEncoding.from_pretrained("pszemraj/flan-ul2-text-encoder")

inputs = tokenizer("Hello, my dog loves memes", return_tensors="pt")
outputs = model(**inputs)

last_hidden_states = outputs.last_hidden_state

usage: semantic similarity

note: this is 'one way' to use the encoder, not 'the only way'. suggestions and ideas welcome.

Below is an example and a set of functions to compute the cosine similarity between the embeddings of different texts with this model

Functions

load_model_and_tokenizer

Loads the model and tokenizer based on model_name, returning a tuple containing the loaded model and tokenizer.

Details

from typing import List, Tuple

import torch
from transformers import AutoModel, AutoTokenizer
from transformers import AutoModelForTextEncoding


def load_model_and_tokenizer(model_name: str) -> Tuple[AutoModel, AutoTokenizer]:
    """
    Load the model and tokenizer based on the given model name.

    Args:
        model_name (str): The name of the model to be loaded.

    Returns:
        Tuple[AutoModelForTextEncoding, AutoTokenizer]: The loaded model and tokenizer.
    """
    model = AutoModelForTextEncoding.from_pretrained(
        model_name, torch_dtype=torch.bfloat16, device_map="auto"
    ).eval()
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    return model, tokenizer

get_embeddings

This computes the embeddings for the given texts given the model and tokenizer via weighted mean pooling across seq_len (as in SGPT)

Details

def get_embeddings(
    model: AutoModel, tokenizer: AutoTokenizer, texts: List[str]
) -> torch.Tensor:
    """
    compute text embeddings via weighted mean pooling across seq_len

    Args:
        model (AutoModel): The model to be used for getting embeddings.
        tokenizer (AutoTokenizer): The tokenizer to be used for tokenizing the texts.
        texts (List[str]): The texts for which embeddings are to be calculated.

    Returns:
        torch.Tensor: The calculated embeddings.
    """
    # Tokenize input texts
    batch_tokens = tokenizer(texts, padding=True, truncation=True, return_tensors="pt")

    # Get the embeddings
    with torch.no_grad():
        last_hidden_state = model(
            **batch_tokens, output_hidden_states=True, return_dict=True
        ).last_hidden_state

    # Get weights
    weights = (
        torch.arange(start=1, end=last_hidden_state.shape[1] + 1)
        .unsqueeze(0)
        .unsqueeze(-1)
        .expand(last_hidden_state.size())
        .float()
        .to(last_hidden_state.device)
    )

    # Get attn mask
    input_mask_expanded = (
        batch_tokens["attention_mask"]
        .unsqueeze(-1)
        .expand(last_hidden_state.size())
        .float()
    )

    # Perform weighted mean pooling across seq_len: bs, seq_len, hidden_dim -> bs, hidden_dim
    sum_embeddings = torch.sum(last_hidden_state * input_mask_expanded * weights, dim=1)
    sum_mask = torch.sum(input_mask_expanded * weights, dim=1)

    embeddings = sum_embeddings / sum_mask

    return embeddings

calculate_cosine_similarity

Helper fn to compute and print out cosine similarity

click to expand

from scipy.spatial.distance import cosine

def calculate_cosine_similarity(embeddings: torch.Tensor, texts: List[str]) -> None:
    """compute and print the cosine sim between the first text and all others"""
    # Calculate cosine similarities
    for i in range(1, len(embeddings)):
        cosine_sim = 1 - cosine(embeddings[0], embeddings[i])
        print(
            'Cosine similarity between "%s" and "%s" is: %.3f'
            % (texts[0], texts[i], cosine_sim)
        )

Usage

Install packages:

pip install transformers accelerate sentencepiece scipy

Then, you can use the functions to compute embeddings and similarity scores:

model_name = "pszemraj/flan-ul2-text-encoder"
model, tokenizer = load_model_and_tokenizer(model_name)

texts = [
    "deep learning",
    "artificial intelligence",
    "deep diving",
    "artificial snow",
]

embeddings = get_embeddings(model, tokenizer, texts)
calculate_cosine_similarity(embeddings, texts)

This will print the cosine similarity between the first text and all other texts in the `texts' list.

References

Inference with this model/the example is based on the ideas and examples in the SGPT repository.

@article{muennighoff2022sgpt,
  title={SGPT: GPT Sentence Embeddings for Semantic Search},
  author={Muennighoff, Niklas},
  journal={arXiv preprint arXiv:2202.08904},
  year={2022}
}