Text Embedding with Transformers

author: Jael Gu

Description

A text embedding operator takes a sentence, paragraph, or document in string as an input and outputs token embeddings which captures the input's core semantic elements. This operator is implemented with pre-trained models from Huggingface Transformers.

Code Example

Use the pre-trained model 'distilbert-base-cased' to generate a text embedding for the sentence "Hello, world.".

Write a pipeline with explicit inputs/outputs name specifications:

from towhee import pipe, ops, DataCollection

p = (
    pipe.input('text')
        .map('text', 'vec', ops.text_embedding.transformers(model_name='distilbert-base-cased'))
        .output('text', 'vec')
)

DataCollection(p('Hello, world.')).show()

Factory Constructor

Create the operator via the following factory method:

text_embedding.transformers(model_name=None)

Parameters:

model_name: str

The model name in string, defaults to None. If None, the operator will be initialized without specified model.

Please note only supported models are tested by us:

Albert

• albert-base-v1
• albert-large-v1
• albert-xlarge-v1
• albert-xxlarge-v1
• albert-base-v2
• albert-large-v2
• albert-xlarge-v2
• albert-xxlarge-v2

Bart

• facebook/bart-large

Bert

• bert-base-cased
• bert-base-uncased
• bert-large-cased
• bert-large-uncased
• bert-base-multilingual-uncased
• bert-base-multilingual-cased
• bert-base-chinese
• bert-base-german-cased
• bert-large-uncased-whole-word-masking
• bert-large-cased-whole-word-masking
• bert-large-uncased-whole-word-masking-finetuned-squad
• bert-large-cased-whole-word-masking-finetuned-squad
• bert-base-cased-finetuned-mrpc
• bert-base-german-dbmdz-cased
• bert-base-german-dbmdz-uncased
• cl-tohoku/bert-base-japanese-whole-word-masking
• cl-tohoku/bert-base-japanese-char
• cl-tohoku/bert-base-japanese-char-whole-word-masking
• TurkuNLP/bert-base-finnish-cased-v1
• TurkuNLP/bert-base-finnish-uncased-v1
• wietsedv/bert-base-dutch-cased

BertGeneration

• google/bert_for_seq_generation_L-24_bbc_encoder

BigBird

• google/bigbird-roberta-base
• google/bigbird-roberta-large
• google/bigbird-base-trivia-itc

BigBirdPegasus

• google/bigbird-pegasus-large-arxiv
• google/bigbird-pegasus-large-pubmed
• google/bigbird-pegasus-large-bigpatent

CamemBert

• camembert-base
• Musixmatch/umberto-commoncrawl-cased-v1
• Musixmatch/umberto-wikipedia-uncased-v1

Canine

• google/canine-s
• google/canine-c

Convbert

• YituTech/conv-bert-base
• YituTech/conv-bert-medium-small
• YituTech/conv-bert-small

CTRL

- ctrl

DeBERTa

• microsoft/deberta-base
• microsoft/deberta-large
• microsoft/deberta-xlarge
• microsoft/deberta-base-mnli
• microsoft/deberta-large-mnli
• microsoft/deberta-xlarge-mnli
• microsoft/deberta-v2-xlarge
• microsoft/deberta-v2-xxlarge
• microsoft/deberta-v2-xlarge-mnli
• microsoft/deberta-v2-xxlarge-mnli

DistilBert

• distilbert-base-uncased
• distilbert-base-uncased-distilled-squad
• distilbert-base-cased
• distilbert-base-cased-distilled-squad
• distilbert-base-german-cased
• distilbert-base-multilingual-cased
• distilbert-base-uncased-finetuned-sst-2-english

Electral

• google/electra-small-generator
• google/electra-base-generator
• google/electra-large-generator
• google/electra-small-discriminator
• google/electra-base-discriminator
• google/electra-large-discriminator

Flaubert

• flaubert/flaubert_small_cased
• flaubert/flaubert_base_uncased
• flaubert/flaubert_base_cased
• flaubert/flaubert_large_cased

FNet

• google/fnet-base
• google/fnet-large

FSMT

• facebook/wmt19-ru-en

Funnel

• funnel-transformer/small
• funnel-transformer/small-base
• funnel-transformer/medium
• funnel-transformer/medium-base
• funnel-transformer/intermediate
• funnel-transformer/intermediate-base
• funnel-transformer/large
• funnel-transformer/large-base
• funnel-transformer/xlarge-base
• funnel-transformer/xlarge

GPT

• openai-gpt
• gpt2
• gpt2-medium
• gpt2-large
• gpt2-xl
• distilgpt2
• EleutherAI/gpt-neo-1.3B
• EleutherAI/gpt-j-6B

I-Bert

• kssteven/ibert-roberta-base

LED

• allenai/led-base-16384

MobileBert

• google/mobilebert-uncased

MPNet

• microsoft/mpnet-base

Nystromformer

• uw-madison/nystromformer-512

Reformer

• google/reformer-crime-and-punishment

Splinter

• tau/splinter-base
• tau/splinter-base-qass
• tau/splinter-large
• tau/splinter-large-qass

SqueezeBert

• squeezebert/squeezebert-uncased
• squeezebert/squeezebert-mnli
• squeezebert/squeezebert-mnli-headless

TransfoXL

• transfo-xl-wt103

XLM

• xlm-mlm-en-2048
• xlm-mlm-ende-1024
• xlm-mlm-enfr-1024
• xlm-mlm-enro-1024
• xlm-mlm-tlm-xnli15-1024
• xlm-mlm-xnli15-1024
• xlm-clm-enfr-1024
• xlm-clm-ende-1024
• xlm-mlm-17-1280
• xlm-mlm-100-1280

XLMRoberta

• xlm-roberta-base
• xlm-roberta-large
• xlm-roberta-large-finetuned-conll02-dutch
• xlm-roberta-large-finetuned-conll02-spanish
• xlm-roberta-large-finetuned-conll03-english
• xlm-roberta-large-finetuned-conll03-german

XLNet

• xlnet-base-cased
• xlnet-large-cased

Yoso

• uw-madison/yoso-4096

checkpoint_path: str

The path to local checkpoint, defaults to None. If None, the operator will download and load pretrained model by model_name from Huggingface transformers.

device: str

The device in string, defaults to None. If None, it will enable "cuda" automatically when cuda is available.

tokenizer: object

The method to tokenize input text, defaults to None. If None, the operator will use default tokenizer by model_name from Huggingface transformers.

Interface

The operator takes a piece of text in string as input. It loads tokenizer and pre-trained model using model name. and then return text embedding(s) in ndarray.

__call__(txt)

Parameters:

data: Union[str, list]

​ The text in string or a list of texts. If data is string, the operator returns token embedding(s) in ndarray. If data is a list, the operator returns token embedding(s) in a list.

Returns:

numpy.ndarray or list

​ The text embedding (or token embeddings) extracted by model.

save_model(format='pytorch', path='default')

Save model to local with specified format.

Parameters:

format: str

​ The format of saved model, defaults to 'pytorch'.

path: str

​ The path where model is saved to. By default, it will save model to the operator directory.

from towhee import ops

op = ops.text_embedding.transformers(model_name='distilbert-base-cased').get_op()
op.save_model('onnx', 'test.onnx')

PosixPath('/Home/.towhee/operators/text-embedding/transformers/main/test.onnx')

supported_model_names(format=None)

Get a list of all supported model names or supported model names for specified model format.

Parameters:

format: str

​ The model format such as 'pytorch', 'torchscript'.

from towhee import ops


op = ops.text_embedding.transformers().get_op()
full_list = op.supported_model_names()
onnx_list = op.supported_model_names(format='onnx')
print(f'Onnx-support/Total Models: {len(onnx_list)}/{len(full_list)}')

2022-12-13 16:25:15,916 - 140704500614336 - auto_transformers.py-auto_transformers:68 - WARNING: The operator is initialized without specified model.
Onnx-support/Total Models: 111/126

Fine-tune

Requirement

If you want to train this operator, besides dependency in requirements.txt, you need install these dependencies.

! python -m pip install datasets evaluate scikit-learn

Get start

We have prepared some most typical use of finetune examples.

Simply speaking, you only need to construct an op instance and pass in some configurations to train the specified task.

import towhee

bert_op = towhee.ops.text_embedding.transformers(model_name='bert-base-uncased').get_op()
data_args = {
    'dataset_name': 'wikitext',
    'dataset_config_name': 'wikitext-2-raw-v1',
}
training_args = {
    'num_train_epochs': 3, # you can add epoch number to get a better metric.
    'per_device_train_batch_size': 8,
    'per_device_eval_batch_size': 8,
    'do_train': True,
    'do_eval': True,
    'output_dir': './tmp/test-mlm',
    'overwrite_output_dir': True
}
bert_op.train(task='mlm', data_args=data_args, training_args=training_args)

For more infos, refer to the examples.

Dive deep and customize your training

You can change the training script in your customer way. Or your can refer to the original hugging face transformers training examples.

More Resources

• The guide to text-embedding-ada-002 model | OpenAI: text-embedding-ada-002: OpenAI's legacy text embedding model; average price/performance compared to text-embedding-3-large and text-embedding-3-small.
• Sentence Transformers for Long-Form Text - Zilliz blog: Deep diving into modern transformer-based embeddings for long-form text.
• Massive Text Embedding Benchmark (MTEB): A standardized way to evaluate text embedding models across a range of tasks and languages, leading to better text embedding models for your app
• Training Your Own Text Embedding Model - Zilliz blog: Explore how to train your text embedding model using the sentence-transformers library and generate our training data by leveraging a pre-trained LLM.
• Tutorial: Diving into Text Embedding Models | Zilliz Webinar: Register for a free webinar diving into text embedding models in a presentation and tutorial
• Tutorial: Diving into Text Embedding Models | Zilliz Webinar: Register for a free webinar diving into text embedding models in a presentation and tutorial
• The guide to jina-embeddings-v2-base-en | Jina AI: jina-embeddings-v2-base-en: specialized embedding model for English text and long documents; support sequences of up to 8192 tokens
• Evaluating Your Embedding Model - Zilliz blog: Review some practical examples to evaluate different text embedding models.
• Training Text Embeddings with Jina AI - Zilliz blog: In a recent talk by Bo Wang, he discussed the creation of Jina text embeddings for modern vector search and RAG systems. He also shared methodologies for training embedding models that effectively encode extensive information, along with guidance o