metadata

license: mit
language:
  - vi
datasets:
  - anti-ai/ViNLI-Zalo-supervised
  - anti-ai/ViNLI-SimCSE-supervised_v2
  - anti-ai/Vietnamese_MsMacro
metrics:
  - repllabs/mean_reciprocal_rank
  - spearmanr
pipeline_tag: sentence-similarity
library_name: rage

Introduce

Installation 🔥

We recommend python 3.9 or higher, torch 2.0.0 or higher, transformers 4.31.0 or higher.
Currently, you can only download from the source, however, in the future, we will upload it to PyPI. RagE can be installed from source with the following commands:

git clone https://github.com/anti-aii/RagE.git
cd RagE
pip install -e .

Quick start 🥮

1. Initialize the model
2. Load model from Huggingface Hub
3. List of pretrained models

We have detailed instructions for using our models for inference. See notebook

1. Initialize the model

Let's initalize the SentenceEmbedding model

>>> import torch 
>>> from pyvi import ViTokenizer
>>> from rage import SentenceEmbedding
>>> device= torch.device('cuda' if torch.cuda.is_available() else 'cpu')
>>> model= SentenceEmbedding(model_name= "vinai/phobert-base-v2", torch_dtype= torch.float32, aggregation_hidden_states= False, strategy_pooling= "dense_first")
>>> model.to(device)
SentenceEmbeddingConfig(model_base: {'model_type_base': 'RobertaModel', 'model_name': 'vinai/phobert-base-v2', 'type_backbone': 'mlm', 'required_grad_base_model': True, 'aggregation_hidden_states': False, 'concat_embeddings': False, 'dropout': 0.1, 'quantization_config': None}, pooling: {'strategy_pooling': 'dense_first'})

Then, we can show the number of parameters in the model.

>>> model.summary_params()
trainable params: 135588864 || all params: 135588864 || trainable%: 100.0
>>> model.summary()
+---------------------------+-------------+------------------+
|        Layer (type)       |    Params   | Trainable params |
+---------------------------+-------------+------------------+
|    model (RobertaModel)   | 134,998,272 |    134998272     |
| pooling (PoolingStrategy) |   590,592   |      590592      |
|       drp1 (Dropout)      |      0      |        0         |
+---------------------------+-------------+------------------+

Now we can use the SentenceEmbedding model to encode the input words. The output of the model will be a matrix in the shape of (batch, dim). Additionally, we can load weights that we have previously trained and saved.

>>> model.load("best_sup_general_embedding_phobert2.pt", key= False)
>>> sentences= ["Tôi đang đi học", "Bạn tên là gì?",]
>>> sentences= list(map(lambda x: ViTokenizer.tokenize(x), sentences))
>>> model.encode(sentences, batch_size= 1, normalize_embedding= "l2", return_tensors= "np", verbose= 1)
2/2 [==============================] - 0s 43ms/Sample
array([[ 0.00281098, -0.00829096, -0.01582766, ...,  0.00878178,
         0.01830498, -0.00459659],
       [ 0.00249859, -0.03076724,  0.00033016, ...,  0.01299141,
        -0.00984358, -0.00703243]], dtype=float32)

2. Load model from Huggingface Hub

First, download a pretrained model.

>>> model= SentenceEmbedding.from_pretrained('anti-ai/VieSemantic-base')

Then, we encode the input sentences and compare their similarity.

>>> sentences = ["Nó rất thú_vị", "Nó không thú_vị ."]
>>> output= model.encode(sentences, batch_size= 1, return_tensors= 'pt')
>>> torch.cosine_similarity(output[0].view(1, -1), output[1].view(1, -1)).cpu().tolist()
2/2 [==============================] - 0s 40ms/Sample
[0.5605039596557617]

3. List of pretrained models

This list will be updated with our prominent models. Our models will primarily aim to support Vietnamese language. Additionally, you can access our datasets and pretrained models by visiting https://huggingface.co/anti-ai.

Model Name	Model Type	#params	checkpoint
anti-ai/ViEmbedding-base	SentenceEmbedding	135.5M	model
anti-ai/BioViEmbedding-base-unsup	SentenceEmbedding	135.5M	model
anti-ai/VieSemantic-base	SentenceEmbedding	135.5M	model

Contacts

If you have any questions about this repo, please contact me (nduc0231@gmail.com)