Mbedder 0.0.4

Mbedder: A pytorch powered framework for seemlessly adding contextual text embeddings from pretrained models

---

Mbedder is a language framework for adding contextual embeddings of pretrained language models to deep learning models.Mbedder is powered by PyTorch and HuggingFace and requires as less as 1 line of code to add embeddings and works similar to how the Embedding Layer works in PyTorch.

List of supported architectures

• Bert
• XLNet
• Albert
• TransfoXL
• DistilBert
• Roberta
• XLM
• XLMRoberta
• GPT
• GPT2
• Flaubert

The pretrained models for the mentioned architecures can be found here.

Features

• Addition of embeddings with 1 line of code
• Embeddings can output Sentence as well as Token level embeddings
• Task specific combination strategies can be applied to hidden states and token embeddings
• Custom pre-trained hugging face transformer models can be used with Mbedder.

Requirements and Installation

• PyTorch version >= 1.6.0
• Python version >= 3.6
• Transformer >= 3.0.2

Mbedder can be using Pip as follows

pip install Mbedder

Getting Started

A basic example of using a Mbedder Bert embedding is shown below:

import torch
from Mbedder import BertEmbedding

class BertClassifier(torch.nn.Module):
    def __init__(self, num_classes):
        super(BertClassifier, self).__init__()
        self.embedding = BertEmbedding.from_pretrained('bert-base-uncased')
        self.fc = torch.nn.Linear(self.embedding.embedding_size, num_classes)
    
    def forward(self, input_ids, attention_mask):
        x = self.embedding(input_ids, attention_mask, output_token_embeddings=False)
        logits = self.fc(x[0])
        return logits

More advanced examples can be found in the examples folder.

License

Mbedder is MIT-licensed.