Simple Representations

This library is based on the Transformers library by HuggingFace. Using this library, you can quickly extract text representations from Transformer models. Only two lines of code are needed to initialize the required model and extract the text representations from it.

Table of contents

• Installation
  • With pip
  • From source
• Usage
  • Minimal Start
  • Default Settings
  • Current Pretrained Models
• Acknowledgements

Installation

This repository is tested on Python 3.8.13 and PyTorch 1.7.1

With pip

First you need to install PyTorch. Please refer to PyTorch installation page regarding the specific install command for your platform.

When PyTorch has been installed, Simple Representation can be installed using pip as follows:

pip install simplerepresentations

From source

Here also, you first need to install PyTorch. Please refer to PyTorch installation page regarding the specific install command for your platform.

When PyTorch has been installed, you can install from source by cloning the repository and running:

pip install .

Usage

Minimal Start

The following example extracts the text representations from BERT Base Uncased model for the sentences Hello Transformers! and It's very simple..

from simplerepresentations import RepresentationModel


def load_data():
	return ['Hello Transformers!', 'It\'s very simple.']


if __name__ == '__main__':
	model_type = 'bert'
	model_name = 'bert-base-uncased'

	representation_model = RepresentationModel(
		model_type=model_type,
		model_name=model_name,
		batch_size=32,
		max_seq_length=10, # truncate sentences to be less than or equal to 10 tokens
		combination_method='cat', # concatenate the last `last_hidden_to_use` hidden states
		last_hidden_to_use=4 # use the last 4 hidden states to build tokens representations
	)

	text_a = load_data()

	all_sentences_representations, all_tokens_representations = representation_model(text_a=text_a)

	print(all_sentences_representations.shape) # (2, 768) => (number of sentences, hidden size)
	print(all_tokens_representations.shape) # (2, 10, 3072) => (number of sentences, number of tokens, hidden size)

You can change the code in load_data function to load your own data from any source you want (e.g. a CSV file).

Default Settings

The default settings for RepresentationModel class are given below:

batch_size (32): integer

The batch size will be used while extracting representations.

max_seq_length (128): integer

Maximum sequence length the model will support.

last_hidden_to_use (1): integer

The number of the last hidden states that will be used to build the representations.

combination_method ('sum'): string ('sum', 'cat')

The method that will be used to combine the last_hidden_to_use.

use_cuda (True): boolean

Whether to use CUDA or not.

process_count (cpu_count() - 2 if cpu_count() > 2 else 1): integer

Number of CPU cores (processes) to use when converting examples to features. Default is (number of cores - 2) or 1 if (number of cores <= 2).

chunksize (500): integer

The number of chunks that the examples will be divided to when converting them to features.

Current Pretrained Models

You can find the complete list of the current pretrained models from Transformers library documentation.

Acknowledgements

None of this would have been possible without the hard work by the HuggingFace team in developing the Transformers library.

Also, a lot of ideas used in this repository inspired from the Simple Transformers library.