textsum

a utility for using transformers summarization models on text docs 🖇

This package provides easy-to-use interfaces for using summarization models on text documents of arbitrary length. Currently implemented interfaces include a python API, CLI, and a shareable demo app.

Tip

For additional details, explanations, and docs, see the wiki

textsum

🔦 Quick Start Guide

Install the package with pip:

pip install textsum

Import the package and create a summarizer:

from textsum.summarize import Summarizer
summarizer = Summarizer() # loads default model and parameters

Summarize a text string:

text = "This is a long string of text that will be summarized."
summary = summarizer.summarize_string(text)
print(f'Summary: {summary}')

Installation

Install using pip with Python 3.8 or later (after creating a virtual environment):

pip install textsum

The textsum package is now installed in your virtual environment. CLI commands are available in your terminal, and the python API is available in your python environment.

Full Installation

For a full installation, which includes additional features such as PDF OCR, Gradio UI demo, and Optimum, run the following commands:

git clone https://github.com/pszemraj/textsum.git
cd textsum
# create a virtual environment (optional)
pip install -e .[all]

Extra Features

The package also supports a number of optional extra features, which can be installed as follows:

8bit: Install with pip install -e "textsum[8bit]"
optimum: Install with pip install -e "textsum[optimum]"
PDF: Install with pip install -e "textsum[PDF]"
app: Install with pip install -e "textsum[app]"
unidecode: Install with pip install -e "textsum[unidecode]"

Replace textsum in the command with . if installing from source. Read below for more details on how to use these features.

Tip

The unidecode extra is a GPL-licensed dependency not included by default with the clean-text package. Installing it should improve the cleaning of noisy input text, but it should not make a significant difference in most use cases.

Usage

There are three ways to use this package:

python API
CLI
Demo App

Python API

To use the python API, import the Summarizer class and instantiate it. This will load the default model and parameters.

You can then use the summarize_string method to summarize a long text string.

from textsum.summarize import Summarizer

summarizer = Summarizer() # loads default model and parameters

# summarize a long string
out_str = summarizer.summarize_string('This is a long string of text that will be summarized.')
print(f'summary: {out_str}')

you can also directly summarize a file:

out_path = summarizer.summarize_file('/path/to/file.txt')
print(f'summary saved to {out_path}')

CLI

To summarize a directory of text files, run the following command in your terminal:

textsum-dir /path/to/dir

There are many CLI flags available. A full list:

Click to expand table

Flag	Description
`--output_dir`	Specify the output directory
`--model`	Specify the model to use
`--no_cuda`	Disable CUDA
`--tf32`	Use TF32 precision
`--force_cache`	Force cache usage
`--load_in_8bit`	Load in 8-bit mode
`--compile`	Compile the model
`--optimum_onnx`	Use optimum ONNX
`--batch_length`	Specify the batch length
`--batch_stride`	Specify the batch stride
`--num_beams`	Specify the number of beams
`--length_penalty`	Specify the length penalty
`--repetition_penalty`	Specify the repetition penalty
`--max_length_ratio`	Specify the maximum length ratio
`--min_length`	Specify the minimum length
`--encoder_no_repeat_ngram_size`	Specify the encoder no repeat ngram size
`--no_repeat_ngram_size`	Specify the no repeat ngram size
`--early_stopping`	Enable early stopping
`--shuffle`	Shuffle the input data
`--lowercase`	Convert input to lowercase
`--loglevel`	Specify the log level
`--logfile`	Specify the log file
`--file_extension`	Specify the file extension
`--skip_completed`	Skip completed files

Some useful options are:

Arguments:

--model: model name or path to use for summarization. (Optional)
--shuffle: Shuffle the input files before processing. (Optional)
--skip_completed: Skip already completed files in the output directory. (Optional)
--batch_length: The maximum length of each input batch. Default is 4096. (Optional)
--output_dir: The directory to write the summarized output files. Default is ./summarized/. (Optional)

To see all available options, run the following command:

textsum-dir --help

Demo App

For convenience, a UI demo¹ is provided using gradio. To ensure you have the dependencies installed, run the following command:

pip install textsum[app]

To launch the demo, run:

textsum-ui

This will start a local server that you can access in your browser & a shareable link will be printed to the console.

Models

Summarization is a memory-intensive task, and the default model is relatively small and efficient for long-form text summarization. If you want to use a different model, you can specify the model_name_or_path argument when instantiating the Summarizer class.

summarizer = Summarizer(model_name_or_path='pszemraj/long-t5-tglobal-xl-16384-book-summary')

You can also use the -m argument when using the CLI:

textsum-dir /path/to/dir -m pszemraj/long-t5-tglobal-xl-16384-book-summary

Any text-to-text or summarization model from the HuggingFace model hub can be used. Models are automatically downloaded and cached in ~/.cache/huggingface/hub.

Advanced Configuration

Parameters

Memory usage can also be reduced by adjusting the parameters for inference. This is discussed in detail in the project wiki.

Important

tl;dr: use the summarizer.set_inference_params() and summarizer.get_inference_params() methods to adjust the inference parameters, passing either a python dict or a JSON file.

Support for GenerationConfig as the primary method to adjust inference parameters is planned for a future release.

8-bit Quantization & TensorFloat32

Some methods of efficient inference² include loading the model in 8-bit precision via LLM.int8 (reduces memory usage) and enabling TensorFloat32 precision in the torch backend (reduces latency). See the transformers docs for more details. Using LLM.int8 requires the bitsandbytes package, which can either be installed directly or via the textsum[8bit] extra:

pip install textsum[8bit]

To use these options, use the --load_in_8bit and --tf32 flags when using the CLI:

textsum-dir /path/to/dir --load_in_8bit --tf32

Or in Python, using the load_in_8bit argument:

summarizer = Summarizer(load_in_8bit=True)

If using the Python API, either manually activate tf32 or use the check_ampere_gpu() function from textsum.utils before initializing the Summarizer class:

from textsum.utils import check_ampere_gpu
check_ampere_gpu() # automatically enables TF32 if Ampere+ available
summarizer = Summarizer(load_in_8bit=True)

Using Optimum ONNX Runtime

Caution

This feature is experimental and might not work as expected. Use at your own risk. ⚠️🧪

ONNX Runtime is a performance-oriented inference engine for ONNX models. It can be used to increase the speed of model inference, especially on Windows and in environments where GPU acceleration is not available. If you want to use ONNX runtime for inference, you need to set optimum_onnx=True when initializing the Summarizer class.

First, install with pip install textsum[optimum]. Then initialize the Summarizer class with ONNX runtime:

summarizer = Summarizer(model_name_or_Path="onnx-compatible-model-name", optimum_onnx=True)

It will automatically convert the model if it has not been converted to ONNX yet.

Notes:

ONNX runtime+cuda needs an additional package. Manually install onnxruntime-gpu if you plan to use ONNX with GPU.
Using ONNX runtime might lead to different behavior in certain models. It is recommended to test the model with and without ONNX runtime the same input text before using it for anything important.

Force Cache

Caution

Setting force_cache=True might lead to different behavior in certain models. Test the model with and without force_cache on the same input text before using it for anything important.

Using the cache speeds up autoregressive generation by avoiding recomputing attention for tokens that have already been generated. If you want to force the model to always use cache irrespective of the model's default behavior³, you can set force_cache=True when initializing the Summarizer class.

summarizer = Summarizer(force_cache=True)

Compile Model

If you want to compile the model for faster inference times, you can set compile_model=True when initializing the Summarizer class.

summarizer = Summarizer(compile_model=True)

Note

Compiling the model might not be supported on all platforms and requires pytorch > 2.0.0.

Contributing

Contributions are welcome! Please open an issue or PR if you have any ideas or suggestions.

See the CONTRIBUTING.md file for details on how to contribute.

Road Map

add CLI for summarization of all text files in a directory
python API for summarization of text docs
add argparse CLI for UI demo
put on PyPI
LLM.int8 inference
optimum inference integration
better documentation in the wiki, details on improving performance (speed, quality, memory usage, etc.)
- in-progress
improvements to the PDF OCR helper module (TBD - may focus more on being a summarization tool)

Other ideas? Open an issue or PR!

The demo is minimal but will be expanded to accept other arguments and options. ↩
if you have compatible hardware. In general, ampere (RTX 30XX) and newer GPUs are recommended. ↩
use_cache can sometimes be disabled due to things like gradient accumulation training, etc., and if not re-enabled will result in slower inference times. ↩

pszemraj/textsum