Skip to main content

Social Media NLP package for PyTorch & PyTorch Lightning.

Project description

A Social Media Natural Language Processing package for PyTorch & PyTorch Lightning.

Key FeaturesAbout MeHow To UseExamples

PyPI - Python Version PyPI Status license

PyTorch Gleam

PyTorch Gleam builds upon PyTorch Lightning for the specific use-case of Natural Language Processing on Social Media, such as Twitter. PyTorch Gleam strives to make Social Media NLP research easier to understand, use, and extend. Gleam contains models I use in my research, from fine-tuning a BERT-based model with Lexical, Emotion, and Semantic information in a Graph Attention Network for stance identification towards COVID-19 misinformation, to using Information Retrieval systems to identify new types of misinformation on Twitter.

About Me

My name is Maxwell Weinzierl, and I am a Natural Language Processing researcher at the Human Technology Research Institute (HLTRI) at the University of Texas at Dallas. I am currently working on my PhD, which focuses on COVID-19 and HPV vaccine misinformation, trust, and more on Social Media platforms such as Twitter. I have built PyTorch Gleam to enable easy reproducibility for my published research, and for my own quick iterations on research ideas.

How To Use

Step 0: Install

Simple installation from PyPI

pip install pytorch-gleam

You may need to install CUDA drivers and other versions of PyTorch. See PyTorch and PyTorch Lightning for installation help.

Step 1: Create Experiment

Create a configs folder with a YAML experiment file. Gleam utilizes PyTorch Lightning's CLI tools to configure experiments from YAML files, which enables researchers to clearly look back and identify both hyper-parameters and model code used in their experiments. This example is from COVID-19 vaccine misinformation stance identification:


seed_everything: 0
  class_path: pytorch_gleam.modeling.models.MultiClassFrameLanguageModel
    learning_rate: 5e-4
    pre_model_name: digitalepidemiologylab/covid-twitter-bert-v2
      No Stance: 0
      Accept: 1
      Reject: 2
      class_path: pytorch_gleam.modeling.thresholds.MultiClassThresholdModule
      class_path: pytorch_gleam.modeling.metrics.F1PRMultiClassMetric
        mode: macro
        num_classes: 3
  max_epochs: 10
  accumulate_grad_batches: 4
  check_val_every_n_epoch: 1
  deterministic: true
  num_sanity_val_steps: 1
  checkpoint_callback: false
    - class_path: pytorch_gleam.callbacks.FitCheckpointCallback
    batch_size: 8
    max_seq_len: 128
    label_name: misinfo
      No Stance: 0
      Accept: 1
      Reject: 2
    tokenizer_name: digitalepidemiologylab/covid-twitter-bert-v2
    num_workers: 8
      - covid19/misinfo.json
      - covid19/stance-train.jsonl
      - covid19/stance-dev.jsonl
      - covid19/stance-test.jsonl

Documentation on available models, datasets, and callbacks will be provided soon.

Details about how to set up YAML experiment files are provided by PyTorch Lightning's documentation.

Annotations for this example are provided in the VaccineLies repository under covid19 as the CoVaxLies collection: CoVaxLies. You will need to download the tweet texts from the tweet ids from the Twitter API.

Step 2: Run Experiment

Create a models folder for your saved TensorBoard logs and model weights. Determine the GPU ID for the GPU you would like to utilize (multi-gpu supported) and provide the ID in a list, with a comma at the end if it is a single GPU ID. You can also just specify an integer, such as 1, and PyTorch Lightning will try to find a single free GPU automatically. Run the following command to start training:

gleam fit \
  --config configs/covid-stance.yaml \
  --trainer.gpus 1 \
  --trainer.default_root_dir models/covid-stance

Your model will train, with TensorBoard logging all metrics, and a checkpoint will be saved upon completion.

Step 3: Evaluate Experiment

You can easily evaluate your system on a test collection as follows:

gleam test \
  --config configs/covid-stance.yaml \
  --trainer.gpus 1 \
  --trainer.default_root_dir models/covid-stance


These are a work-in-progress, as my original research code is a bit messy, but they will be updated soon!

COVID-19 Vaccine Misinformation Detection on Twitter
COVID-19 Vaccine Misinformation Stance Identification on Twitter
COVID-19 Misinformation Stance Identification on Twitter
Vaccine Misinformation Transfer Learning
Vaccine Hesitancy Profiling on Twitter
  • TODO

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

pytorch-gleam-0.6.2.tar.gz (3.7 MB view hashes)

Uploaded Source

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page