Skip to content

Latest commit

 

History

History
87 lines (71 loc) · 3.13 KB

DETAILS.md

File metadata and controls

87 lines (71 loc) · 3.13 KB

Details

Folder structure

  • __main__.py is the entry point for the program.
  • __params__.py contains the hyperparameters, configuration and command line arguments.
  • src is where the source code is stored.
    • data.py contains the data loading and preprocessing functions.
    • train.py contains the training loop.
    • eval.py contains the evaluation functions.
    • pred.py contains the prediction functions.
    • visualize.py contains the visualizations.
    • crawl contains the web crawling functions (bluesky, twitter, youtube).
  • res is where the resources are stored.
  • .out is where the generated files are stored.
  • .devcontainer stores a VSCode devcontainer configuration.
  • requirements.txt lists the required packages with versions.

Setup

pip install -r requirements.txt

Data crawling

python climate-opinions --crawl=<platform> \
  --api_key=<api-key> \
  --query=<query> \
  --results=<results-directory>

Defaults are configured in __params__.py.

  • 'bluesky|twitter|youtube' (platform selection)
  • api key for youtube.
  • query (default is "global warming", "climate crisis", "climate emergency", "global heating", "climate change").
  • results directory to store the crawled posts.

Model training, prediction, and visualization

Hyperparameters and configuration

python climate-opinions --model=<model> \
  --epochs <epochs> \
  --batch_size <batch_size> \
  --sample
  --results=<results-directory>

Defaults can be configured in __params__.py.

  • 'baseline|blank|sentiment' (model selection).
  • epochs and batch size for training (defaults are 10 and 32).
  • seed for random number generation (for reproducibility) (default: 42).
  • results directory to store the best model, evaluation results and predictions.

Sample mode

When developing, you can use the --sample flag to only use a small subset of the data to see if everything works before starting long training processes.

Data Preprocessing

  • news label is removed.
  • text is lowercased.
  • links are removed.
  • labels are shifted to 0, 1, 2 (easier to work with non-negative numbers)

Model

  • Blank and baseline are based on huggingface BERT-base-uncased.
  • Sentiment model is based on BERTweet (BERTweet is a BERT model pre-trained on a large corpus of English tweets).

Training

  • epochs/batch size can be configured via command line arguments or in code.
  • best model is selected by (validation) loss and saved.

Evaluation

  • best model is loaded.
  • evaluation will be run if training is interrupted.
  • evaluation metrics are: accuracy, precision/recall/f1 for each class (negative=0, neutral=1, positive=2).
  • evaluation results are stored in csv with model name and metadata.

Prediction

  • best model is loaded.
  • prediction is done on the crawled data.
  • prediction results are stored in csv for selected model.

Visualization

  • visualize share of climate change opponents over time.
  • …

README