Project structure of: openai/weak-to-strong

• weak-to-strong Weak-to-strong learning experiments with language and vision tasks.
  • notebooks
    • Plotting.py Python script for data plotting and analysis.
  • pyproject.toml Python project using Hatch, depends on OpenAI's "weak_to_strong" with torch, numpy.
  • README.md Weak-to-strong learning setup for binary tasks, using pretrained language models and various loss functions.
  • setup.py Sets up Python package with setuptools
  • train_weak_to_strong.py Trains weak models to strong, compares performance, stores results.
  • vision Experiments weak-to-strong learning on ImageNet using various models with data pipeline and logistic regression.
    • data.py Image data pipeline for ImageNet dataset
    • models.py Deep learning models creation and modification.
    • README.md Experiments weak-to-strong learning on ImageNet with various models
    • run_weak_strong.py Trains logistic regression model with weak and strong labels, evaluates accuracy.
  • weak_to_strong Weak to strong contrastive learning for data augmentation.
    • common.py PyTorch tokenizer & memory clear functions
    • datasets.py Load, split, tokenize datasets in DatasetConfig class.
    • eval.py Evaluates model accuracy on dataset with batches.
    • logger.py WandbLogger: Weights & Biases logging, dumping, configuring.
    • loss.py Loss function base class with custom loss functions.
    • model.py Transformer model with linear head
    • train.py Trains and saves model, manages GPUs, checks accuracy, logs results.