dropwise 0.2.0

Monte Carlo Dropout-based uncertainty estimation for Transformers

Dropwise

Dropwise is a lightweight PyTorch/HuggingFace wrapper for performing Monte Carlo Dropout–based uncertainty estimation in Transformers. It enables confidence-aware decision making by revealing how certain a model is about its predictions — all with just a few lines of code.

---

🚀 Features

• ✅ Enable dropout during inference for Bayesian-like uncertainty estimation
• ✅ Compute predictive entropy, confidence, and per-class standard deviation
• ✅ Works seamlessly with Hugging Face Transformers and PyTorch
• ✅ Supports batch inference, CPU/GPU, and customizable num_passes
• ✅ Cleanly packaged and extensible for research or production

---

🤖 Supported Tasks

Task Type	Example Model
sequence-classification	distilbert-base-uncased-finetuned-sst-2-english
token-classification	dslim/bert-base-NER
question-answering	deepset/bert-base-cased-squad2
regression	roberta-base (custom heads)

⚠️ Your model must contain dropout layers for MC sampling to work (most HF models do).

---

🤖 Supported Models

Dropwise works with any AutoModelForSequenceClassification compatible model including:

• bert-base-uncased, bert-large-uncased
• roberta-base, roberta-large
• deberta-v3-base, deberta-v2-xlarge
• albert-base-v2, distilbert-base-uncased
• Any custom fine-tuned Hugging Face model for classification

⚠️ The model must contain dropout layers (most pretrained transformers do).

---

📦 Installation

pip install dropwise

Or install locally for development:

git clone https://github.com/aryanator01/dropwise.git
cd dropwise
pip install -e .

---

🧠 Example Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
from dropwise import DropwisePredictor

model = AutoModelForSequenceClassification.from_pretrained("bert-base-uncased")
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")

predictor = DropwisePredictor(model, tokenizer, num_passes=30)
result = predictor.predict("The performance was unexpectedly good!")

print("Predicted class:", result['predicted_class'].item())
print("Entropy:", result['entropy'].item())
print("Confidence:", result['probs'].max().item())
print("Per-class std dev:", result['std_dev'])

---

📊 Output Dictionary

• predicted_class: index of most probable class
• entropy: predictive entropy (higher = less confident)
• std_dev: standard deviation across MC passes for each class
• mean_logits: average logits before softmax
• probs: softmax probabilities

---

🧠 Why Dropwise?

Unlike deterministic predictions, Dropwise estimates uncertainty via stochastic forward passes — enabling confidence-aware applications.

Use cases include:

• Filtering low-confidence predictions
• Active learning and semi-supervised setups
• Detecting ambiguous, adversarial, or out-of-distribution inputs
• Enhancing interpretability and robustness in real-world deployment

---

📂 Folder Structure

dropwise/
├── dropwise/
│   ├── __init__.py
│   ├── predictor.py
│   └── tasks/
│       ├── __init__.py
│       ├── sequence_classification.py
│       ├── token_classification.py
│       ├── question_answering.py
│       └── regression.py
├── tests/
│   └── test_predictor.py
├── setup.py
├── README.md
├── LICENSE

---

🧪 Running Tests

python tests/test_predictor.py

---

📝 License

MIT License

---

Made with ❤️ for uncertainty-aware, explainable AI.