Skip to content

Installation

AdaptShot v0.2.0 is a Python package for CPU-first few-shot image classification with conformal prediction, contrastive learning, and human-in-the-loop continual learning.

Requirements

  • Python 3.9+
  • A CPU-only environment is supported by default (PyTorch is optional)
  • ~15 MB disk space for core dependencies (numpy + Pillow)
  • ~45 MB additional for ImageNet pretrained backbone weights (auto-downloaded on first use)
  • Internet access for the first run only (backbone weight download)

Version

These commands target AdaptShot v0.2.0.

Install From PyPI

pip install adaptshot

Fast install: Core dependencies (numpy, Pillow) install in under 60 seconds. No GPU drivers, no CUDA, no 2 GB downloads. PyTorch is optional.

Install From GitHub Release

pip install https://github.com/johnson2006christopher/adaptshot/releases/download/v0.2.0/adaptshot-0.2.0-py3-none-any.whl

Optional Extras

# PyTorch for training, fine-tuning, and custom backbones
pip install "adaptshot[torch]"

# FAISS-CPU similarity search (recommended for >100 support images)
pip install "adaptshot[faiss]"

# Gradio UI dependencies for the Pilot Dashboard
pip install "adaptshot[ui]"

# Offline Studio GUI (includes ONNX Runtime for torch-free inference)
pip install "adaptshot[gui]"

# Development tools (testing, linting, benchmarking)
pip install "adaptshot[dev]"

# Everything
pip install "adaptshot[all]"

Install From Source

git clone https://github.com/johnson2006christopher/adaptshot.git
cd adaptshot
pip install -e ".[dev]"

ONNX Runtime Support (Torch-Free Inference)

AdaptShot v0.2.0 supports ONNX Runtime as a lightweight backend when PyTorch is not installed. To use this:

# Export backbones to ONNX format (requires torch)
python scripts/export_backbones.py

# Install with ONNX Runtime support
pip install "adaptshot[gui]"  # Includes onnxruntime

The ONNX backend automatically activates when torch is unavailable and ONNX models are present in src/adaptshot/data/.

Verify The Install

python - <<'PY'
import adaptshot
from adaptshot import FewShotLearner, ConformalEngine

print(adaptshot.__version__)
print(FewShotLearner.__name__)
print(ConformalEngine.__name__)
PY

Expected output:

0.2.0-dev
FewShotLearner
ConformalEngine

Pretrained Weights

AdaptShot v0.2.0 uses ImageNet-pretrained backbone weights by default. The weights are downloaded automatically on first use (~45 MB for ResNet-18, ~10 MB for MobileNetV3-Small). This ensures embeddings are meaningful and match the ImageNet-normalized preprocessing pipeline.

Minimal Dependency Setup

For ultra-lightweight deployments (no PyTorch, no ONNX Runtime):

pip install adaptshot
# Core dependencies: numpy + Pillow only (~15 MB total)

The library falls back gracefully: ONNX backend is tried first if available, then a clear error message guides you to install either [torch] or [gui] extras.

Verification Checklist

  • [ ] python --version shows Python 3.9 or newer.
  • [ ] pip install adaptshot completes in under 60 seconds.
  • [ ] The verification command prints 0.2.0-dev.
  • [ ] You can import FewShotLearner and ConformalEngine.

Next Steps