conformal-clip-data

A companion data package providing benchmark datasets for the clip-conformal package.

Overview

This package bundles the simulated textile image dataset used in Megahed et al., 2025 for demonstrating conformal prediction with CLIP-based few-shot image classification in manufacturing quality control applications.

This is a data-only package designed to work seamlessly with the clip-conformal package (coming soon to PyPI), which provides the core implementation of conformal prediction methods for CLIP models. By separating data from implementation, we keep the main package lightweight while providing easy access to reproducible benchmark datasets.

Installation

Install from PyPI (data only):

pip install conformal-clip-data

Optional image preview tools (Pillow + matplotlib) as an extra:

pip install "conformal-clip-data[standalone]"

Install from source:

git clone https://github.com/fmegahed/conformal-clip-data.git
cd conformal-clip-data
pip install -e .
# with preview tools: pip install -e .[standalone]

Quick Start

from conformal_clip_data import textile_simulated_root, nominal_dir, local_dir, global_dir

# Access the textile dataset root
textile_path = textile_simulated_root()
print(f"Textile dataset location: {textile_path}")

# Access specific image categories
print(f"Nominal images: {nominal_dir()}")
print(f"Local defect images: {local_dir()}")
print(f"Global defect images: {global_dir()}")

Quick Image Peek (optional)

If you installed the optional extras ([standalone]), you can quickly visualize a few samples:

from conformal_clip_data import nominal_dir, local_dir, global_dir
import random
from PIL import Image
import matplotlib.pyplot as plt

samples = []
for d in [nominal_dir(), local_dir(), global_dir()]:
    files = list(d.glob("*.jpg"))
    samples += random.sample(files, k=min(3, len(files)))

cols = 3
rows = (len(samples) + cols - 1) // cols
plt.figure(figsize=(12, 4 * rows))
for i, f in enumerate(samples, 1):
    plt.subplot(rows, cols, i)
    plt.imshow(Image.open(f), cmap="gray")
    plt.title(f.parent.name + " / " + f.name)
    plt.axis("off")
plt.tight_layout()
plt.show()

Verify Installation

Note on names (PEP 503 normalization):

• We publish as conformal-clip-data; pip also accepts conformal_clip_data and resolves to the same project.

Quick check in your shell:

python -c "import conformal_clip_data as c, sys; print(c.__version__); print(c.__file__); print(sys.executable)"

Colab tip: use %pip install conformal-clip-data in one cell, then import in a new cell.

Dataset Provenance

These images were originally generated using the R script below and were previously released under an MIT License in our repository:

• Image generation script: https://raw.githubusercontent.com/fmegahed/qe_genai/refs/heads/main/data/textile_images/extract_textile_images_from_r_textile_pkg.R

• Original dataset release: https://github.com/fmegahed/qe_genai/tree/main/data/textile_images

Dataset Summary

To systematically evaluate CLIP's performance on STS image classification, we used the spc4sts R package to create a controlled dataset of simulated textile fabric textures. This approach allowed us to precisely model both nominal and defective weave structures and to control defect type and severity.

Our dataset contains:

Class	Description	Count
Nominal	Standard textile weave patterns	1,000
Local defects	Localized disruptions in the weave	500
Global defects	Systematic shifts in weave parameters	500

Each image is 250 × 250 px, generated using spc4sts recommended parameters:

• Nominal images:
  Spatial autoregressive parameters ϕ₁ = 0.6, ϕ₂ = 0.35

• Global defects:
  Both parameters reduced by 5%

• Local defects:
  Generated using the package's defect-insertion functions

Relationship with clip-conformal

This data package is designed as a companion to the clip-conformal package, which will be released to PyPI shortly. The separation of concerns provides several benefits:

• Lightweight installation: The clip-conformal package remains small and fast to install
• Reproducibility: Benchmark datasets are versioned and distributed consistently
• Extensibility: Additional datasets can be added without modifying the core package
• Optional usage: Users can work with clip-conformal using their own data without downloading benchmark datasets

For the full implementation of conformal prediction methods for CLIP models and complete examples using this dataset, please install the clip-conformal package (coming soon).

Citation

If you use this dataset in your research, please cite:

@misc{megahed2025adaptingopenaisclipmodel,
      title={Adapting OpenAI's CLIP Model for Few-Shot Image Inspection in Manufacturing Quality Control: An Expository Case Study with Multiple Application Examples},
      author={Fadel M. Megahed and Ying-Ju Chen and Bianca Maria Colosimo and Marco Luigi Giuseppe Grasso and L. Allison Jones-Farmer and Sven Knoth and Hongyue Sun and Inez Zwetsloot},
      year={2025},
      eprint={2501.12596},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2501.12596},
}

And the original spc4sts package used to generate the textile images:

@article{bui2020spc4sts,
  title={spc4sts: Statistical process control for stochastic textured surfaces in R},
  author={Bui, Anh Tuan and Apley, Daniel W},
  journal={Journal of Quality Technology},
  volume={53},
  number={3},
  pages={219--242},
  year={2020},
  doi={10.1080/00224065.2019.1707730}
}

License

MIT License. These images were generated by the authors and are released under the MIT License. See LICENSE for details.