conformal-clip 0.2.2

Few-shot CLIP classification with conformal prediction and calibrated probabilities

conformal_clip

Few-shot vision classification with conformal prediction and optional probability calibration.

This package runs CLIP-like models via open-clip-torch (any model in Hugging Face OpenCLIP) and also supports timm vision-only models for few-shot conformal prediction (Hugging Face timm models).

---

Features

• Backends: OpenCLIP (CLIP-like, vision+text) and timm (vision-only)
• Few-shot classification using exemplar image banks
• Conformal prediction (global and Mondrian) with finite-sample coverage
• Optional calibration (isotonic or sigmoid/Platt)
• Zero-shot baseline for CLIP-like models
• Benchmark utility to compare backends and settings across the same splits

Backends overview

Backends are provided in two broad categories: CLIP-like vision–language encoders (via open-clip-torch) and image-only encoders (via timm). All of them can be loaded with load_backend(backend, backend_model_id, device); the recommendations below are approximate and assume small batch sizes and fp16 where possible.

CLIP-like (vision–language) encoders

Family	Backend key	OpenCLIP model id	Recommended environment	Notes
ViT (small)	clip_b32	ViT-B-32-quickgelu	Low–medium (4–8 GB GPU or CPU)	Lightest ViT CLIP; good starting point when resources are tight.
ViT (small)	clip_b16	ViT-B-16-quickgelu	Medium (≥8 GB GPU)	More accurate than B/32 at modest extra cost.
ViT (base)	siglip2	ViT-B-16-SigLIP2	Medium (≥8 GB GPU)	ViT-B model with SigLIP2 loss; strong trade-off between accuracy and cost.
ViT (mobile)	mobileclip2	MobileCLIP2-S4	Low (CPU, 4–8 GB GPU, edge)	Mobile-optimized CLIP; preferred for low-power or edge deployments.
ViT (large)	openai	ViT-L-14-quickgelu	Medium–high (≥8–12 GB GPU)	Classic CLIP baseline; strong general performance.
ViT (xlarge)	openclipbase	ViT-H-14-quickgelu	High (≥16 GB GPU)	Larger ViT-H encoder; use when memory is ample.
ViT (giant)	vitg	ViT-bigG-14	Very high (≥24 GB GPU)	Extremely large model; for offline or benchmark use only.
ResNet CLIP	resnet50	RN50x64-quickgelu	Medium (≥8 GB GPU or strong CPU)	Deep CNN CLIP; useful as a non-ViT baseline.
EVA family	eva02	EVA02-E-14-plus	High (≥16–24 GB GPU)	High-capacity ViT-style model; heavy but strong.
ConvNeXt	convnext	convnext_xxlarge	Very high (≥24 GB GPU)	Very large ConvNeXt CLIP; avoid on small GPUs.
CoCa	coca	coca_ViT-L-14	High (≥12–16 GB GPU)	Captioning-oriented CLIP variant; strong but memory-hungry.

Image-only (vision encoders via timm)

Family	Backend key	timm model id	Recommended environment	Notes
Lightweight	mobilenetv4	mobilenetv4_conv_aa_large.e230_r448_in12k_ft_in1k	Low (CPU, 4–8 GB GPU)	Very efficient mobile-style CNN; best when resources are tight.
Lightweight	resnet18	resnet18.a1_in1k	Low (CPU, 4–8 GB GPU)	Classic small ResNet; easy to run and debug.
Lightweight	efficientnet_b0	efficientnet_b0.ra_in1k	Low–medium (CPU, 4–8 GB GPU)	Strong accuracy/efficiency balance among small CNNs.
ViT-L	dinov3	vit_large_patch16_dinov3.lvd1689m	Medium–high (≥8–16 GB GPU)	Self-supervised ViT-Large; heavy but strong general-purpose features.

---

Install

Core package:

pip install conformal-clip

With example dataset (textile images):

pip install "conformal-clip[data]"

Notes:

• Set HF_TOKEN in your environment (or a .env file) if you need access to gated models (e.g., DINOv3). The loader forwards it to HUGGINGFACE_HUB_TOKEN.
• PyTorch with CUDA is recommended for speed but not required.

---

Environment Setup

• Hugging Face token (for gated repos like some DINOv3 builds):
  • In shell: export HF_TOKEN=hf_... (Linux/macOS) or set HF_TOKEN=hf_... (Windows)
  • Or create a .env file next to your script with HF_TOKEN=hf_....
  • The loader maps HF_TOKEN to HUGGINGFACE_HUB_TOKEN automatically.
• CUDA (optional): Install a CUDA-enabled PyTorch build from PyTorch Get Started (Locally) then use device = torch.device("cuda").

---

Setup

All Quickstart blocks below build on a single setup flow: pick a device, load a backend, gather image paths, and split them into few-shot banks, a calibration set, and a test set. Run Step 1, then either Step 2A or Step 2B, then Step 3. After that you can run any Quickstart block.

Step 1 — Device + backend

import torch
from conformal_clip import load_backend

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model, preprocess_fn, tokenize_fn = load_backend("openclipbase", None, device)

Step 2 — Image paths (pick A or B)

Both options produce the same six path lists (nom_train, nom_calib, nom_test, def_train, def_calib, def_test) that Step 3 consumes.

Option A — example textile dataset (installed via conformal-clip[data])

import os
from conformal_clip_data import nominal_dir, local_dir, global_dir

def list_paths(d):
    exts = {".jpg", ".jpeg", ".png"}
    return [os.path.join(d, f) for f in os.listdir(d) if os.path.splitext(f)[1].lower() in exts]

nominal_paths = list_paths(nominal_dir())
local_paths   = list_paths(local_dir())
global_paths  = list_paths(global_dir())

nom_train, nom_calib, nom_test = nominal_paths[:50], nominal_paths[50:100], nominal_paths[100:150]
def_train = local_paths[:25]   + global_paths[:25]
def_calib = local_paths[25:50] + global_paths[25:50]
def_test  = local_paths[50:75] + global_paths[50:75]

Option B — your own local folders

Expects a nominal/ and a defective/ directory; adjust slice sizes to your dataset.

import os

base_dir      = "./data/textile_images/simulated"
nominal_dir   = os.path.join(base_dir, "nominal")
defective_dir = os.path.join(base_dir, "defective")

def list_paths(d):
    exts = {".jpg", ".jpeg", ".png"}
    return [os.path.join(d, f) for f in os.listdir(d) if os.path.splitext(f)[1].lower() in exts]

nominal_paths   = list_paths(nominal_dir)
defective_paths = list_paths(defective_dir)

nom_train, nom_calib, nom_test = nominal_paths[:50],   nominal_paths[50:100],   nominal_paths[100:150]
def_train, def_calib, def_test = defective_paths[:50], defective_paths[50:100], defective_paths[100:150]

Step 3 — Build PIL banks, labels, and filenames

This block runs after either Option A or Option B.

import os
from PIL import Image

pil_nom_bank = [Image.open(p).convert("RGB") for p in nom_train]
pil_def_bank = [Image.open(p).convert("RGB") for p in def_train]
pil_calib    = [Image.open(p).convert("RGB") for p in (nom_calib + def_calib)]
pil_test     = [Image.open(p).convert("RGB") for p in (nom_test  + def_test)]

calib_labels   = ["Nominal"] * len(nom_calib) + ["Defective"] * len(def_calib)
test_labels    = ["Nominal"] * len(nom_test)  + ["Defective"] * len(def_test)
test_filenames = [os.path.basename(p) for p in (nom_test + def_test)]

PIL images vs. preprocessed tensors. The APIs in this package differ in the form they expect:

• benchmark_models consumes PIL images directly and applies each backend's preprocessing internally as it sweeps.
• few_shot_fault_classification_conformal and evaluate_zero_shot_predictions consume preprocessed tensors — pass PIL images through preprocess_fn from load_backend(...) first.

Setup ends with PIL banks (pil_*). Each Quickstart block below applies preprocess_fn when (and only when) it's needed.

---

Quickstart

Few-shot classification with conformal prediction

Continues from Setup above. Few-shot conformal expects preprocessed tensors, so we first run each PIL bank through preprocess_fn (returned by load_backend in Step 1).

from conformal_clip import few_shot_fault_classification_conformal

nominal_images   = [preprocess_fn(img) for img in pil_nom_bank]
defective_images = [preprocess_fn(img) for img in pil_def_bank]
calib_images     = [preprocess_fn(img) for img in pil_calib]
test_images      = [preprocess_fn(img) for img in pil_test]

results = few_shot_fault_classification_conformal(
    model=model,
    test_images=test_images,
    test_image_filenames=test_filenames,
    nominal_images=nominal_images,
    nominal_descriptions=["nominal textile sample"] * len(nominal_images),
    defective_images=defective_images,
    defective_descriptions=["defective textile sample"] * len(defective_images),
    calib_images=calib_images,
    calib_labels=calib_labels,
    alpha=0.1,
    mondrian=True,
    prob_calibration="isotonic",  # or "sigmoid" or None
)

Zero-shot classification (CLIP-like backends only)

Continues from Setup above. Zero-shot also expects preprocessed tensors and additionally needs tokenize_fn. Setup loaded openclipbase (a CLIP-like backend), so tokenize_fn is non-None.

from conformal_clip import evaluate_zero_shot_predictions

test_images = [preprocess_fn(img) for img in pil_test]

metrics_df, results_df = evaluate_zero_shot_predictions(
    labels=["Nominal", "Defective"],
    label_counts=[test_labels.count("Nominal"), test_labels.count("Defective")],
    test_images=test_images,
    test_image_filenames=test_filenames,
    model=model,
    device=device,
    tokenize_fn=tokenize_fn,
    save_confusion_matrix=True,
)

Benchmark across backends, calibration, and conformal settings

Continues from Setup above. benchmark_models takes PIL images directly — no preprocess_fn step is needed, since the function loads each backend internally and applies that backend's own preprocessing as it sweeps.

from conformal_clip import benchmark_models

cls_df, cp_df, cls_style, cp_style = benchmark_models(
    train_nominal_images=pil_nom_bank,
    train_defective_images=pil_def_bank,
    calib_images=pil_calib,
    calib_labels=calib_labels,
    test_images=pil_test,
    test_labels=test_labels,
    device=device,
    seed=2025,
    # resource_tier controls which backends are run when backends is None.
    # Defaults to "low" (small models suitable for low-resource environments).
    resource_tier="low",  # or "medium" or "high"
    calibration_methods=(None, "isotonic", "sigmoid"),
    conformal_modes=(None, "global", "mondrian"),
    alpha_list=(0.1,),
)

Resource warning: many CLIP/timm backbones are large (hundreds of MB to multiple GB per model). Running benchmark_models over many backends can require tens of gigabytes of RAM and substantial disk cache space. We do not recommend running the full benchmark configuration on machines with limited memory or storage; instead, restrict the backends argument to a small subset of models.

---

Examples

The examples/ folder contains runnable end-to-end scripts. Four manufacturing-inspection datasets are demonstrated: textile (the default, ships via conformal-clip[data]), extrusion, microstructure, and pipe.

Per-backend few-shot scripts (each runs one backend on the textile dataset and shows the full pipeline — load, split, calibrate, conformal):

• textile_mobileclip2.py — MobileCLIP2 (lightweight CLIP variant)
• textile_mobilenetv4.py — MobileNetV4 (vision-only timm)
• zero_shot_openclip.py — Zero-shot baseline with an OpenCLIP backbone

Benchmark suites (sweep multiple backends, calibration methods, and conformal modes; emit CSV + styled HTML):

• benchmark_textile.py, benchmark_extrusion.py, benchmark_microstructure.py, benchmark_pipe.py

Discovery and customization:

• custom_openclip_example.py — using custom-clip and custom-clip-hf
• list_models_openclip_timm.py — browse available OpenCLIP and timm models

Each _shared_<dataset>.py module (e.g., _shared_textile.py) provides dataset-specific path resolution and reproducible train/calibration/test splits — the same pattern you can copy when adapting these scripts to your own data.

---

Discover and use custom models

Beyond the named backends in the tables above, you can load any OpenCLIP-compatible model or any timm vision backbone using the custom-clip, custom-clip-hf, or custom-vision backends. The snippets in this section are standalone — they do not depend on the Setup flow.

Custom OpenCLIP (vision–language) models

custom-clip accepts an OpenCLIP built-in model name (optionally with a pretrained tag via "model@tag"). custom-clip-hf accepts a Hugging Face repo id ("hf-hub:org/repo"). Browse models: Hugging Face OpenCLIP.

List available built-in model names and their tags:

import open_clip

# Dict[str, Set[str]] mapping model_name -> available pretrained tags
by_model = open_clip.list_pretrained_tags_by_model()

print("Some available model names (built-ins):")
for name in sorted(by_model.keys())[:20]:
    print(" ", name)

# Inspect tags for a specific model
model_name = "ViT-L-14-quickgelu"
print("Available pretrained tags for", model_name, ":", sorted(by_model.get(model_name, [])))

Load a custom built-in model (optionally specify a tag):

import torch
from conformal_clip import load_backend

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Using just the model name (the loader picks a sensible pretrained tag)
model, preprocess_fn, tokenize_fn = load_backend(
    backend="custom-clip", backend_model_id="ViT-L-14-quickgelu", device=device,
)

# Or pin a specific tag with "model@tag"
model, preprocess_fn, tokenize_fn = load_backend(
    backend="custom-clip", backend_model_id="ViT-L-14-quickgelu@openai", device=device,
)

Load from Hugging Face (OpenCLIP-compatible weights) using custom-clip-hf:

# Gated repos require HF_TOKEN in the environment or a .env file (see Environment Setup).
model, preprocess_fn, tokenize_fn = load_backend(
    backend="custom-clip-hf", backend_model_id="hf-hub:org-or-user/repo-id", device=device,
)

Custom timm (vision-only) models

custom-vision accepts any timm model id. The returned tokenize_fn is None because timm backbones do not have a text encoder, so vision-only backends are usable with few_shot_fault_classification_conformal and benchmark_models but not with evaluate_zero_shot_predictions. Browse models: Hugging Face timm.

List timm model names programmatically:

import timm

# All models that have pretrained weights available
names = timm.list_models(pretrained=True)
print(f"Found {len(names)} pretrained timm models")
print("First 20:", names[:20])

# Filter by family/pattern
print("mobilenet*:", timm.list_models("mobilenet*", pretrained=True)[:10])
print("convnext*:", timm.list_models("convnext*", pretrained=True)[:10])

Load a custom timm model:

import torch
from conformal_clip import load_backend

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

model, preprocess_fn, tokenize_fn = load_backend(
    backend="custom-vision",
    backend_model_id="convnext_base.fb_in22k_ft_in1k",
    device=device,
)
assert tokenize_fn is None  # vision-only backbones have no text encoder

---

Project structure

conformal_clip/
├── conformal_clip/          # Main package
│   ├── __init__.py          # Public API surface (see __all__)
│   ├── backends.py          # load_backend(...) for OpenCLIP and timm
│   ├── wrappers.py          # CLIPWrapper: sklearn-compatible few-shot classifier
│   ├── conformal.py         # Global and Mondrian conformal prediction
│   ├── zero_shot.py         # Zero-shot evaluation for CLIP-like models
│   ├── metrics.py           # Classification and conformal-set metrics
│   ├── benchmark.py         # benchmark_models(...) sweep utility
│   ├── viz.py               # Confusion matrix plotting
│   ├── image_io.py          # Image loading from disk or URL
│   ├── io_github.py         # List image URLs from GitHub directories
│   └── data_utils.py        # Train/calibration/test split helpers
├── examples/                # Runnable end-to-end scripts (see "Examples" above)
└── tests/                   # Unit and integration tests (pytest)

---

Citation

If you use this package 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}
}

---

License

MIT License (see LICENSE).