unicornwhodev/lite-rt_models_cadrylpro

LiteRT Models for Cadryl Pro

Seven specialized model families in 41 LiteRT files, with input/output contracts, labels, artifact hashes and CPU inference examples. The repository is public; access to its files requires manual approval on Hugging Face.

Model catalogue

Model	Formats	Input	Test images	Recorded FP16 result
Isolated waste classification	FP32, FP16	224 px	238	Balanced accuracy: 0.8521
Synthetic material classification	FP32, FP16	320 px	1,480	Balanced accuracy: 0.9993
PCB1 anomaly classification, 448 px	FP32, FP16	448 px	442	Balanced accuracy: 0.9713
Helmets, heads and people	FP32, FP16	320 px	478	mAP50:95: 0.2775
Pothole segmentation	FP32, FP16, INT8	256 px	60	Foreground IoU: 0.6750
PCB1 anomaly classification, 320 px	FP32, FP16	320 px	442	Balanced accuracy: 0.9325

Scores are specific to each model's test set and metric; they are not a common cross-task ranking. The 320 px PCB1 model is a separate compact variant of the 448 px model. See catalogue.json for exact values and source revisions.

Get started

Request access on this page. After approval, authenticate and download the pinned package:

hf auth login
hf download unicornwhodev/lite-rt_models_cadrylpro --revision 83df995280e11d881804474d2edfae5225277e47 --local-dir cadryl-models
cd cadryl-models
python -m pip install tensorflow-cpu==2.15.1 numpy==1.26.4 pillow==10.4.0
python verify_package.py
python infer.py models/trashnet your-photo.jpg --precision fp16

For a binary mask or bounding boxes:

python infer.py models/nids-de-poule road.jpg --precision int8 --mask-output mask.png
python infer.py models/casques worksite.jpg --precision fp16

Each model directory contains .tflite files, labels.txt, contract.json, format-specific metrics and CPU inference evidence. FP16 is the default weight precision; its input remains float32. The pothole INT8 variant accepts uint8 input and can retain floating-point operations.

Images are decoded as RGB and normalization is inside the graph. Classification and segmentation use direct bilinear resizing. The detector preserves aspect ratio, pads with gray 127.5, and places the resized image at the top left. Its named SSD outputs are mapped back to the original photograph's coordinates.

Verification and limitations

The retained 5 October 2026 verification record reports that all 13 files were loaded and executed on Windows CPU with TensorFlow Lite 2.15.1, with checks of hashes, signatures, shapes and dtypes. The original package review also reran the full waste and pothole test sets (238 and 60 images); the other scores come from retained complete Colab evaluations. This card refresh does not repeat those executions. verify_package.py checks and runs a downloaded copy.

Use the models within their documented domains: isolated waste objects, synthetic material renders, the PCB1 board family and road imagery for potholes. Transfer of the material classifier to real photographs is unmeasured. Helmet/head/person detections require human review and do not establish individual safety compliance.

Android phone qualification, on-device latency and validation on your own task images remain open. Cadryl mask and SSD adapters require their own integration checks. A test-set score is not a guarantee on another domain.

Attribution and terms

Read NOTICES.md, the notices in licenses/ and the sources recorded in catalogue.json. The existing Apache-2.0 code licence applies separately; no new blanket licence for model weights is granted. Upstream terms remain applicable. The pothole source's recorded CC-BY-4.0/Apache-2.0 declaration discrepancy remains documented in the notices. Access approval does not replace a licence.

Documentation and revision

Product catalogue updated on 7 October 2026. The download command pins the complete package containing all 41 LiteRT files. Each model retains its source revision, training seed, full test results and input/output contract. The additional 24 files were hash-checked and executed on Colab CPU before publication.

Publication du 7 octobre 2026

Onze entraînements supplémentaires sont disponibles : TrashNet, matériaux et nids-de-poule graines 42/2026, casques graines 42/2026, blocs de texte OCR-D graines 17/42/2026. Chaque modèle conserve son corpus et ses partitions. Les variantes présentes ont passé l'évaluation complète de leur test et une nouvelle vérification d'inférence CPU.

Modèle	Formats	Test	Score FP16
Tri de dechets isoles · graine 42	FP32, FP16	238	balanced_accuracy : 0.8774
Masque de nids-de-poule · graine 42	FP32, FP16, INT8	60	foreground_iou : 0.6694
Bois, plastique, metal et verre synthetiques · graine 42	FP32, FP16	1480	balanced_accuracy : 1.0000
Tri de dechets isoles · graine 2026	FP32, FP16	238	balanced_accuracy : 0.8640
Masque de nids-de-poule · graine 2026	FP32, FP16, INT8	60	foreground_iou : 0.6661
Bois, plastique, metal et verre synthetiques · graine 2026	FP32, FP16	1480	balanced_accuracy : 1.0000
Casques, tetes et personnes sur chantier · graine 42	FP32, FP16	478	map_50_95 : 0.2824
Blocs de texte dans les imprimes historiques · graine 17	FP32, FP16	48	map_50_95 : 0.2754
Casques, tetes et personnes sur chantier · graine 2026	FP32, FP16	478	map_50_95 : 0.2812
Blocs de texte dans les imprimes historiques · graine 42	FP32, FP16	48	map_50_95 : 0.3058
Blocs de texte dans les imprimes historiques · graine 2026	FP32, FP16	48	map_50_95 : 0.3455

OCR-D propose des boîtes de blocs de texte, sans reconnaissance des caractères. La classe person du modèle casques reste peu précise. Les matériaux sont issus de rendus synthétiques. La qualification Android et la validation métier restent à réaliser.

VisA PCB1 — graines 42 et 2026

Deux entraînements supplémentaires de classification normal/anomalie sont disponibles pour la catégorie PCB1 du corpus VisA. Les quatre variantes FP32/FP16 ont passé l’évaluation complète sur les mêmes 442 images et une nouvelle vérification d’inférence CPU. Les modèles existants sont conservés.

Modèle	Formats	Test	Score FP16
Controle visuel d'une famille de circuits · graine 42	FP32, FP16	442	balanced_accuracy : 0.9125
Controle visuel d'une famille de circuits · graine 2026	FP32, FP16	442	balanced_accuracy : 0.9751

Test : 402 images normales et 40 anomalies. Consulter les métriques par classe ; un score de classification ne localise pas le défaut et ne prouve pas la conformité industrielle. La qualification Android et la validation sur vos pièces restent à réaliser. Les conditions et attributions VisA accompagnent chaque modèle ; aucune licence nouvelle des poids n’est attribuée.

Cadryl extension models

29 completed trials and their LiteRT contracts.