Templates

Start from a model that already trains

Train an Auspex model from scratch on your own labels, or fine-tune YOLO, ResNet or HRNet from public weights. Every template is a working script you can read and change — not a wrapper you have to trust.

Auspex models

Trained from scratch on your own labels.

GPU

Auspex RT

Bounding-box detection, trained from scratch on your labels.

COCO JSONYOLO

pip install auspex-rt==0.2.0

GPU

Auspex Seg

Instance segmentation: a polygon, class and box for every object.

COCO polygons

pip install auspex-seg==0.2.1

GPU

Auspex Vision

Boxes, polygons, keypoints, polylines and tags, all in one model.

COCO JSON

pip install auspex-vision==0.7.0

GPU

Auspex Cursus

Polyline detection: an ordered sequence of points per instance.

CursusJSONCVAT XML

pip install auspex-cursus==0.8.2

GPU

Auspex Engine

An auto-annotator: one specialist per annotation type, bundled into one model.

COCO JSONYOLOCVAT XML

pip install auspex-engine==0.3.1

CPU

Auspex Test

Trains nothing: runs any Auspex model you trained over images and draws what it finds, with the package and version that trained it.

A trained best.ptAn Engine bundleImages

Popular models

Fine-tuned from public weights, so they need far fewer epochs; plus a blank project to test the loop.

GPU

YOLO

Pick YOLO26, YOLO11 or YOLOv8 for detection, segmentation, pose or classification, or YOLO12, v10, v9, v5 and RT-DETR for boxes, in every size.

YOLO data.yamlCOCO JSONClass folders

pip install ultralytics>=8.4.0

CPU

YOLO Test

Trains nothing: runs a model you trained over images and draws what it finds, so you see the boxes before you trust the score.

A trained best.ptImages

pip install ultralytics>=8.4.0

GPU

ResNet

Image classification with any of timm’s 1,078 pretrained models: ResNet-18 to ResNet-200, ConvNeXt, EfficientNet, ViT and more.

Class folders

pip install timm>=1.0.28,<2

GPU

HRNet Pose

Keypoints inside the boxes you already have, with any of the 11 HRNets from w18-small to w64.

COCO keypoints

pip install timm>=1.0.28,<2

GPU

SAM Annotate

Trains nothing: SAM 3, SAM 2.1, SAM 2, SAM or MobileSAM turns your images into polygon masks to train the others on.

ImagesCOCO boxesText prompts

pip install ultralytics>=8.3.237

CPU

Blank script or notebook

Trains nothing: prints the environment and writes proof.txt, to check the whole loop at zero GPU cost.

No dataset needed

Start here on your first run

What a template does for you

The same five things, whichever model it trains. They are the difference between a script that runs on your machine and one that survives a Kaggle session.

It installs what it needs, pinned

Kaggle's image does not carry these packages, and the newest release of one is not always the one a template was written against. Each template installs an exact version, so a run you make in six months trains the same model as the one you make today.

It finds your data, whatever the export looks like

Point it at a COCO JSON, a YOLO data.yaml, a CVAT XML or a folder of class folders. It searches /kaggle/input, unpacks an archive Kaggle left zipped, and tells you what it found and what it had to skip — before the first epoch, not after an hour.

It checks the settings against your data first

Epochs, image size and batch are suggested from the number of images you actually attached and the rate the template was measured at on a T4. A number that cannot fit is refused with a reason, rather than discovered when the session ends.

It carries on across sessions

Kaggle ends a session at 12 hours and keeps nothing when it does. Every template stops itself at 11, saves its checkpoint and writes down how far it got — and the next session resumes from there by itself, until the epochs you asked for are done. A 40-hour training is a run you start once.

It leaves the outputs where you can use them

best.pt and last.pt stay under runs/, and the metrics, plots and config land in artefacts/ as flat files. Nothing worth keeping is buried in a path you have to guess at.

What you need to bring

Labelled pictures, as a Kaggle dataset, in one of these. Export from whatever tool you label in — no conversion step of ours in between.

COCO JSON: One .json of images, annotations and categories. Boxes, polygons and keypoints.
YOLO: A data.yaml beside images/ and labels/, with one .txt a picture. Boxes and polygons.
CVAT 1.1 XML: An annotations.xml export. Boxes, polygons, polylines, points and tags.
Class folders: One folder a class, pictures inside. What a classifier wants.
CursusJSON: Ordered points per instance, for polyline models.

Each card above says which of these it reads. If your pictures are unlabelled, Auspex Engine trains an auto-annotator from the ones you have labelled and then labels the rest.

Your next training run is one click away.

Start your free trial See pricing