---
license: unlicense
library_name: onnx
pipeline_tag: image-classification
tags:
- onnx
- image-classification
- ritual-recognition
---

# RitualCNN

RitualCNN recognizes the seven ritual drawing classes used by the [Rituals game](https://github.com/methil-group/rituals).

## Use the model

The ONNX file is `RitualCNN.onnx`. It accepts a grayscale image as a `float32` tensor named `input`, with shape `[batch, 1, 128, 128]`.

Before inference, convert the image to grayscale, resize it to 128 by 128 pixels, set pixels above 127 to 1 and all other pixels to 0. The model returns seven raw scores named `output`. The highest score is the predicted class.

The output order is: `Acervus`, `Avarus`, `Flosculus`, `Mutus`, `Not`, `Pudor`, `Regium`.

## Example predictions

These are model inputs from the held-out validation set after preprocessing. When a class has no correct validation predictions, its row shows the highest-scoring example. The top softmax score is not a calibrated probability.

| Input image | True label | Predicted label | Top softmax score | Result |
|---|---|---|---:|---|
| ![Acervus example](examples/acervus.png) | Acervus | Acervus | 19% | Correct |
| ![Avarus example](examples/avarus.png) | Avarus | Avarus | 61% | Correct |
| ![Flosculus example](examples/flosculus.png) | Flosculus | Flosculus | 51% | Correct |
| ![Mutus example](examples/mutus.png) | Mutus | Regium | 39% | Incorrect |
| ![Not example](examples/not.png) | Not | Not | 65% | Correct |
| ![Pudor example](examples/pudor.png) | Pudor | Pudor | 27% | Correct |
| ![Regium example](examples/regium.png) | Regium | Regium | 52% | Correct |

## Confusion matrix

Rows show the true label. Columns show the predicted label. This matrix uses all 33 held-out validation images.

![Confusion matrix for the validation set](confusion_matrix.png)

## Training and evaluation
The model was trained on 163 labeled source image files. A stratified 80/20 split was made by class before augmentation. Augmentation runs on the fly for training images only: rotations up to 15 degrees and translations up to 12 pixels. The split contains 130 training images and 33 validation images. 0 byte-identical duplicate groups and 0 duplicate model-input groups were found.

Validation accuracy is 21/33 (63.6%); macro F1 is 59.8%. The validation set has 3 to 7 images per class. Early stopping selected epoch 47 using validation loss, so these results are for model selection rather than an independent final test.

A separate flat test folder had 6 files. 2 exact duplicates were excluded, leaving 4 images from 3 classes; 3/4 were correct. This set is too small and incomplete to estimate general performance.

The split is by file. The repository has no creator or session IDs for a grouped split, and drawings of the same ritual naturally look similar. Validation results are an estimate from a small dataset, not a guarantee of accuracy on new drawings. See `metrics.json` for the exact file lists, dataset fingerprint, per-class results, and ONNX compatibility checks.

## Files

- `RitualCNN.onnx` — trained model, ONNX opset 11.
- `config.json` — input, output, and class information.
- `metrics.json` — training details and validation results.
- `examples/` — validation input images shown above.
- `confusion_matrix.png` — validation confusion matrix.
- `README.md` — this model card.

The original drawings and the game source are in the [Rituals GitHub repository](https://github.com/methil-group/rituals). The exact training split and configuration for this model are listed in `metrics.json`.