---
license: mit
pipeline_tag: image-classification
tags:
  - onnx
  - onnxruntime
  - image-classification
  - binary-classification
library_name: onnx
---

# cstria0106/shit-detector

ONNX image classifier for detecting whether an image contains feces or toilet waste/staining under this project's labeling policy.

## Try It In Your Browser

**Live browser demo:** [https://huggingface.co/spaces/cstria0106/shit-detector-browser](https://huggingface.co/spaces/cstria0106/shit-detector-browser)

Runs fully client-side: the demo downloads this ONNX model in the browser and runs inference locally with ONNX Runtime Web.


## Files

- `shit_detector.onnx`: ONNX model exported with dynamic batch axis.
- `metadata.json`: preprocessing values, threshold, class names, evaluation metrics, and deployment notes.
- `inference.py`: minimal ONNX Runtime inference helper.
- `requirements.txt`: runtime dependencies for the helper.

## Labels

Class order: `['shit', 'not_shit']`.

The positive class is `shit`. The default decision threshold is `0.149` on the positive-class probability.

## Dataset

The model was trained on a project-specific binary image dataset.

| Split | shit | not_shit |
|---|---:|---:|
| Train | 483 | 2,894 |
| Validation | 120 | 686 |
| Test | 189 | 1,003 |

## Training

- Objective: 2-class cross entropy
- Optimizer: AdamW
- Learning rate: `1.5e-4`
- Weight decay: `0.01`
- Epochs: up to 80
- Best checkpoint selected by validation F1
- Backbone frozen for first 5 epochs, then unfrozen
- Balanced sampling enabled with positive sampling fraction `0.40`
- Augmentation: random resized crop, horizontal flip, color jitter
- Positive-only color augmentation was used to improve robustness to color-shifted positive examples.

## Model Specs

- Base model: `tf_efficientnet_b0.ns_jft_in1k`
- Format: ONNX, opset `17`
- Input size: `256 x 256` RGB image
- Parameters: `4010110`
- ONNX size: `15.34 MB`
- Class order: `['shit', 'not_shit']`
- Positive threshold: `0.149`

## Evaluation Snapshot

- Precision: 93.51%
- Recall: 91.53%
- F1: 92.51%
- ROC-AUC: 99.09%
- PR-AUC: 97.42%

These metrics come from the local evaluation set recorded in `metadata.json`; they are not a public benchmark.

## Intended Use

This model is intended for project-specific binary image triage. It may produce false positives on visually similar brown food, stains, mud, and other hard negatives, and false negatives on small, occluded, color-shifted, or unusual positive cases.

## Runtime

```bash
pip install -r requirements.txt
python inference.py path/to/image.jpg
```

## Deployment Target

- Target: `lattepanda_n4120`
- CPU: `Intel Celeron N4120`
- Model parameters: `4010110`
- ONNX size MB: `15.341004371643066`