---
license: mit
library_name: pytorch
pipeline_tag: image-classification
tags:
  - image-to-text
  - ocr
  - cipher
  - moon-cipher
  - pytorch
  - cnn
  - image-classification

# Shown in model card widget (with model.safetensors)
model_size: "13.6M params"
tensor_type: "F32"
format: Safetensors
training_epochs: 150

---

# Moon Cipher Detector (Classifier)

CNN classifier for **moon cipher** glyph recognition: 27 classes (A–Z plus `~`). Input: 128×128 grayscale glyph crop. Use with a detector for full image decoding.

**Model card (right sidebar):** This repo includes **Safetensors** (`model.safetensors`), so the Hub shows **Model size**, **Tensor type (F32)**, and format. Best checkpoint from training (up to 150 epochs).

## Model metadata

| Model | Format | Size | Params | Tensor type |
|-------|--------|------|--------|--------------|
| Classifier (`model.safetensors`) | Safetensors | 51.98 MB | 13,616,347 params | F32 |

## Usage

### 1. Install

```bash
pip install torch torchvision huggingface_hub safetensors
```

### 2. Download from Hugging Face

```python
from huggingface_hub import hf_hub_download

# Safetensors (preferred; used for Hub widget)
model_path = hf_hub_download(
    repo_id="nhellyercreek/moon-cipher-detector",
    filename="model.safetensors"
)
mappings_path = hf_hub_download(
    repo_id="nhellyercreek/moon-cipher-detector",
    filename="mappings.json"
)
```

### 3. Load and run

Use `MoonClassifier` from the Moon-Cipher-Detector repo (`models/moon_classifier.py`). Load Safetensors with `safetensors.torch.load_file(model_path)` and `model.load_state_dict(state_dict)`, or use `best_classifier.pth` with `torch.load(..., weights_only=True)`. Input: 128×128 grayscale glyph crops.

## Config

- **Classes**: 27 (A–Z + `~`)
- **Architecture**: Custom CNN (MoonClassifier), 128×128 input.
- **Format**: Safetensors + PyTorch `.pth`, **tensor type**: F32.
- **Training**: Best checkpoint (up to 150 epochs).

See `config.json` for machine-readable settings (params, size, format, training_epochs).