# color space 

**Repository Path**: genvex/color-space

## Basic Information

- **Project Name**: color space 
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-08-21
- **Last Updated**: 2026-08-21

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# Caviar Color Grading System

## Structure

```
caviar-grading/
├── PAPER.md                 ← Paper (v1.0, verified)
├── grade.py                 ← Grading CLI (production)
├── models/
│   ├── model.joblib         ← RandomForest (RGB mean+std → 6 classes)
│   └── mixed_cnn.pth        ← ResNet18 (mixed vs non-mixed)
├── fig1_pca_2d.png          ← PCA scatter
├── fig2_pc1_hist.png        ← Brightness histogram
├── fig3_confusion.png       ← Confusion matrix
├── fig4_metrics.png         ← Precision/Recall/F1
└── experiments/             ← Full experiment scripts (numbered)
    01_extract_features.py
    02_pca.py
    03_visualize.py
    04_variance_analysis.py
    05_rf_classifier.py
    06_cnn_baseline.py
    07_cnn_focalloss.py
    08_cnn_5fold.py
    09_generate_figs.py
    10_validate_pipeline.py
```

## Usage

### Grading (production)

```bash
# Prerequisites
pip install torch torchvision scikit-learn joblib pillow numpy

# Grade a single image
python3 grade.py /path/to/caviar.jpg

# Grade batch
python3 grade.py img1.jpg img2.jpg img3.jpg ...
```

**Output**: Prints CNN mixed probability, RF 6-class predictions with confidence, and final verdict (grade or REVIEW).

### Reproduce experiments

```bash
# Step 0: Data at /data1/usb_backup/ (7-class JPG folders)

# Run in order
python3 experiments/01_extract_features.py   # → /tmp/fish_features.npy
python3 experiments/02_pca.py                 # → /tmp/fish_pca.npy
python3 experiments/03_visualize.py           # → figs
python3 experiments/05_rf_classifier.py       # → /tmp/caviar_model/
python3 experiments/07_cnn_focalloss.py       # → /tmp/mixed_cnn_v2.pth
python3 experiments/10_validate_pipeline.py   # → full validation
```

## Dependencies

| Package | Min Version |
|---|---|
| torch | 2.11.0+cu130 |
| torchvision | ≥0.16 |
| scikit-learn | 1.9 |
| numpy | 1.26 |
| Pillow | 10 |
| matplotlib | 3.11 |

## Requirements

- GPU: NVIDIA (RTX 5090 used, but any CUDA GPU works)
- Disk: ~60 GB for models + data
- Training CNN: ~10 min on 5090

## Key Numbers

| Metric | Value |
|---|---|
| CNN mixed detection | 92.1% recall (58/63) |
| RF 6-class (5-fold CV) | 84.2% ± 0.97% |
| Full pipeline (3027 imgs) | 90.3% |
| Single image latency | < 200ms |