Skip to main content

​
Overview

YOLO-Pi uses the YAD2K (Yet Another Darknet to Keras) conversion tool to transform YOLO models from the Darknet framework into Keras-compatible .h5 model files. This guide explains how to download YOLO weights and configurations, and convert them for use with YOLO-Pi.

​
Prerequisites

  • Python 3.6 or higher
  • YAD2K conversion tool (included in src/yad2k/)
  • YOLO weights and configuration files from the official YOLO site

​
Supported Models

YOLO-Pi supports the following YOLO model configurations:

Tiny YOLO VOC

Lightweight model optimized for speed. Best for Raspberry Pi and resource-constrained devices.
  • Model: tiny-yolo-voc.h5
  • Classes: PASCAL VOC (20 classes)
  • Performance: ~1 frame per 2 seconds on MacBook Pro

Full YOLO

Full-size model for better accuracy at the cost of performance.
  • Model: yolo.h5
  • Classes: COCO (80 classes)
  • Performance: Slower but more accurate

​
Download YOLO Models

1

Visit the YOLO website

Download the YOLO weights and configuration files from the official YOLO site.
2

Choose your model

For Raspberry Pi and embedded devices, we recommend Tiny YOLO VOC:
wget https://pjreddie.com/media/files/tiny-yolo-voc.weights
wget https://raw.githubusercontent.com/pjreddie/darknet/master/cfg/tiny-yolo-voc.cfg
3

Verify downloads

Ensure both files are downloaded successfully:
ls -lh tiny-yolo-voc.*
You should see both .weights and .cfg files.

​
Converting Models with YAD2K

​
Basic Conversion

The YAD2K conversion process transforms Darknet weights and configuration into a Keras model: 
./yad2k.py tiny-yolo-voc.cfg tiny-yolo-voc.weights model_data/tiny-yolo-voc.h5

​
Conversion Command Format

./yad2k.py [config_file] [weights_file] [output_path]
Parameters:
  • config_file: Darknet configuration file (.cfg)
  • weights_file: Pre-trained YOLO weights (.weights)
  • output_path: Destination path for the Keras model (.h5)

​
Full YOLO Model Conversion

For the full YOLO model with COCO classes: 
./yad2k.py yolo.cfg yolo.weights model_data/yolo.h5
The conversion process may take several minutes depending on your hardware and the model size.

​
Model Files Structure

After conversion, you’ll need three essential files for each model:

​
1. Model File (.h5)

The converted Keras model containing the neural network architecture and weights. Location: model_data/tiny-yolo-voc.h5 or model_data/yolo.h5

​
2. Anchors File

Pre-defined anchor boxes used for object detection. These are specific to each model. Tiny YOLO VOC Anchors (model_data/tiny-yolo-voc_anchors.txt): 
1.08,1.19,  3.42,4.41,  6.63,11.38,  9.42,5.11,  16.62,10.52
Full YOLO Anchors (model_data/yolo_anchors.txt): 
0.57273, 0.677385, 1.87446, 2.06253, 3.33843, 5.47434, 7.88282, 3.52778, 9.77052, 9.16828
The anchor values are defined in the keras_yolo.py file:
voc_anchors = np.array(
    [[1.08, 1.19], [3.42, 4.41], [6.63, 11.38], [9.42, 5.11], [16.62, 10.52]])
See src/yad2k/models/keras_yolo.py:17-18

​
3. Classes File

List of object classes the model can detect. PASCAL VOC Classes (model_data/pascal_classes.txt): 
aeroplane
bicycle
bird
boat
bottle
bus
car
cat
chair
cow
diningtable
dog
horse
motorbike
person
pottedplant
sheep
sofa
train
tvmonitor
PASCAL VOC contains 20 object classes, while COCO contains 80 classes. Choose based on your detection needs.

​
Model Architecture

The YAD2K conversion creates a Keras model using the Darknet-19 architecture: 
def darknet_body():
    """Generate first 18 conv layers of Darknet-19."""
    return compose(
        DarknetConv2D_BN_Leaky(32, (3, 3)),
        MaxPooling2D(),
        DarknetConv2D_BN_Leaky(64, (3, 3)),
        MaxPooling2D(),
        bottleneck_block(128, 64),
        MaxPooling2D(),
        bottleneck_block(256, 128),
        MaxPooling2D(),
        bottleneck_x2_block(512, 256),
        MaxPooling2D(),
        bottleneck_x2_block(1024, 512))
Source: src/yad2k/models/keras_darknet19.py:51-65

​
Verification

After conversion, verify your model files are in place: 
ls -lh model_data/
You should see:
  • tiny-yolo-voc.h5 (or yolo.h5)
  • tiny-yolo-voc_anchors.txt (or yolo_anchors.txt)
  • pascal_classes.txt (or coco_classes.txt)
Ensure the model file, anchors, and classes match. Mixing Tiny YOLO VOC with COCO classes will result in errors.

​
Troubleshooting

​
YAD2K Not Found

If the yad2k.py script is not found, ensure you’re in the correct directory: 
cd src
python -m yad2k.yad2k [config] [weights] [output]

​
Conversion Errors

If you encounter errors during conversion:
  1. Verify the .cfg and .weights files match the same YOLO version
  2. Check that you have sufficient disk space for the output .h5 file
  3. Ensure all YAD2K dependencies are installed (Keras, TensorFlow, NumPy)

​
Model Validation

To validate the converted model loads correctly: 
from keras.models import load_model

# Load the converted model
yolo_model = load_model('model_data/tiny-yolo-voc.h5')
print(f'{yolo_model.layers[0].input_shape} model loaded successfully')

​
Next Steps

Running Inference

Learn how to use your converted model for object detection

MQTT Integration

Set up MQTT messaging for detection results

Build docs developers (and LLMs) love

Get started for freeTalk to us