tfold - Tensorflow Object Detector

Introduction

This simple wrapper aims to provide a more programmatic alternative to the training pipeline provided by Tensorflow object detection research project.

Installation

pip install git+https://github.com/telize-ai/tfold.git

Usage

Detection

# Load the inference graph
detector = Detector(
    graph_path='frozen_inference_graph.pb',
    label_path='label_map.config'
)  

# Run the detection
detections = detector.detect_in_file('image.jpg')

# Draw an image with bounding boxes
detector.draw_vis(threshold=0.4)

Training

json_data = open(os.path.join(data_dir, 'samples.json')).read()
samples = json.loads(json_data)

train = Trainer(
    model_name='detection_model',
    model_dir='models/detection_model',
    model_type='faster'
)

for sample in samples['train']:
    filename = os.path.join(data_dir, sample['filename'])
    train.add_train_sample(filename, sample['boxes'])

for sample in samples['eval']:
    filename = os.path.join(data_dir, sample['filename'])
    train.add_eval_sample(filename, sample['boxes'])

train.start()

The samples.json file contains a json with the following format:

[
    {
        "filename": "image.jpg",
        "boxes": [
            {
                "xmin": 100,
                "xmax": 300,
                "ymin": 100,
                "ymax": 200,
                "class_text": "car",
                "class_int": 1
            }
        ]
    }
]

Todo

Contributions

If you would like to contribute please contact me at marius@mi.www.ro. For any other questions, please use the same address, or alternatively open an issue. I'd love to hear your thoughts, so don't heistate to write me.