LightNet: Bringing pjreddie's DarkNet out of the shadows

LightNet provides a simple and efficient Python interface to DarkNet, a neural network library written by Joseph Redmon that's well known for its state-of-the-art object detection models, YOLO and YOLOv2. LightNet's main purpose for now is to power Prodigy's upcoming object detection and image segmentation features. However, it may be useful to anyone interested in the DarkNet library.

LightNet's features include:

• State-of-the-art object detection: YOLOv2 offers unmatched speed/accuracy trade-offs.
• Easy-to-use via Python: Pass in byte strings, get back numpy arrays with bounding boxes.
• Lightweight and self-contained: No dependency on large frameworks like Tensorflow, PyTorch etc. The DarkNet source is provided in the package.
• Easy to install: Just pip install lightnet and python -m lightnet download yolo.
• Cross-platform: Works on OSX and Linux, on Python 2.7, 3.5 and 3.6.
• 10x faster on CPU: Uses BLAS for its matrix multiplications routines.
• Not named DarkNet: Avoids some potentially awkward misunderstandings.

🌓 Installation

LightNet requires an installation of OpenBLAS:

sudo apt-get install libopenblas-dev

LightNet can be installed via pip:

pip install lightnet

Once you've downloaded LightNet, you can install a model using the lightnet download command. This will save the models in the lightnet/data directory. If you've installed LightNet system-wide, make sure to run the command as administrator.

python -m lightnet download tiny-yolo
python -m lightnet download yolo

The following models are currently available via the download command:

🌓 Usage

An object detection system predicts labelled bounding boxes on an image. The label scheme comes from the training data, so different models will have different label sets. YOLOv2 can detect objects in images of any resolution. Smaller images will be faster to predict, while high resolution images will give you better object detection accuracy.

Images can be loaded by file-path, by JPEG-encoded byte-string, or by numpy array. If passing in a numpy array, it should be of dtype float32, and shape (width, height, colors).

import lightnet

model = lightnet.load('tiny-yolo')
image = lightnet.Image.from_bytes(open('eagle.jpg', 'rb').read())
boxes = model(image)

METHOD lightnet.load

Load a pre-trained model. If a path is provided, it shoud be a directory containing two files, named {name}.weights and {name}.cfg. If a path is not provided, the built-in data directory is used, which is located within the LightNet package.

model = lightnet.load('tiny-yolo')
model = lightnet.load(path='/path/to/yolo')

🌓 Network

The neural network object. Wraps DarkNet's network struct.

CLASSMETHOD Network.load

Load a pre-trained model. Identical to lightnet.load().

METHOD Network.__call__

Detect bounding boxes given an Image object. The bounding boxes are provided as a list, with each entry (class_id, class_name, prob, [(x, y, width, height)]), where `x` and y` are the pixel coordinates of the center of the centre of the box, and width and height describe its dimensions. class_id is the integer index of the object type, class_name is a string with the object type, and prob is a float indicating the detection score. The thresh parameter controls the prediction threshold. Objects with a detection probability above thresh are returned. We don't know what hier_thresh or nms do.

boxes = model(image, thresh=0.5, hier_thresh=0.5, nms=0.45)

METHOD Network.update

Update the model, on a batch of examples. The images should be provided as a list of Image objects. The box_labels should be a list of BoxLabel objects. Returns a float, indicating how much the models prediction differed from the provided true labels.

loss = model.update([image1, image2], [box_labels1, box_labels2])

🌓 Image

Data container for a single image. Wraps DarkNet's image struct.

METHOD Image.__init__

Create an image. data should be a numpy array of dtype float32, and shape (width, height, colors).

image = Image(data)

CLASSMETHOD Image.blank

Create a blank image, of specified dimensions.

image = Image.blank(width, height, colors)

CLASSMETHOD Image.load

Load an image from a path to a jpeg file, of the specified dimensions.

image = Image.load(path, width, height, colors)

CLASSMETHOD Image.from_bytes

Read an image from a byte-string, which should be the contents of a jpeg file.

image = Image.from_bytes(bytes_data)

🌓 BoxLabels

Data container for labelled bounding boxes for a single image. Wraps an array of DarkNet's box_label struct.

METHOD BoxLabels.__init__

Labelled box annotations for a single image, used to update the model. ids should be a 1d numpy array of dtype int32, indicating the correct class IDs of the objects. boxes should be a 2d array of dtype float32, and shape (len(ids), 4). The 4 columns of the boxes should provide the relative x, y, width, height of the bounding box, where x and y are the coordinates of the centre, relative to the image size, and width and height are the relative dimensions of the box.

box_labels = BoxLabels(ids, boxes)

CLASSMETHOD BoxLabels.load

Load annotations for a single image from a text file. Each box should be described on a single line, in the format class_id x y width height.

box_labels = BoxLabels.load(path)