python-dlclib

A wrapper of DeepLabCut for its ease of use.

dlclib is supposed to work with almost all single-camera setups of all DLC versions.

Installation

It is better building your DeepLabCut environment first, before running:

$ pip install dlclib

Classes

dlclib.estimate.TFSession

A thin wrapper over the DeepLabCut inference network based on TensorFlow.

• session = TFSession.from_config(cfg) will return a TFSession object using the specified config.yaml file of the DeepLabCut project.
• session.get_pose(image) will return a numpy.ndarray corresponding to the estimation. for a batch-processing network, use a list of images, or a 4-D uint8 array whose axis 0 correspond to indexes of images.
• session.open_writer(path) opens a CSVWriter at the specified path (see below).

dlclib.posture.PostureMapper

A utility class that maps the output of TFSession to a dictionary. Note that the current implementation:

• does not care anything about conversion speed.

• currently only supports non-batch models.

• mapper = PostureMapper.from_config(cfg) will return a PostureMapper object using the specified config.yaml file of the DeepLabCut project.

• posedict = mapper.map(pose) receives the output of TFSession.get_pose(), and returns a dictionary. posedict[part] is a named tuple that has x, y and prob attributes.

dlclib.writer.CSVWriter

A simple class for writing the estimation into a CSV file. This class does support batch-processing networks.

• writer = session.open_writer(path) from a TFSession object initializes the writer associated with the inference session.
• For every video frame, you can call writer.push(frame). It automatically calls TFSession.get_pose() internally, and writes its outputs into the file.
• Calling writer.flush() will flush the pended video frames (in case of batch modes; it does not wait for more frames and runs get_pose() to writes the estimation).
• Calling writer.close() will flush its internal buffer, close the file and invalidates itself. You can keep using the originating TFSession object.

This class supports the context-manager model, so you can write:

with session.open_writer(path) as out:
    for frame in video_frames:
        out.push(frame)

License

2020-2021 Keisuke Sehara, the MIT License