dlclibrary 0.0.7

Lightweight library supporting universal functions for the DeepLabCut ecosystem

DLClibrary

DLClibrary is a lightweight library supporting universal functions for the DeepLabCut ecosystem.

Supported functions (at this point):

• API for downloading model weights from the model zoo

Quick start

Install

The package can be installed using pip:

pip install dlclibrary

:warning: warning, the closely named package dlclib is not an official DeepLabCut product. :warning:

Example Usage

Downloading a pretrained model from the model zoo:

from pathlib import Path
from dlclibrary import download_huggingface_model

# Creates a folder and downloads the model to it
model_dir = Path("./superanimal_quadruped_model")
model_dir.mkdir()
download_huggingface_model("superanimal_quadruped", model_dir)

PyTorch models available for a given dataset (compatible with DeepLabCut>=3.0) can be listed using the dlclibrary.get_available_detectors and dlclibrary.get_available_models methods. The datasets for which models are available can be listed using dlclibrary.get_available_datasets. Example use:

>>> import dlclibrary
>>> dlclibrary.get_available_datasets()
['superanimal_bird', 'superanimal_topviewmouse', 'superanimal_quadruped']

>>> dlclibrary.get_available_detectors("superanimal_bird")
['fasterrcnn_mobilenet_v3_large_fpn', 'ssdlite']

>>> dlclibrary.get_available_models("superanimal_bird")
['resnet_50']

How to add a new model?

TensorFlow models

Pick a good model_name. Follow the (novel) naming convention (modeltype_species), e.g. superanimal_topviewmouse.

1. Add the model_name with path and commit ID to: https://github.com/DeepLabCut/DLClibrary/blob/main/dlclibrary/dlcmodelzoo/modelzoo_urls.yaml

2. Add the model name to the constant: MODELOPTIONS https://github.com/DeepLabCut/DLClibrary/blob/main/dlclibrary/dlcmodelzoo/modelzoo_download.py#L15

3. For superanimal models also fill in the configs!

PyTorch models (for deeplabcut >= 3.0.0)

PyTorch models are listed in dlclibrary/dlcmodelzoo/modelzoo_urls_pytorch.yaml. The file is organized as:

my_cool_dataset:  # name of the dataset used to train the model
  detectors:
    detector_name: path/to/huggingface-detector.pt  # add detectors under `detector`
  pose_models:
    pose_model_name: path/to/huggingface-pose-model.pt  # add pose models under `pose_models`
    other_pose_model_name: path/to/huggingface-other-pose-model.pt

This will allow users to download the models using the format datatsetName_modelName, i.e. for this example 3 models would be available: my_cool_dataset_detector_name, my_cool_dataset_pose_model_name and my_cool_dataset_other_pose_model_name.

To add a new model for deeplabcut >= 3.0.0, simply:

• add a new line under detectors or pose models if the dataset is already defined
• add the structure if the model was trained on a new dataset

The models will then be listed when calling dlclibrary.get_available_detectors or dlclibrary.get_available_models! You can list the datasets for which models are available using dlclibrary.get_available_datasets.