Solos Python library


Clone the repository in your local:

git clone

Install the requirements:

pip install -r requirements.txt

Call the downloader:

python from_json /path_to_your_dst

If some video fails to be downloaded, a backup copy with the remaining videos is created at:


So that you can resume from:

python from_json /path_to_your_dst /path_to_your_dst/backup.json

You can directly import Solos from appending that folder to pythonpaths or installing via pip


For simplicity solos can be installed via pip. (Realize youtube script to download videos is not part of the package)

pip install solos

Skeletons and Numpy

Skeletons are provided as a numpy memory map file in order to have a versatile scalable way of reading them.
The class Solos.SkReader automatically downloads the refined skeletons and exposes a np.mmap instance. You can query SkReader with video keys to obtain a np.mmap array with the skeletons of the chosen video. Besides you can slice the original array which cointains all the skeletons.

import solos
sk_npy = solos.SkReader(download=True,in_ram=False)
sk = sk_npy[youtube_id] # This is a np.mmap if in_ram = False or np.ndarray if in_ram=True
#sk shape is Nx3x47 --> N,[x,y,c],47  where N is the total amount of frames

YouTube IDs

YouTube IDs are exposed by calling Solos.get_solos_ids() as a dictionary whose keys are the categories (instruments) present in the dataset and its values are lists of YouTube IDs (corresponding to the key category).
Dataset timestamps are exposed by calling Solos.get_solos_timestamps().
Corresponding paths to the files can be found at Solos.SOLOS_IDS_PATH and Solos.SOLOS_TIMESTAMPS_PATH.

Downloading the data (Optional)

The numpy array should be automatically downloaded. In case anything fails the .npy file can be downloaded from GDrive 3Gb. Index json file is already included in the python package. To open it in reading mode:
npy = np.memmap(path, dtype=np.float32, mode='r', shape=(N, 3, 47))
For extra info check np.mmap docs