search

danbider / lightning-pose

Accelerated pose estimation and tracking using semi-supervised convolutional networks.
MIT License
235 stars 34 forks source link

Starts training from specific checkpoint #143

Closed NgD-Huy closed 5 months ago

NgD-Huy commented 5 months ago

Hello and happy Friday,

I couldn't find these topics in the documentation so some clarification would be appreciated.

When adding new videos to an existing project,

  1. Is it possible to label novel videos using the current model or do we still need to label frames?
  2. Is it possible to continue training from an existing checkpoint?

Thank you very much for your work!

themattinthehatt commented 5 months ago

Hi @NgD-Huy, thanks for the questions.

  1. can you clarify what you mean by "label novel videos"? If you're asking about the model producing predictions on every frame in a new video, yes absolutely - once you have a trained model you can run inference on any new video you wish; see docs here: https://lightning-pose.readthedocs.io/en/latest/source/user_guide/inference.html Or are you asking if you can use a model to create new "ground truth" labels that you then use for training a new model?

  2. yes! in the config file, under the "model" section, add a new parameter named "checkpoint", and set the parameter to the absolute path of whatever checkpoint file you want config.yaml:

    model:
checkpoint: /path/to/model.ckpt

    I will make sure to update the documentation to include this info!

NgD-Huy commented 5 months ago

That answered all of my questions! Thank you very much!

themattinthehatt commented 5 months ago

closed by #147