Jessy Lauer

Hi @auesro, I've run a video analysis for multiple hours on COLAB (the free version) with `use_shelve=True` and have had stable system & GPU RAM usage all along; RAM usage...

@auesro, I came up with a solution in #1820. `use_shelve` will now also affects the behavior of `convert_detections2tracklets`, which should result in very low memory usage! 😃

@auesro, right now 😄 In the notebook, just replace `!pip install https://github.com/DeepLabCut/DeepLabCut/archive/master.zip` by `!pip install git+https://github.com/DeepLabCut/DeepLabCut.git@assembler_shelve` and it should install specifically the patch.

`use_shelve` should be passed to `convert_detections2tracklets` https://github.com/DeepLabCut/DeepLabCut/blob/2f8a87b3fce288c671fd2282fc64506dfcdac4ca/deeplabcut/pose_estimation_tensorflow/predict_videos.py#L629-L640 You must be getting a *_el.pickle alongside the video, could you tell me if it's heavy? Although we're only storing 2D coordinates, I'll...

Hi @QueenKrokuz! Finalizing the assembler_shelve branch still requires some work on my side before it's ready (notably to fix the error your reported and tracklet stitching). I'll also look further...

Hey @kylecnewton, do you have unique body parts defined in the config? There are apparently 2 tracklets to reconstruct a full track, and that should not be a problem unless...

Hey @kylecnewton, could you share with me the faulty pickle file at [email protected]? I could not reproduce the error 😕

I figured out what's wrong and I can't believe I didn't think about it earlier. 😅 When we attempt to reconstruct a single track, `split_tracklets` is set to False with...

Why would you pass `n_tracks=2`? I thought you had only an animal visible in the scene.

Unless debris are meant to be tracked, pass n_tracks=1; DeepLabCut knows how to get rid of false detections. I tried it on the 3 problematic files you shared with me...