Isolating labelled frames

[rpartan]

I am trying to train a model on very large videos (over 1,000,000 frames). The training never got past Epoch 1 because the computer does not have enough memory to run the whole video. Is there a way I can isolate just the frames I have already labelled to start training?

Thanks,
RP

[gitttt-1234]

Hi @rpartan,

There are a couple of options here:

1. If you know the indices of the labelled frames, then you can explicitly specify the indices for training and validation split in the training_config.json file, as given below. However, this option is only available with CLI (and not GUI). Related issue: Editing previously labelled frames #1962

       "labels": {
           "training_labels": null,
           "validation_labels": null,
           "validation_fraction": 0.1,
           "test_labels": null,
           "split_by_inds": false,
           "training_inds": [2,3,19,24,25,31,34,41,45,53,56,60,...],
           "validation_inds": null,
           "test_inds": null,
           "search_path_hints": [],
           "skeletons": []
       },

2. If you don't have the list of indices, then you can get a new labels file by isolating user_labelled_instances with the below code, using sleap_io:

import sleap_io as sio

labels = sio.load_slp("train.slp")
filtered_lfs = []
for lf in self.labels:
   if lf.user_instances is not None and len(lf.user_instances) > 0:
       lf.instances = lf.user_instances
       filtered_lfs.append(lf)
new_labels = sio.Labels(
   videos=labels.videos,
   skeletons=labels.skeletons,
   labeled_frames=filtered_lfs,
)
new_labels.save("train_labels.slp", embed=True)

The new train_labels.slp will only contain frames with user-labelled instances.

Let me know if you have any questions!

Thanks,

Divya