Multianimal training dataset does not contain all cropped images

I have previously worked with the single animal workflow, but in attempting to use maDLC to track a single animal I’ve encountered an issue.

Extract 20 frames from each of 2 videos and label. Check labels - all looks good.
If I run deeplabcut.create_multianimaltraining_dataset at this stage I get what I expect: a training set of 38 uncropped images (40*0.95)
However, if I run deeplabcut.cropimagesandlabels and verify that the cropped images were produced… then deeplabcut.create_multianimaltraining_dataset I get this bizarre mixture of uncropped images from the first video and cropped images from the second.

Could the issue be related to this error that deeplabcut.cropimagesandlabels throws?

Should the training set used in batch-based training of the network contain only croppped, or cropped and uncropped labeled images?

Thanks for any help,
Aki

Neither issue occurs when working with a single video - as opposed to >1 in a project. The OrderedDict error of cropimagesandlabels, nor the training set issue of create_multianimaltraining_dataset.

I suspect it is my video path and video file names that are incompatible with the dictionary operations used by cropimagesandlabels. I will try to adjust them tomorrow and see if that fixes the issue.

FYI: after cropimagesandlabels, only cropped images should make it to the training dataset, so yes, that error is the problem.

You seem to have a full stop (".") in your file path. That might be a problem.