Adjusting Original Labels

I’m trying to figure out how to adjust the original labels. I can extract outlier frames and then refine labels but I don’t see the option or checkbox in the GUI to adjust original labels. Am i looking in the wrong place? Anyone have any ideas?


1 Like

I think you can just use deeplabcut.label_frames(config_path), select the folder where your extracted frames are and the GUI should show you the labels you already create it and you should be able to adjust them.

1 Like

@cgutierrez-Ibanez: yes, indeed. Sorry for the confusion, the new version (newer than 2.0.6) is different from the biorxiv paper in that sense, but the updated paper is in press :slightly_smiling_face:

1 Like

thank you! when you run the label frames, does it change anything if I already had refined frames in the previous iteration? I ran deeplabcut.label_frames and created a training set but it seems like that rewrote the most recent iteration instead of creating a new one.
Would that be because I didn’t merge the training sets?
Thanks for your help!

yes, if you didn’t merge then the iteration # is not updated in the config.yaml file.

I wonder if I could get some specific guidance on adjusting the original labels. Occasionally (<5% of the time), the labels appear in the entirely incorrect portion of the image. When I use deeplabcut.label_frames(config_path), the image appears with non-movable crosshairs and by right clicking, I can put filled circles onto the correct locations. Unfortunately, those circles now appear in all subsequent frames of the video. What am I doing wrong? Also, is there a way to simply exclude those specific frames from the training?

seems you are opening the labeled-data folder that is created when you run “check labels”, not the original folder of unmarked images.

You can also delete images from the folder you don’t want in go into the training set. Then you need to run (the undocumented function):


Signature: deeplabcut.dropannotationfileentriesduetodeletedimages(config)
Drop entries for all deleted images in annotation files, i.e. for folders of the type: /labeled-data/*folder*/CollectedData_*scorer*.h5
Will be carried out iteratively for all *folders* in labeled-data.

config : string     
    String containing the full path of the config file in the project.

Thank you!

Using deeplabcut.dropannotationfileentriesduetodeletedimages, the frames were successfully dropped. However, when I subsequently ran deeplabcut.train_network(path_config_file), I got the following error message:

UnknownError Traceback (most recent call last)
C:\Users\harrisa\AppData\Local\Continuum\anaconda3\envs\dlc-windowsGPU\lib\site-packages\tensorflow\python\client\ in _do_call(self, fn, *args)
1333 try:
-> 1334 return fn(*args)
1335 except errors.OpError as e:

C:\Users\harrisa\AppData\Local\Continuum\anaconda3\envs\dlc-windowsGPU\lib\site-packages\tensorflow\python\client\ in _run_fn(feed_dict, fetch_list, target_list, options, run_metadata)
1318 return self._call_tf_sessionrun(
-> 1319 options, feed_dict, fetch_list, target_list, run_metadata)

C:\Users\harrisa\AppData\Local\Continuum\anaconda3\envs\dlc-windowsGPU\lib\site-packages\tensorflow\python\client\ in _call_tf_sessionrun(self, options, feed_dict, fetch_list, target_list, run_metadata)
1406 self._session, options, feed_dict, fetch_list, target_list,
-> 1407 run_metadata)

UnknownError: Failed to get convolution algorithm. This is probably because cuDNN failed to initialize, so try looking to see if a warning log message was printed above.
[[{{node resnet_v1_50/conv1/Conv2D}}]]
[[{{node sigmoid_cross_entropy_loss/value}}]]

I’m not sure but either TF error or ResNet not there?

Well, I rebooted and training is underway! fingers crossed, looks like it worked. Thanks again!

1 Like