Custom our data


I already labeled my images by another tool and I wonder to use them by DLC. As there is no video file available from the images, I create a training csv file such as multi-mouse project by myself. I am wondering how can I train a model from the available labeled images with DLC without video file?


Hi! It seems like the only thing you’re missing is creating a folder under labeled-data that contains your frames as well as a CollectedData.h5 file of annotations. You’d then need to add a path to a “fake” video having the name of the folder to the video_sets entry in the config.yaml; this will guarantee the folder is found when creating the training dataset

Hi Jessy,

Thank you so much for your reply! Yes, I already created a folder under labeled-data with the name of fake video consists of a CSV annotation file and our images. How can I generate h5 file from CSV?

Hi Fahimeh,

There’s a few ways of doing this. It can be done from the terminal:

python -c "import pandas as pd; pd.read_csv('input_file.csv').to_hdf('output_file.h5', key='data')"

For other ways, have a look here (all Python implementations):

HI, also this is on our wiki, see here!

1 Like

Thanks, it is done :slight_smile: