I have been trying out the DL training tool and I have a few feature requests you may wish to consider:
1. Allow image uploads to support subfolders within the "Label Folders".
When dealing with very large image data sets often the images for any one label will be stored in a multitude of subfolders. Rather than having to flatten this out before uploading it would be good if the tool considered any images in subfolders to the label folder to be associated with that label.
2. Allow fine tuning with custom pre-trained models and custom topologies.
Sometimes one wants to reference weights from a pretrained model other than imagenet.
3. Allow model analysis to be filtered by epoch.
Sometimes the last epoch is not the best, therefore it would be good if you could review, in the model analysis tab, results from earlier epochs.
4. Allow downloading of earlier snapshots other than the latest snapshot.
Sometimes the best snapshot is not the last snapshot, so being able to download an earlier snapshot would be useful.
If any of these features are already available in the current version please let me know.