The app is beautifully done. I'm really impressed by how well it works given the knobs available.
However I tried to train it to recognize some images of characters from an anime (so a little different than facial recognition), and I managed to break the model: achieving 64% error with significant number of examples per class. I think one downside is Lobe doesn't expose how potentially overconfident the model is. I would love the ability to take the existing model and test it on a new image that I can import into the app.
EDIT: I would love to see the following in a future version:
1. What are the percentages associated with each image per class. I see that an image was misclassified, but did it at least include my desired class in its top 5 predicted classes?
2. Test the model on unlabeled inputs directly in the app to see how well the model might generalize. I would like to see a "Test" tab on the left once training is complete.
3. View other metrics of model goodness like F-1 score and training details like CV partitions in the app somehow.
Hey there! Thank you so much for the feedback, we are planning improvements like this.
Here's a few tips for now:
1. You can view by "Test Images" on the Train tab (view options). So you can see how well your model is performing on your test images (a random 20% split from all of your images).
2. You can test your model on the Play tab, by dragging in new images your model has not seen, to see how well it is performing. You can also tell Lobe if it was correct or not and iteratively improve your model.
However I tried to train it to recognize some images of characters from an anime (so a little different than facial recognition), and I managed to break the model: achieving 64% error with significant number of examples per class. I think one downside is Lobe doesn't expose how potentially overconfident the model is. I would love the ability to take the existing model and test it on a new image that I can import into the app.
EDIT: I would love to see the following in a future version:
1. What are the percentages associated with each image per class. I see that an image was misclassified, but did it at least include my desired class in its top 5 predicted classes?
2. Test the model on unlabeled inputs directly in the app to see how well the model might generalize. I would like to see a "Test" tab on the left once training is complete.
3. View other metrics of model goodness like F-1 score and training details like CV partitions in the app somehow.
Again, this is a really cool idea :)