I tested out Lobe

If you haven’t heard about it yet, Lobe is a sleek new tool from Microsoft that lets you train models without writing any code. You can watch an impressive walktrough video on their website: lobe.ai.

What does it actually do? For now you can only do image classification and have two models to chose from resnet-50 mobilenetV2. The training happens on your local computer and it doesn’t provide GPU support (at least yet).

I was curious what the performance and accuracy are so I tested it on the Oxford Pets dataset. It took around an hour to train on my PC and reached 85% accuracy. Hour worth of training with fastai on the same PC gives 88%. Of course you can do much better and faster if you move to a GPU.

My conclusion is that it’s a cool tool for beginners to prototype quickly on small datasets, doesn’t seem to be very useful beyond that.

3 Likes

I had the same impression, I am building a video analysis and tried the software, first you can not export the model and it takes hours to train but one beautiful feature is the continual learning to improve the model

Do you mean improving as you add more samples? I wonder how it’s done exactly, my initial idea would be to run another epoch with the new images oversampled. But maybe there is a more clever way.