We haven't released pre-trained models yet. Mostly for two reasons: 1/ The PyTorch checkpoints depend on the specific Python model class. Even if you refactor only e.g. a MaxPool layer into a direct functional.max_pool function call, loading old checkpoints will no longer work. We have an ONNX model exporter now (rs export) which allows for self-contained and portable protobuf model and weight files. This workflow needs some more time and careful evaluation, though. 2/ The models for Tanzania I was working on in my spare time I can open up for sure. If there is community interest maybe we can come up with a publicly available model catalogue hosting ONNX models and metadata where folks can easily upload and download models. For our internal models and the data we extract we are thinking through a broader strategy since a lot of time and resources are going into creating and cleaning datasets, doing hard-negative mining, running multiple training iterations and so on. They're also bound to the Mapbox aerial imagery on specific zoom levels.
The model architecture is kept simple on purpose. It used to be an encoder-decoder U-Net'ish architecture which we trained from scratch. Recently (https://github.com/mapbox/robosat/pull/46) I switched out the encoder to a pre-trained ResNet, as proposed by Alexander Buslaev. It's a mix of the papers listed in the docstring at the top with a focus on simplicity and maintainability:
Internally we were also exploring a multi-class PSPNet but decided not to move forward with it right now: the RoboSat model is currently a binary model (feature vs. background) which makes a few things easier in practice, such as efficiently storing results which is needed when scaling it up e.g. to all of North America.
Personally I always try pre-trained models very very useful.
If I'm working in a new domain (which this is to me) then I prefer to get the workflow right (files in the right directories etc) before changing the NN architecture. It's a pretty big time investment to train a NN just to try it.
The model architecture is kept simple on purpose. It used to be an encoder-decoder U-Net'ish architecture which we trained from scratch. Recently (https://github.com/mapbox/robosat/pull/46) I switched out the encoder to a pre-trained ResNet, as proposed by Alexander Buslaev. It's a mix of the papers listed in the docstring at the top with a focus on simplicity and maintainability:
https://github.com/mapbox/robosat/blob/1e687552fe9b254a14d55...
Internally we were also exploring a multi-class PSPNet but decided not to move forward with it right now: the RoboSat model is currently a binary model (feature vs. background) which makes a few things easier in practice, such as efficiently storing results which is needed when scaling it up e.g. to all of North America.