The initial goal was to decouple the multi-view aspects of this work and model training. However, I keep writing boilerplate code to train and run inference with mmsegmentation models. I think it might be useful to package this up as a derived segmenter. I've been doing os calls to other command line scripts to actually train the model because the dependencies for mmsegmentation are quite complicated. I think it should stay that way.
Since this is not part of the core workflow, and the idea of calling other scripts at a given location on disk is pretty brittle, it might be worth separating this functionality back out at some point. One option would be to have have a separate repo/package that extends the core functionality. But for now, I'm ok with it being in one package that's not perfectly streamlined.
The initial goal was to decouple the multi-view aspects of this work and model training. However, I keep writing boilerplate code to train and run inference with mmsegmentation models. I think it might be useful to package this up as a derived segmenter. I've been doing
oscalls to other command line scripts to actually train the model because the dependencies for mmsegmentation are quite complicated. I think it should stay that way.Since this is not part of the core workflow, and the idea of calling other scripts at a given location on disk is pretty brittle, it might be worth separating this functionality back out at some point. One option would be to have have a separate repo/package that extends the core functionality. But for now, I'm ok with it being in one package that's not perfectly streamlined.