azavea / raster-vision-backend-plugin Goto Github PK

If someone wishes to work on a more robust implementation of this GluonCV backend for Raster Vision, below are some ideas:

• Support more GluonCV models. Currently only ResNet50v2 is supported.
• Benchmarking with existing Keras Classification backend.
• Support more tasks. Currently, only chip classification is implemented; however, GluonCV also supports object detection, semantic segmentation, and instance segmentation.
• Use full Spacenet Vegas dataset. Currently, the largest subset used for training was about one third of the full dataset.
• Use new dataset. Recommendation: COWC Potsdam
• Improve batch size. Currently, something about CUDA's configuration does not work with a batch size of greater than 32.

Time estimate: 2 weeks

You will also need to support:

• the bundle command by implementing save_bundle_files and load_bundle_files in the config.
• resuming training a model from a checkpoint and the replace_model option (see other backends for what this means)

The goal is to implement the methods in the backend (and any stuff needed in the config and configbuilder) that allow the chip command to generate training chips in the gluoncv format to be consumed by the training routine. Time estimate: 1 week

prep tasks:

• install gluoncv
• download data
• study keras_classification backend and config

tips:

• get it to work locally first (might want to make a PR for this, and a second PR that adds and tests functionality for running remotely)
• try to use "directory for each class" dataset format if possible. you'll need to think about how the train method will work to see if this is possible.

Time estimate: 1 week

As part of this you'll need to:

• test prediction using a prediction package