geo-deep-learning
geo-deep-learning copied to clipboard
NRCan
Deep learning applied to georeferenced datasets
Features from #134 are still present in GDL. Have they been used lately? Maybe time for deprecation.
https://github.com/NRCan/geo-deep-learning/blob/develop/train_segmentation.py#L114 The visualization tool is designed to let user see outputs of model as it is being trained and see the evolution of the predictions on the same samples epoch...
This is the Epic for all aspects of GDL that relate to operations. The ultimate goal is to provide services for 1) inference and 2) training. See this link for...
Currently, models are predicting on the input image, even in NaN areas. There should be a quick filtering of the prediction values for those areas. They should be overwritten to...
Currently, GDL uses Data Parallel (DP). According to the pytorch documentation, Distributed Data Parallel (DDP) is much faster the DP. From [pytorch documentation](https://pytorch.org/docs/master/notes/cuda.html#cuda-nn-ddp-instead) "There are significant caveats to using CUDA...
Currently, training is performed on a list of GeoTIFF input images using reference data in GeoPackage files. That list of inputs is stored [in csv files](https://github.com/NRCan/geo-deep-learning/blob/0f63e25f2e4ed67e2b5f9038462403b2cf97ca63/README.md#csv-preparation). For the results we...
Once sat and label data is written to hdf5, it becomes more opaque and difficult to verify. Minor misalignements and other defects, minor or major, between gpkg and source rasters...
Multiband VRT data is created and returned as a raster input (String XML) this is a departure from the previous functionality in which a raster file is opened as a...
