geo-deep-learning
geo-deep-learning copied to clipboard
NRCan
Deep learning applied to georeferenced datasets
The data augmentation take a long time to do, here some idea for speed that up: https://pytorch-lightning.readthedocs.io/en/latest/notebooks/lightning_examples/augmentation_kornia.html https://towardsdatascience.com/fast-data-augmentation-in-pytorch-using-nvidia-dali-68f5432e1f5f
Currently, when using inference.py with ground truth files, we average metrics over all predictions made. It would be important to keep per-image metrics for benchmarking purposes. For example, this could...
[Torchgeo's glossary](https://torchgeo.readthedocs.io/en/latest/user/glossary.html) Some examples: - sample -> chip/patch - Use "aoi" and "tile" in documentation: - aoi: Synonym for [region of interest (ROI)](https://torchgeo.readthedocs.io/en/latest/user/glossary.html#term-region-of-interest-ROI). A particular spatial area to focus on....
At inference, our [smoothing window](https://github.com/NRCan/geo-deep-learning/blob/develop/inference_segmentation.py#L181) multiplies raw chip probability maps to set higher confidence to predictions in center of chip. As a result of the convolution with a high overlap...
will let us use: - Distributed Data Parallel (https://pytorch-lightning.readthedocs.io/en/latest/multi_gpu.html) - Float 16 (https://pytorch-lightning.readthedocs.io/en/latest/apex.html) - computing cluster (SLURM) (https://pytorch-lightning.readthedocs.io/en/latest/slurm.html) - Early stopping (https://pytorch-lightning.readthedocs.io/en/latest/early_stopping.html) - export to ONNX (https://pytorch-lightning.readthedocs.io/en/latest/production_inference.html) etc. https://pytorch-lightning.readthedocs.io/en/latest/#common-use-cases
# Model checkpoint naming convention A naming convention for models' weights checkpoints written during training has become essential for: - quickly understanding what flavor of model checkpoint we are dealing...
…on (add returns to train_segmentation.py and run_gdl()) metrics.py: add try/except if missing key model_choice.py: temporarily raise gpu threshold to 100% until issue #246 is addressed train_segmentation.py: - fix bug for...
- [ ] [Copying CONDA_DIR from stage 1](https://github.com/NRCan/geo-deep-learning/blob/37cd99dc016df873ef363a1092e41e01b50eff42/Dockerfile#L73) takes more than 10 GB of temporary disk space; investigate - [x] Build using our GitHub repo as (remote) context `docker build...
Fixing a bug about the `dontcare` value from the segmentation dataset, this value is call in [train_segmentation](https://github.com/NRCan/geo-deep-learning/blob/develop/train_segmentation.py) line [411](https://github.com/NRCan/geo-deep-learning/blob/37cd99dc016df873ef363a1092e41e01b50eff42/train_segmentation.py#L411) and [416](https://github.com/NRCan/geo-deep-learning/blob/37cd99dc016df873ef363a1092e41e01b50eff42/train_segmentation.py#L416). The value was not set in the dataset I...
Geo-deep-learning provides sensible defaults to output filenames of the inference process e.g. in the form of appending `_inference` to the input filename. This is reasonable when running the inference process...