Adam J. Stewart
Adam J. Stewart
**Environment:** 1. Framework: PyTorch 2. Framework version: 1.12.0 3. Horovod version: 0.25.0 4. MPI version: OpenMPI 4.1.4 5. CUDA version: N/A 6. NCCL version: N/A 7. Python version: 3.9.13 8....
Original authors got RMSE: 0.657, MAE: 0.248 with ResCeption, I got RMSE: 0.597, MAE: 0.290 with ResNet-50 Optimal hyperparams are pre-trained ResNet-50 with a learning rate of 0.0001. Still need...
I believe this was a holdover from 0.1.0 back when `IntersectionDataset` didn't have its own index. Current best practice is to use `IntersectionDataset` instead of one of the individual datasets...
### Summary Our current `train.py` script only logs to TensorBoard. We should add a CSV logger. ### Rationale In order to determine the optimal hyperparams in a hyperparam sweep, we...
See #674 and #675. I identified several other places where this bug may still occur: https://github.com/microsoft/torchgeo/pull/675#issuecomment-1184802967 We should investigate and fix these.
### Summary We should make our tutorial notebook tests faster. ### Rationale Our tutorial notebooks currently take 40+ minutes to test. We hope to greatly expand the number of tutorials...
### Summary We should remove dataset-specific arguments from datamodules and instead pass them directly to the dataset through kwargs. ### Rationale As an example, we'll look at `TropicalCycloneWindEstimation` and `CycloneDataModule`....
I never know the best way to create true color images from satellite imagery. QGIS normalizes images based on percentiles, but that blows up depending on whether or not your...
### Summary We should add a `torchgeo.datasets.utils.DatasetNotFoundError` class for consistent error messages when datasets aren't found. ### Rationale There is currently very little consistency between the errors reported by different...