Refactor

[nshaud]

• Rewrite disjoint sampling method
• Move sklearn models into their own file
• Rewrite the Dataset
• Add parallelization option (see Parallelize data loading #32)
• Use a unique IGNORED_INDEX value for all ignored pixels
• Use sklearn.metrics everywhere needed (especially in validation)
• Move data exploration/data visualization functions into their own file
• Rewrite the build_dataset function
• Main script uses a main function
• Unify val and test function
• Deal with the varying tensor sizes when using : spectra (1D), images (2D), cubes (3D).
• Unify segmentation and classification datasets
• Use Sequential API for simple models
• Add --overlap options for training and test
• Use scheduler/auto LR reduction (see Use ReduceLROnPlateau as a default scheduler #22)
• Save output image after training
• Rewrite data augmentation as torchvision.transforms (see Defined hyperspectral-specific Transforms #33)
• Move downloaders into their own script
• Add other class balancing schemes (see Class balancing/resampling #39)
• Add IoU/dice score loss
• Improve cross-validation support
• Optional: Simplify dataset configuration
• Optional: Support defining a dataset as a collection of HSI images and GT masks

[3097108366]

Where is the output image after training saved,I really can't find it