Nicolas Audebert
Nicolas Audebert
Since PyTorch 4.0, [some things have been deprecated and should be fixed](https://pytorch.org/2018/04/22/0_4_0-migration-guide.html): - [ ] `Variable` does not exist anymore - [ ] `loss.data[0]` should now be `loss.item()` - [...
Refactor
- [x] Rewrite disjoint sampling method - [x] Move sklearn models into their own file - [x] Rewrite the Dataset - [x] Add parallelization option (see #32) - [x] Use...
Current approach to class balancing is to use inverse median frequency loss reweighting. Other options could be: ### Resampling Resample the dataset (e.g. upsample minority classes or downsamples majority classes)...
`torchvision` defines [Transforms](https://pytorch.org/docs/stable/torchvision/transforms.html) objects to apply data augmentation and other transformations to data. We could and should define our own custom Transforms. Pros: - easier to reuse the data augmentation...
Currently the dataset is [normalized into (0,1)](https://github.com/nshaud/DeepHyperX/blob/99792c5aec51a9602099b4b1c6618af01b652e09/datasets.py#L319). This is mostly fine but we should be able to use other normalizations or even no normalization at all. TODO: - [ ]...
_Maybe out of scope for this toolbox_. The toolbox currently work with the assumption that the models are supervised. Working with unsupervised models (e.g. autoencoders) could be helpful. - [...
Currently, the torch DataLoader uses blocking data loading. Although loading is very fast (we store the NumPy arrays in-memory), transfer to GPU and data augmentation (which is done on CPU)...
The current spatially disjoint train/test split divides the image in 2 for each class. However there might be spatial correlations between the pixels for those regions and this approach is...