Austin

trainer does not work with FolderDataset

1

```AttributeError: 'FolderDataset' object has no attribute 'num_inputs'``` when attempting to call `fit_loader` using an FolderDataset dataset. Same problem with the default torch ImageFolder dataset

running s2p on image crops

3

I'm trying to run s2p on worldview3 pairs without downloading the entire 40k by 40k `full-images`. I can do this with nasa ASP by placing my crops of interest in...

add install from dev instructions

1

install from dev instructions for all notebooks

ModelHparams isn't a general interface as it implies

7

```python @dataclass class ModelHparams(hp.Hparams, ABC): initializers: List[Initializer] = hp.optional( default_factory=lambda: [], doc="The initialization strategy for the model", ) num_classes: Optional[int] = hp.optional( doc="The number of classes. Needed for classification tasks",...

avoid double forward pass for train metrics

5

Our current training loop requires an extra forward pass to compute training metrics which creates a considerable slowdown. Since training metrics are pretty useful, this slowdown could be a dealbreaker...

Generalized Mixup

3

Our current implementation of mixup works directly on targets.. that's why we need a dense loss, or to convert class indexes to 1 hot, like so: ```python if check_for_index_targets(y): y_onehot...