Removing the transformations in the validation phase

[victoresque]

Removing the transformations in the validation phase

[SunQpark]

train / valid splitting of BaseDataLoader in this project, or torch.utils.data.random_split only split the indices of dataset. I think the best method of doing this is to have different dataset object for train and validation set.

1. modify your dataset class to have training=True in __init__, and use different data by this flag.
2. initialize train_dataset and valid_dataset respectively.
3. self.valid_loader = torch.utils.data.DataLoader(valid_dataset, ...)
4. override the split_validation method to just return self.valid_loader

[denadai2]

Thank you