How to use Randomizable third party transforms or customered transforms for CacheDataset?

[Jingnan-Jia]

This jupyter shows how to insert third party transforms (from torchIO and rising) into monai. But all of the transforms are not Randomizable, and it used the normal Dataset instead of CacheDataset or SmartCacheDataset.

I noted that The CacheDataset or SmartCacheDataset will cache the fixed transforms at first, leave the Randomizable transforms to be implemented by Dataloader.

The Randomizable transforms from Monai inherite Randomizable class which would be cached in this line.

It means that only the transforms which inherite Randomizable and Transforms would be possible to be cached.

However our customed transforms or the transforms from torchIO does not inherite Randomizable and Transforms, which means our customed transforms and the torchIO.transforms can not be cached by CacheDataset or SmartCacheDataset.

For instance, this is my own customed transform to load data (I hope it could be cached):

class LoadDatad:
    def __call__(self, data: Mapping[str, Union[np.ndarray, str]]) -> Dict[str, np.ndarray]:
        path = data['path_key']
        x = load_itk(fpath, require_ori_sp=True)

        data = {'image_key': x,
                    'path_key': path}  
        return data

If I use this transform as the first transform in my composed transform list, I found this transform will not be cached at all.

So I think it is necessary to remind users in the documentation that "If you want your own transforms be cached by monai.CacheDataset/SmartCacheDataset/PersistentDataset, you have to ensure: 1. your transforms inherite monai.Transform 2. your transforms can not be Randomizable "

Another solution is: changing this line:
from
if isinstance(_transform, Randomizable) or not isinstance(_transform, Transform):
to
if isinstance(_transform, Randomizable):

Is my understanding right?

[Nic-Ma]

Hi @Jingnan-Jia ,

Thanks for your investigation and analysis & suggestions.
I submitted PR to enhance the doc-string of cache datasets: #2799.

Thanks.