Cache invalidation for PersistentDataset

[mattwarkentin]

Hi,

I am wondering if the expectation is that the user would have to manually invalidate the cache when the transforms are updated/changed?

For example, when I run the code below, when I access the first subject (e.g. ds[0]), there is a normal wait time while the transforms are performed. The next and all subsequent times are based on the cache and are very fast. But then when I change one of the transforms, I would expect this to invalidate the cache, but in fact the cached value is returned.

import monai
from monai.transforms import (
    AddChanneld,
    Compose,
    LoadImaged,
    ToTensord,
    Orientationd,
    ScaleIntensityRanged,
    CenterSpatialCropd,
    Spacingd
)

subject = {
  "img": "image.nrrd",
  'mask': "mask.nrrd"
}

keys = list(subject.keys())

trans = Compose([
  LoadImaged(keys),
  Orientationd(keys, 'RAS'),
  Spacingd(keys, pixdim=(1, 1, 1)),
  AddChanneld(keys),
  ScaleIntensityRanged('img', -1000, 1000, 0, 1, True),
  CenterSpatialCropd(keys, roi_size=[50, 50, 50]),
  ToTensord(keys)
])

ds = monai.data.PersistentDataset([subject], transform=trans, cache_dir="cache/")

ds[0] # slow
ds[0] # fast

## Change crop size
trans = Compose([
  LoadImaged(keys),
  Orientationd(keys, 'RAS'),
  Spacingd(keys, pixdim=(1, 1, 1)),
  AddChanneld(keys),
  ScaleIntensityRanged('img', -1000, 1000, 0, 1, True),
  CenterSpatialCropd(keys, roi_size=[75, 75, 75]),
  ToTensord(keys)
])

ds = monai.data.PersistentDataset([subject], transform=trans, cache_dir="cache/")

ds[0] # fast from cache
ds[0]['img'].shape
# torch.Size([1, 50, 50, 50])

[wyli]

this is also discussed in #999