Comparision between different Datasets

[Jingnan-Jia]

At first, thanks very much for the development of MONAI. I love it!

I found MONAI provided CacheDataset, LMDBDataset, PersistentDataset, and SmartCache for the acceleration of data loading and transform.

My understanding of the 4 Datasets:
If I have 1000 3D CT scans for training,

1. CacheDataset will cache cache_num cases in memory before random transforms for training. So the data loading will be very fast for the first cache_num training samples and become slower for the rest 1000-cache_num samples in each epoch. The disadvantage is it require a lot of cpu memory if we want to cache all 1000 3D CT scans in one shot.
2. SmartCache will cache cache_num cases in memory before random transforms for training. But part of the cache_num cases will be replaced before next epoch according to replace_rate. So the data loading speed is stable for each batch of training data. This one seems better if we can not cache all 1000 3D CT scans in one shot.
3. PersistentDataset will save all cases in disk before random transforms. And load them again during random transforms. It will still load data from disk, but the loading time would be shorter because loading tensors seems faster than loading medical images.
4. I know little on LMDBDataset because I do not have the experience of LMDB database.

My question is:
Among the last 3 Datasets, do you have a recommendation on which Dataset is the fastest one? SmartCache, PersistentDataset or LMDBDataset ?

[Nic-Ma]

Hi @Jingnan-Jia ,

Thanks for your interest here.
You can find detailed comparison and tutorial in this notebook:
https://github.com/Project-MONAI/tutorials/blob/master/acceleration/dataset_type_performance.ipynb

Thanks.