Frequently asked questions and answers

A list of frequently asked questions and answers maintained by the core developer team, based on the users' feedback and discussions.

Q1. The randomised data transformations generate insufficiently shuffled outputs. How can I resolve this issue?

Q2. How do I sample 2D slices from 3D volumes?

---

Q1. The randomised data transformations generate insufficiently shuffled outputs. How can I resolve this issue?

This is likely because the random number generator in MONAI is duplicated when multiple workers are initialised by torch.utils.data.Dataloader (see also Pytorch FAQ, #398).

A possible solution is to set up random seeds in workers via the worker_init_fn option:

def worker_init_fn(worker_id):
    worker_info = torch.utils.data.get_worker_info()
    try:
        worker_info.dataset.transform.set_random_state(worker_info.seed % (2 ** 32))
    except AttributeError:
        pass

dataloader = torch.utils.data.DataLoader(..., worker_init_fn=worker_init_fn)

Q2. How to sample 2D slices from 3D volumes?

This could be achieved by setting a 3D window size, followed by "squeezing" the length one spatial dimension (see also: #299). For example,

train_transforms = Compose([
    ...
    RandSpatialCropd(keys=['img', 'seg'], roi_size=[96, 96, 1], random_size=False),
    SqueezeDimd(keys=['img', 'seg'], dim=-1),  # remove the last spatial dimension
    ...
])