Shape mismatch ValueError when using RandGaussianNoised with more than one key

[LucasFidon]

Describe the bug
When using RandGaussianNoised during training with more than one key the error "ValueError: shape mismatch: objects cannot be broadcast to a single shape" is returned.

To Reproduce
I have encountered the error when using RandGaussianNoised in my training pipeline for BraTS.

I have in my composition of training transforms:
RandGaussianNoised(keys=["flair", "t1", "t1ce", "t2"], mean=0., std=0.1, prob=0.15),

For this example, in

MONAI/monai/transforms/intensity/dictionary.py

Line 113 in 8207e1e

self.mean = ensure_tuple_size(mean, len(self.keys))

we obtain self.mean = (0.,0,0,0).

This creates a mismatch between the shape of self.mean and im_shape in

MONAI/monai/transforms/intensity/dictionary.py

Line 120 in 8207e1e

self._noise = self.R.normal(self.mean, self.R.uniform(0, self.std), size=im_shape)

where im_shape=(1,128,128,128).

A Simple Solution
By replacing

MONAI/monai/transforms/intensity/dictionary.py

Line 113 in 8207e1e

self.mean = ensure_tuple_size(mean, len(self.keys))

by self.mean = mean the error does not appear anymore.

Expected behavior
The reason why the tests pass with the current version of the code is that the test https://github.com/Project-MONAI/MONAI/blob/master/tests/test_rand_gaussian_noised.py only contains examples with len(keys)==1.

In addition, in

MONAI/monai/transforms/intensity/dictionary.py

Lines 131 to 133 in 8207e1e

	for key in self.keys:
	dtype = dtype_torch_to_numpy(d[key].dtype) if isinstance(d[key], torch.Tensor) else d[key].dtype
	d[key] = d[key] + self._noise.astype(dtype)

the same sample of Gaussian noise is used for all the keys.
I would have expected that different noise samples would be used with different keys.
Is it the intended behavior for this transformation?

I can work on a pull request if you can confirm which behavior you would like to see here.

Thank you,
Lucas

[saruarlive]

@LucasFidon,
Based on the bug, so it handles a single dictionary key, right!
If you stack ["flair", "t1", "t1ce", "t2"] images in one tensor, (4, :, : , : ) before, then use the RandGaussianNoised class using one dictionary key, does it raise error then?

[LucasFidon]

Hi @saruarlive
Yes, that seems to be a good workaround.
But I suppose RandGaussianNoised is intended to work with any number of keys so that that does not solve the issue.