RandGaussianNoise changes the type of the input to np.float64

[murraycutforth]

First, thanks for creating a great library!

Describe the bug
Applying the RandGaussianNoise tranform to a np.float32 array changes the dtype to np.float64. This causes a crash down the line since pytorch expects np.float32 inputs.

Expected behavior
The transformed image's dtype does not change

Screenshots

Environment (please complete the following information):

• OS: Centos 7
• Python version: 3.6
• MONAI version: 0.1.0

This could be an easy fix, by changing
return img + self._noise if self._do_transform else img
to
return img + self._noise.astype(img.dtype) if self._do_transform else img .

[murraycutforth]

I forgot to add, it's trivial to work around this by adding a CastToType transform in the transformation pipeline.

[Nic-Ma]

Hi @murraycutforth ,

Thanks for your bug report.
I will submit a PR to fix this issue soon.
Thanks.