RandZoom change the input type

[bhashemian]

Describe the bug
When using RandZoom with the input type of torch.Tensor it will output np.array!

To Reproduce

import torch
from monai.transforms import RandZoom
a = torch.randn(10, 3, 20, 20)
type(RandZoom(prob=1.0, min_zoom=0.9, max_zoom=1.1)(a))

Expected behavior
expected correct output:

<class 'torch.Tensor'>

but this will output:

<class 'numpy.ndarray'>

Environment

================================
Printing MONAI config...
================================
MONAI version: 0.6.0+117.g97fd2b92
Numpy version: 1.21.1
Pytorch version: 1.10.0a0+ecc3718
MONAI flags: HAS_EXT = False, USE_COMPILED = False
MONAI rev id: 97fd2b9201ea5e21692126539b0bddb99f1f7b23

Optional dependencies:
Pytorch Ignite version: 0.4.5
Nibabel version: 3.2.1
scikit-image version: 0.15.0
Pillow version: 8.2.0
Tensorboard version: 2.5.0
gdown version: 3.13.0
TorchVision version: 0.11.0a0
tqdm version: 4.53.0
lmdb version: 1.2.1
psutil version: 5.8.0
pandas version: 1.1.4
einops version: 0.3.0

For details about installing the optional dependencies, please visit:
    https://docs.monai.io/en/latest/installation.html#installing-the-recommended-dependencies


================================
Printing system config...
================================
System: Linux
Linux version: Ubuntu 20.04.2 LTS
Platform: Linux-5.8.0-59-generic-x86_64-with-glibc2.10
Processor: x86_64
Machine: x86_64
Python version: 3.8.10
Process name: python
Command: ['python', '-c', 'import monai; monai.config.print_debug_info()']
Open files: []
Num physical CPUs: 8
Num logical CPUs: 16
Num usable CPUs: 16
CPU usage (%): [1.4, 0.7, 96.5, 1.4, 3.5, 0.7, 2.1, 3.4, 0.7, 2.0, 0.7, 0.7, 0.0, 0.0, 0.0, 0.0]
CPU freq. (MHz): 1206
Load avg. in last 1, 5, 15 mins (%): [2.8, 3.6, 11.3]
Disk usage (%): 93.5
Avg. sensor temp. (Celsius): UNKNOWN for given OS
Total physical memory (GB): 31.1
Available memory (GB): 25.3
Used memory (GB): 5.3

================================
Printing GPU config...
================================
Num GPUs: 2
Has CUDA: True
CUDA version: 11.4
cuDNN enabled: True
cuDNN version: 8202
Current device: 0
Library compiled for CUDA architectures: ['sm_52', 'sm_60', 'sm_61', 'sm_70', 'sm_75', 'sm_80', 'sm_86', 'compute_86']
GPU 0 Name: NVIDIA TITAN RTX
GPU 0 Is integrated: False
GPU 0 Is multi GPU board: False
GPU 0 Multi processor count: 72
GPU 0 Total memory (GB): 23.7
GPU 0 CUDA capability (maj.min): 7.5
GPU 1 Name: NVIDIA TITAN RTX
GPU 1 Is integrated: False
GPU 1 Is multi GPU board: False
GPU 1 Multi processor count: 72
GPU 1 Total memory (GB): 23.7
GPU 1 CUDA capability (maj.min): 7.5

[bhashemian]

@rijobro @wyli @Nic-Ma

[wyli]

had a quick look, the outcome will be a tensor only if the padding mode is constant

MONAI/monai/transforms/croppad/array.py

Line 113 in 281c17e

if isinstance(img, torch.Tensor) and mode == "constant" and not self.np_kwargs:

perhaps we can make the condition less strict here

[Nic-Ma]

Hi @wyli ,

Thanks for your analysis, let me make a quick fix ASAP.

[rijobro]

The default padding for Zoom is Edge, but torch.pad doesn't support that type of padding. See modes in https://pytorch.org/docs/stable/generated/torch.nn.functional.pad.html and https://numpy.org/doc/stable/reference/generated/numpy.pad.html.

[Nic-Ma]

The default padding for Zoom is Edge, but torch.pad doesn't support that type of padding. See modes in https://pytorch.org/docs/stable/generated/torch.nn.functional.pad.html and https://numpy.org/doc/stable/reference/generated/numpy.pad.html.

Hi @rijobro ,

I think it's OK to use the edge as default value, if passing Tensor to the transform, user must set suitable mode, otherwise the transform can raise mode error.
Just like what my PR did: list all the supported modes for numpy and all the supported modes for Tensor, if not matching, raise an error.
What do you think?

Thanks.

[rijobro]

You're right that we want to minimise the amount of checks that we need to do. But this means that the Zoom will no longer work with torch input using the default padding. Seems like a slight backwards step to me. As mentioned in your comment, maybe our defaults should be constant. Generally speaking, we can try to use defaults that work for both torch and numpy.