`Pad` uses numpy though pytorch backend is available

[function2-llx]

Describe the bug
According to the doc string of Pad:

MONAI/monai/transforms/croppad/array.py

Lines 88 to 89 in f519699

	`torch.nn.functional.pad` is used unless the mode or kwargs are not available in torch,
	in which case `np.pad` will be used.

However, sometimes Pad uses numpy though pytorch backend is available.

To Reproduce
Run the following script, which wraps pad of numpy and torch by printing corresponding information first, and pads the input with Monai's transform to see the output:

import torch

from monai.transforms import Pad
from monai.transforms.croppad import functional as F
from monai.utils import PytorchPadMode

class Wrap:
    def __init__(self, func, info: str):
        self.func = func
        self.info = info

    def __call__(self, *args, **kwargs):
        print(self.info)
        return self.func(*args, **kwargs)

if __name__ == '__main__':
    F._np_pad = Wrap(F._np_pad, 'np')
    F._pt_pad = Wrap(F._pt_pad, 'pt')
    pad = Pad(mode=PytorchPadMode.CONSTANT)
    x = torch.randn(1, 5, 5)
    print(pad(x, [(0, 1), (1, 1), (0, 0)]).shape)

Expected behavior

PyTorch is available in the script, so it should output pt (indicating that PyTorch pad is used)

Actual behavior
The script outputs np (indicating that numpy pad is used).

Environment

================================
Printing MONAI config...
================================
MONAI version: 1.1.0+88.gf519699f
Numpy version: 1.24.2
Pytorch version: 1.13.1.post200
MONAI flags: HAS_EXT = False, USE_COMPILED = False, USE_META_DICT = False
MONAI rev id: f519699f4b709aa69f23c252842e4798cac293d8

Optional dependencies:
Pytorch Ignite version: 0.4.10
ITK version: 5.3.0
Nibabel version: 5.0.0
scikit-image version: 0.19.3
Pillow version: 9.2.0
Tensorboard version: 2.12.0
gdown version: 4.6.2
TorchVision version: 0.14.1a0+b69fce3
tqdm version: 4.64.1
lmdb version: 1.4.0
psutil version: 5.9.4
pandas version: 1.5.3
einops version: 0.6.0
transformers version: 4.21.3
mlflow version: 2.1.1
pynrrd version: 1.0.0

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 22.04.2 LTS
Platform: Linux-5.19.0-32-generic-x86_64-with-glibc2.35
Processor: x86_64
Machine: x86_64
Python version: 3.10.9
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 (%): [2.0, 2.7, 5.3, 3.4, 4.1, 2.1, 2.7, 0.0, 3.4, 4.2, 7.2, 12.2, 4.8, 6.2, 2.1, 97.9]
CPU freq. (MHz): 3733
Load avg. in last 1, 5, 15 mins (%): [2.6, 2.4, 1.9]
Disk usage (%): 47.3
Avg. sensor temp. (Celsius): UNKNOWN for given OS
Total physical memory (GB): 62.7
Available memory (GB): 32.0
Used memory (GB): 29.7

================================
Printing GPU config...
================================
Num GPUs: 1
Has CUDA: True
CUDA version: 11.2
cuDNN enabled: True
cuDNN version: 8401
Current device: 0
Library compiled for CUDA architectures: ['sm_35', 'sm_50', 'sm_60', 'sm_61', 'sm_70', 'sm_75', 'sm_80', 'sm_86', 'compute_86']
GPU 0 Name: NVIDIA GeForce GTX 1080 Ti
GPU 0 Is integrated: False
GPU 0 Is multi GPU board: False
GPU 0 Multi processor count: 28
GPU 0 Total memory (GB): 10.9
GPU 0 CUDA capability (maj.min): 6.1

Additional context

I think the condition for PyTorch here is too strict.

MONAI/monai/transforms/croppad/functional.py

Lines 70 to 74 in f519699

	_pad = (
	_pt_pad
	if mode in {"reflect", "replicate"} and img.dtype not in {torch.int16, torch.int64, torch.bool, torch.uint8}
	else _np_pad
	)

[wyli]

thanks for reporting, would you like to help enhance the logic? the main testing cases are listed here https://github.com/Project-MONAI/MONAI/blob/dev/tests/test_pad_mode.py

[function2-llx]

Hi @wyli, I'd like to, but maybe two weeks later, since only then will I have enough time.

#4749 introduced this behavior for fixing another issue. Could you please give some hints on how to keep these compatible?

[function2-llx]

I see, according to pytorch/pytorch#40763 (comment), torch.nn.functional.pad does not support input tensors with dtype other than float32 or float64 for modes "replicate" and "reflect". The fix will be simple, change this condition:

MONAI/monai/transforms/croppad/functional.py

Lines 70 to 74 in f519699

	_pad = (
	_pt_pad
	if mode in {"reflect", "replicate"} and img.dtype not in {torch.int16, torch.int64, torch.bool, torch.uint8}
	else _np_pad
	)

to:

 _pad = ( 
     _np_pad 
     if mode in {"reflect", "replicate"} and img.dtype in {torch.int16, torch.int64, torch.bool, torch.uint8} 
     else  _pt_pad
 ) 

Or, we can just use PyTorch pad here first and use NumPy pad if any exception occurs. I'm not sure which one is desired.

[wyli]

ok, I remember the _pt_pad was not very robust, it's only compatible with a few combinations of dtypes + padding modes. I can have a look soon.