nn.ZeroPad2d 行为与 pytorch 不一致

[triple-Mu]

问题

nn.ZeroPad2d 行为与 pytorch 不一致

报错

loaded library: /lib/x86_64-linux-gnu/libibverbs.so.1
torch.Size([2, 10, 1, 64, 58, 58])
Traceback (most recent call last):
  File "/home/ubuntu/work/oneflow/vision/flowvision/models/ttest.py", line 16, in <module>
    fOut = fPad(flowTensor)
  File "/home/ubuntu/miniconda3/envs/flow/lib/python3.8/site-packages/oneflow/nn/module.py", line 115, in __call__
    res = self.forward(*args, **kwargs)
  File "/home/ubuntu/miniconda3/envs/flow/lib/python3.8/site-packages/oneflow/nn/modules/padding.py", line 414, in forward
    return flow._C.pad(x, pad=self.padding, mode="constant", value=self.value)
RuntimeError: 
  File "/home/ci-user/runners/release/_work/oneflow/oneflow/oneflow/core/functional/impl/nn_functor.cpp", line 1858, in operator()
    Check failed: (ndim) <= (5) (6 vs 5) Dimension of input tensor should less than or equal to 5

最小复现

import torch
import torch.nn as tnn
import oneflow as flow
import oneflow.nn as fnn
import numpy as np

inp = np.random.randn(2, 10, 1, 64, 56, 56).astype(np.float32)

torchTensor = torch.from_numpy(inp)
tPad = tnn.ZeroPad2d(1)
tOut = tPad(torchTensor)
print(tOut.shape)

flowTensor = flow.from_numpy(inp)
fPad = fnn.ZeroPad2d(1)
fOut = fPad(flowTensor)
print(fOut.shape)

[Flowingsun007]

这个问题是因为functor内部错误地check了input的dim，这个pr修复了此问题：Oneflow-Inc/oneflow#8354