Inconsistency between pytorch and onnxruntime

[maybeLee]

I build an onnx graph with only one BatchNormalization layer and two transpose layer as follows:

However, the output between onnxruntime and pytorch's result is different when converting onnx to pytorch:

To reproduce: Please access the model through this link: https://drive.google.com/file/d/1KQ-ZvdghB2Fw0b1U42M1Q8DFfJzKMKuF/view?usp=sharing

import onnx
from onnx2pytorch import ConvertModel
import numpy as np
import torch
input = np.random.rand(1,3,3,3)
onnx_model = onnx.load("incon.onnx")
torch_model = ConvertModel(onnx_model, experimental=True)
torch_input = torch.from_numpy(input)
torch_model.double()
torch_pred = torch_model(torch_input)


import onnxruntime as ort
providers = [
    ('CUDAExecutionProvider', {
        'device_id': 0,
        'arena_extend_strategy': 'kNextPowerOfTwo',
        'gpu_mem_limit': 10 * 1024 * 1024 * 1024,  # 10G
        'cudnn_conv_algo_search': 'EXHAUSTIVE',
        'do_copy_in_default_stream': True,
    }),
    'CPUExecutionProvider',
]
onnx_path = "incon.onnx"
onnx_model = ort.InferenceSession(onnx_path, providers=providers)

input_name = onnx_model.get_inputs()[0].name
output_name = onnx_model.get_outputs()[0].name
input = input.astype('float32')
onnx_pred = onnx_model.run([output_name], {input_name: input})[0]
print("When running on CPU: ", onnx_pred[0])

Result of ONNXRuntime:

Result of PyTorch:

tensor([[[[ 0.8651,  0.4476,  1.0992],
          [-1.4390, -1.9661,  0.6654],
          [-0.0661,  0.8117,  0.4977]],

         [[ 1.4929, -1.0736, -1.8301],
          [-0.3819,  0.1353, -1.3534],
          [-0.5915,  1.0055,  0.1152]],

         [[ 1.4507,  0.9089,  0.7818],
          [-0.2749,  0.6187,  0.7877],
          [-1.0552, -0.8879, -0.7635]]]], dtype=torch.float64,
       grad_fn=<PermuteBackward0>)

Result of ONNXRuntime:

[[[0.6175244  0.7708563  0.99578035]
  [0.06858005 0.03348556 0.86044073]
  [0.39565903 0.8820858  0.8081321 ]]

 [[0.7670855  0.30615166 0.08200629]
  [0.32042617 0.6754674  0.2307212 ]
  [0.27048445 0.94129366 0.6888292 ]]

 [[0.75703746 0.9117755  0.89674425]
  [0.3459281  0.8231245  0.89860445]
  [0.16000679 0.3628597  0.41474345]]]

[koon-kai]

I have the same problem.