Some models don't work in C++.

[c-m-k-miyoshi]

I'm trying to run DirectML via onnxruntime, but some models don't work in C++ sample.

• My exection environment
  • In C++
    • OnnxRuntime DirectML : version==1.15.1
    • DirectML : version==1.12.0
    • Graph optimization level : ORT_DISABLE_ALL
    • MemPattern : Disable
    • Execution mode : ORT_SEQUENTIAL
  • In Python
    • Python version==3.9.18
    • onnxruntime-directml==1.15.1

One of the models that does not work is the one below.

Code for model creation

import os
import torch
import torch.nn as nn

model_name_onnx = 'model_avg_512.onnx'

class simple_model(nn.Module):
    def __init__(self):
        super().__init__()
        self.avg_pool_01 = nn.AvgPool2d((2,2), stride=(2,2), padding=0)

    def forward(self, img):
        y0 = self.avg_pool_01(img)

        return y0

model = simple_model()

input0 = torch.tensor(torch.randint(255, size=(1, 1, 512, 512)), dtype=torch.float32)
y = model(input0)

input_names = ['input0']
output_names = ['output']
dynamic_axes = {'input0': {0: 'batch_size'}, 'output': {0: 'batch_size'}}

torch.onnx.export(
        model,
        (input0),
        model_name_onnx,
        export_params=True,
        opset_version=11,
        input_names=input_names,
        output_names=output_names,
    )

• Input size: [1x1x512x512]
• Layer information: single average_pool2d layer only
  • kernel_size: (2,2)
  • strides: (2,2)
  • padding: (0,0)
• opset_version==11

The above model works fine with the Python sample, but it does not work with the C++ sample.
Also, the model with the input size changed to [1x1x256x256] works correctly with both Python and C++ samples.

Is this a known issue?
My goal is to make the model work with input size [1,1,1080,1920].

I'd be grateful for any help, thanks!

[JohnnyWang19921007]

@jstoecker
Good day!
My colleague @c-m-k-miyoshi faced this issue, if it's possible, may I check that there is any workaround or fix?
This issue will delay our customer Acer to release their latest AI PC, according to the schedule, this feature should be provided in the mid of Jan 2024.

[JohnnyWang19921007]

@jstoecker
May I check there is any update?