Resize with nearest mode have inconsistent results compared to PyTorch and TVM

[lazycal]

Describe the bug
The result is one pixel off:

ORT=   [... 12. 12. 12. 13. ...]
Torch= [... 12. 12. 13. 13. ...]

when resizing a tensor a with shape [1, 1, 1, 1, 26] and values a[...,i]=i to [1, 1, 1, 1, 64] with nearest mode.

Urgency
None

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Ubuntu 18.04
• ONNX Runtime installed from (source or binary): binary
• ONNX Runtime version: 1.11.1
• Python version: 3.7.11
• Visual Studio version (if applicable):
• GCC/Compiler version (if compiling from source):
• CUDA/cuDNN version:
• GPU model and memory:

To Reproduce

import torch
from onnx import checker
import onnxruntime as ort
import numpy as np
ash = [1, 1, 1, 1, 26]
bsh = [1, 1, 1, 1, 64]


class Model(torch.nn.Module):
    def __init__(self) -> None:
        super().__init__()

    @torch.no_grad()
    def forward(self, a):
        b = torch.nn.functional.interpolate(
            a, size=bsh[2:], mode='nearest')
        return b


anp = np.reshape(np.arange(26), ash).astype(np.float32)
model = Model()
model.eval()
torch.onnx.export(model, (torch.from_numpy(anp),), "output.onnx",
                  input_names=["i0"], output_names=["o0"], opset_version=14)
checker.check_model("output.onnx", full_check=True)
print('model checked')
b_tch = model(torch.from_numpy(anp)).numpy()

sess_opt = ort.SessionOptions()
sess_opt.graph_optimization_level = ort.GraphOptimizationLevel.ORT_DISABLE_ALL
sess = ort.InferenceSession("output.onnx", sess_opt, providers=[
                            "CPUExecutionProvider"])
b_ort_gpu = sess.run(["o0"], {"i0": anp})[0]


from tvm import relay
import onnx
mod, param = relay.frontend.from_onnx(onnx.load("output.onnx"))
b_tvm = relay.create_executor(mod=mod).evaluate()(
    **{"i0": anp}).numpy()
np.testing.assert_allclose(
    b_tch, b_tvm, err_msg="torch vs tvm", atol=1e-2, rtol=1e-2)
print('----------------> torch is consistent with tvm')
print('ORT=  ', b_ort_gpu[0, 0, 0, 0, 30:34])
print('Torch=', b_tch[0, 0, 0, 0, 30:34])
torch.testing.assert_allclose(
    b_ort_gpu, b_tvm, atol=1e-2, rtol=1e-2)
print('pass')

The model file: model.onnx.zip

Expected behavior
They should be consistent either TVM and PyTorch are both wrong or ORT is wrong.

Screenshots

Additional context
Initiall I thought related to #12091, but now it looks like a separate issue as the inconsistent value is not constant, but just the value from the pixel right next.

[ytaous]

@hariharans29 @yihonglyu - any comment/feedback is more than welcome.

[yihonglyu]

@lazycal Is this issue on GPU? If so, could you provide system information?

[lazycal]

Hi @yihonglyu, this is on CPU:

sess = ort.InferenceSession("output.onnx", sess_opt, providers=[
                            "CPUExecutionProvider"])

Sorry for my bad variable naming.

[AsiaCao]

would appreciate any update on this issue.
i also encountered a similar issue.