TensorRT 7 Resize produce inconsistent result with PyTorch Interpolate

[xpngzhng]

I have an HRNet semantic segmentation model trained with Pytorch, using interpolate bilinear align_corners=True
Then I convert it for C++ inference with align_corners=True and align_corners=False, I find that the output of the two mode are almost the same.

Using PyTorch1.5, onnx build from source (almost the latest version, showed 1.7.0), set interpolate with align_corners=False, export onnx model with opset version 10.

Then I use TensorRT 7.0.0.11 for inference in Windows. I find the output is quite different from those of PyTorch, the main problem is the result is almost a shifted and smaller version than the PyTorch version.

Is writing a plugin the only way to solve this problem?

Here are the result
torch align

torch no align

torch diff

trt no align

trt torch diff

[OtisZZ]

I met the same problem.
I am using trt7.1 and inference a model, the result is some pixels movement difference from onnx model inference. And the onnx inference is good to align with the tensorflow inference results.
Is there any settings to make upsample functions the same?

[sixcheng]

I have the same problem.
Have you solved it?

[donrax]

The issue is connected to the interpolate (resize) layer in Pytorch. Also, take a look at #361 (comment)

You can try and solve it by using the following:
x = F.interpolate(x, size=(H, W), mode='bilinear', align_corners=True)

Pytorch 1.5.1
Tensorrt 7.1.3 (nvidia doc)
Opset 11

[kevinch-nv]

@xpngzhng thanks for the report. Can you clarify a few things to help us narrow down the issue?

• Is this reproducible on the latest TRT version? We've made some updates on what resize modes of ONNX we actually support.
• Is this reproducible on ONNX opset >= 11? ONNX has increased their range of resize modes since opset 11, and that may the cause of the discrepancy seen between Pytorch and ONNX.

[kevinch-nv]

Closing this issue due to inactivity - please open a new issue if the problem persists.