[Question] Is it possible to convert "interpolation with align_corners=True" from Pytorch to TensorRt

[ycchanau]

Description

When I convert a PyTorch model to onnx and then use TensorRT to parse it, I get this error.

[8] Assertion failed: (transformationMode == "asymmetric") && "This version of TensorRT only supports asymmetric resize!"

From this issue #273, I notice that tensorRt has an IResizeLayer with align_coerners and gives the same output as Pytorch when align_corners=True. I just wonder if it is possible to convert "interpolation with align_corners=True" from Pytorch to TensorRt

Environment

TensorRT Version: 7.0
GPU Type: T4
Nvidia Driver Version:
CUDA Version: 10.2
CUDNN Version: 7.6.5
Operating System + Version: ubuntu 18.04
Python Version (if applicable): 3.6
TensorFlow Version (if applicable):
PyTorch Version (if applicable): 1.4.0
Baremetal or Container (if container which image + tag): TensorRT Release 20.02

Relevant Files

Steps To Reproduce

The code of the model

class Resize_Test(nn.Module):
    def __init__(self):
        super(Resize_Test, self).__init__()


    def forward(self, x):
        x = F.interpolate(x, size=(128, 128), mode='bilinear', align_corners=True)
        return x

[rmccorm4]

Hi @ycchanau ,

1. You could try editing the ONNX parser source like this PR, building it, and seeing if it parses your model successfully: https://github.com/onnx/onnx-tensorrt/pull/418/files

2. It seems like torch2trt might support interpolate since this directory exists: https://github.com/NVIDIA-AI-IOT/torch2trt/tree/master/torch2trt/converters/interpolate, you could try to use that.

3. It also seems like another user contributed a PR to torch2trt with support for interpolate, you could check this out as well: A whole lot of extra operations (including interpolate, 3d convs/deconvs/pooling) NVIDIA-AI-IOT/torch2trt#249

Please share your results if any of these options work for you.

[WIll-Xu35]

@rmccorm4 Thank you so much for your information. Your first solution solved my problem.

[rajeevsrao]

To be closed pending onnx/onnx-tensorrt#418