NHWC to NCHW input NCHW but output NHWC?

[AromaticJ]

I used command:
python -m tf2onnx.convert --graphdef ./model.pb --output ./model.onnx --inputs input_tensor:0 --inputs-as-nchw input_tensor:0 --outputs output_1:0,output_2:0,output_3:0,output_4:0,output_5:0
converted an onnx file successfully, and used it to generate an tensorrt engine on Nvidia Orin device with tensorrt8 by trtexec tools,
however, trtexec logs:

[10/02/2019-23:59:47] [I] Created input binding for input_tensor:0 with dimensions 1x3x512x512
[10/02/2019-23:59:47] [I] Using random values for output output_0:0
[10/02/2019-23:59:47] [I] Created output binding for output_0:0 with dimensions 1x512x512x3
[10/02/2019-23:59:47] [I] Using random values for output output_1:0
[10/02/2019-23:59:47] [I] Created output binding for output_1:0 with dimensions 1x512x512x3
[10/02/2019-23:59:47] [I] Using random values for output output_2:0
[10/02/2019-23:59:47] [I] Created output binding for output_2:0 with dimensions 1x512x512x3
[10/02/2019-23:59:47] [I] Using random values for output output_3:0
[10/02/2019-23:59:47] [I] Created output binding for output_3:0 with dimensions 1x512x512x3
[10/02/2019-23:59:47] [I] Using random values for output output_4:0
[10/02/2019-23:59:47] [I] Created output binding for output_4:0 with dimensions 1x512x512x3
[10/02/2019-23:59:47] [I] Starting inference

you can see, my input is already format nchw, but outputs are still format nhwc.
How to convert my outputs to format nchw as same as input?

System information

• OS Ubuntu 18.04:
• Tensorflow 1.14:
• Python version 3.6:
• tensorrt 8.2.1:

[fatcat-z]

Working on it.

[hwangdeyu]

Same feature request with #1973

[hwangdeyu]

Hi @AromaticJ The --outputs_as_nchw option is supported it in latest main branch. Please tell me if it's work for your model demand Thanks!