Using TensorRT file (end2end.engine) outside MMDeploySDK

[k-ishi8]

I would like to use the end2end.engine(MMPose TensorRT) obtained by MMDeploy with Triton Inference Server.
I am aware that end2end.engine only includes inference processing, so I think I need to implement the preprocessing myself, is that correct?
Also, I am outputting pipeline.json to understand the contents of the preprocessing, but I have a question about the value of Normalize.

I used this as a reference to get end2end.engine and pipeline.json.
In config, I specified topdown_heatmap/coco/hrnet_w48_coco_256x192.py.
In topdown_heatmap/coco/hrnet_w48_coco_256x192.py, type='NormalizeTensor' mean=[0.485, 0.456, 0.406],std=[0.229, 0.224, 0.225]. In pipeline.json, "type": "Normalize","mean": [123.675,116.28,103.53],"std": [58.395,57.120000000000005,57.375] is set.
Is there a difference between 'NormalizeTensor' and 'Normalize'?

[lzhangzz]

You can find the definition of NormalizeTensor here.

It seems that NormalizeTensor is working on torch.Tensor in CHW format, while Normalize is working on HWC numpy.Array

[k-ishi8]

@lzhangzz Thank you for your response.
If I use end2end.engine, should I convert with HWC or CHW?
Since what is listed in Pipeline.json is Normalize, should I convert it to HWC?

[lvhan028]

As far as I know, pipeline.json usually includes Normalize followed by ImageToTensor.
Normalize request the input being in HWC format.
After Normalize, Image2Tensor will transpose the tensor from HWC to CHW.

[k-ishi8]

Thank you for your response.
I was able to get the preprocessed data using Compose and test_pipeline.