How to clarify unexpected "Labels" field of results from Tensorrt based Dino (exported from mmDet-5scale-SwinL) , thanks.

[kidd6566]

Thanks for all your great support.

We've successfully trained model with custom dataset based on Dino (mmDet powered dino-5scale_swin-l_8xb2-12e_coco), and we're trying to export this model to ONNX/Tensorrt with the aid from mmDeply so far. When we try to inference images by processing the output from the trt-model(.engine), for static-640x640 configuration, we have expected formatted of outputs as below,

  ["dets" 1x300x5] : bbox and score  (result as expected)
  ["labels" 1x300] : classes  (suppose to be integer-like formatted)

But when we're checking the "labels" fields directly output from TRT-model, it seems that they would be values in float-type.
May we have your advices for how we could figure it out at this time, and thanks for your great help.

What we have from "labels" field so far :

    Ex: array([0.0e+00, 1.1e-44, 0.0e+00, 1.4e-45, 4.2e-45, 0.0e+00, 2.4e-44,
           ...)

Outputs sample exported by TRTEXEC :

    { "name" : "dets"
      , "dimensions" : "1x300x5"
      , "values" : [ 634.75, 360.563, 640, 514.705, 0.0122363, 634.75,...]}
      
    , { "name" : "labels"
      , "dimensions" : "1x300"
      , "values" : [ 3, 2...]}
    }

Thanks.

[kidd6566]

Sorry to bother again. It seems that there should be more post-processing we have to handle here before further visualizing results.
Close this discussion with more clues at this time , thanks