EfficientNMS plugin doesn't support fp16

[lucasjinreal]

I have an onnx model, injected EfficientNMS plugin using graph-surgon, the inference result like this in fp16:

(array([38], dtype=int32), array([[448.      , 310.75    , 557.5     , 413.25    ],
       [324.25    , 308.      , 346.25    , 326.5     ],
       [193.25    , 302.5     , 225.      , 327.5     ],
       [371.5     , 288.25    , 381.      , 301.75    ],
       [448.      , 310.      , 559.5     , 414.      ],
       [329.75    , 308.75    , 350.75    , 328.25    ],
       [397.75    , 306.      , 454.25    , 363.      ],
       [368.      , 311.      , 387.      , 342.5     ],
       [422.25    , 321.      , 464.75    , 378.5     ],
       [379.5     , 309.5     , 410.      , 351.5     ],
       [541.      , 305.25    , 640.      , 478.25    ],
       [306.5     , 267.25    , 310.      , 276.25    ],
       [225.25    , 305.5     , 247.75    , 324.5     ],
       [348.      , 309.25    , 369.      , 334.25    ],
       [258.25    , 306.      , 274.25    , 324.      ],
       [372.5     , 310.25    , 403.5     , 345.75    ],
       [194.875   , 303.25    , 224.625   , 326.25    ],
       [255.375   , 312.25    , 274.25    , 327.25    ],
       [256.5     , 309.75    , 274.5     , 326.75    ],
       [360.5     , 311.25    , 383.      , 340.25    ],
       [399.5     , 305.5     , 454.5     , 363.5     ],
       [541.      , 303.25    , 640.      , 482.75    ],
       [421.      , 320.5     , 466.5     , 377.      ],
       [446.75    , 309.5     , 558.5     , 412.5     ],
       [354.75    , 310.25    , 371.25    , 334.25    ],
       [321.      , 307.5     , 334.5     , 323.      ],
       [330.5     , 308.5     , 347.5     , 327.5     ],
       [255.5     , 307.      , 274.5     , 326.5     ],
       [378.75    , 314.25    , 406.25    , 350.25    ],
       [ 12.296875, 299.5     ,  22.390625, 324.      ],
       [194.125   , 303.25    , 223.875   , 326.25    ],
       [275.75    , 261.25    , 283.25    , 275.75    ],
       [354.75    , 311.      , 372.25    , 336.5     ],
       [343.75    , 310.      , 360.75    , 330.5     ],
       [346.5     , 310.      , 369.      , 334.      ],
       [364.75    , 310.      , 394.75    , 344.      ],
       [447.75    , 310.75    , 557.5     , 413.25    ],
       [371.75    , 310.25    , 404.25    , 345.75    ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ],
       [  0.      ,   0.      ,   0.      ,   0.      ]], dtype=float32), array([-0.7505, -0.505 , -0.7524, -0.754 , -0.511 , -0.511 , -0.5156,
       -0.5215, -0.5225, -0.525 , -0.5254, -0.765 , -0.7656, -0.538 ,
       -0.7695, -0.7715, -0.546 , -0.786 , -0.574 , -0.5938, -0.1953,
       -0.21  , -0.2183, -0.2261, -0.619 , -0.6323, -0.2988, -0.3066,
       -0.3384, -0.6694, -0.3599, -0.6963, -0.4214, -0.7227, -0.4497,
       -0.742 , -0.746 , -0.746 ,  0.    ,  0.    ,  0.    ,  0.    ,
        0.    ,  0.    ,  0.    ,  0.    ,  0.    ,  0.    ,  0.    ,
        0.    ,  0.    ,  0.    ,  0.    ,  0.    ,  0.    ,  0.    ,
        0.    ,  0.    ,  0.    ,  0.    ,  0.    ,  0.    ,  0.    ,
        0.    ,  0.    ,  0.    ,  0.    ,  0.    ,  0.    ,  0.    ,
        0.    ,  0.    ,  0.    ,  0.    ,  0.    ,  0.    ,  0.    ,
        0.    ,  0.    ,  0.    ,  0.    ,  0.    ,  0.    ,  0.    ,
        0.    ,  0.    ,  0.    ,  0.    ,  0.    ,  0.    ,  0.    ,
        0.    ,  0.    ,  0.    ,  0.    ,  0.    ,  0.    ,  0.    ,
        0.    ,  0.    ], dtype=float16), array([ 7,  2,  2,  9,  2,  2,  2,  2,  2,  2,  2,  9,  2,  2,  2,  2,  2,
        2,  2,  2,  2,  2,  2,  2,  2,  2,  2,  2,  2, 12,  2,  9,  2,  2,
        2,  2,  7,  2,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,
        0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,
        0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,
        0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0],
      dtype=int32))

the boxes and labels are normal, but scores are wrong, it was negative values.

using fp32 without this problem.

My env:

• TensorRT 8.2
• CUDA11.5
• 2060

[wraveane]

This seems to be related to an optimization that's in place for sorting in fp16 where it wrongly assumes all scores are within [0-1] range. If your scores are outside of this range, i.e. negative values, it can give wrong results. There will be a fix for this soon.

[lucasjinreal]

@wraveane Does it means must wait until next tensorrt version release? that would be a long time.

[wraveane]

@jinfagang An update to the EfficientNMS plugin has been posted yesterday which, among other things, fixes this FP16 problem on non-sigmoid activated input scores.
Yes, the fix should be available on the next TensorRT release, but until then, you can also build this repository from source now, and replace the libnvinfer_plugin.so.8.x.x library installed by TensorRT with the one built from this repository.
Please let me know if that resolves the issue you were having.

[lucasjinreal]

@wraveane thanks.

[liuchangf]

you can just +1 for each score in older version of tensorRT , it will very close to original detection score