grimoire/mmdetection-to-tensorrt

test.py

Closed this issue · 4 comments

Yuanye-F commented

Thanks for your helpful work.

I have a problem when I tested trt_model in a dateset.

I found that using tool/test.py to test the COCO2017 dataset causes problems, and I think it's because of an error related to mmdection 'single_gpu_test' and the result from 'ModelWarper'.

The result of the error is.
400000/5000, 1053.8 task/s, elapsed: 380s, ETA: -374s

It should be stopped at 5000 not 400000. Could you please help me with this problem?

The trt_model I am using is transferred from mmdetection 'faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth'.

grimoire commented

Hi
Thanks for your bug report. This test module is created before mmdetection 2.3.0, They must have change something after that.
Change the return value of ModelWarp.forward() to return convert_to_mmdet_result(result, self.num_classes) (remove [0]) should give you the right result.

Yuanye-F commented

Thanks for your reply.

This might work, but there's an error:

[TensorRT] ERROR: Parameter check failed at: engine.cpp::setBindingDimensions::1046, condition: profileMinDims.d[i] <= dimensions.d[i]

grimoire commented

That's caused by small opt_shape_param range.
some images might have large w/h ratio after preprocess. set opt_shape_param like below should fix this:

    opt_shape_param=[
        [
            [1,3,224,224],    # min tensor shape
            [1,3,800,1312],  # shape used to do int8 calib
            [1,3,1344,1344], # max tensor shape
        ]
    ]

Worth to mention that, large opt_shape_param need more memory.

Yuanye-F commented

Thanks, the problem is solved.