ctdet detect boxes is wrong

[facerless]

I trained dla_512 on CenterNet. pytorch version: 1.0.0

After i convert model to tensorrt engine, and i test in a image. Predict for class is right, but predict boxes is all wrong. What happend? Can u help me! Thanks you.

detect box x1,y1,x2,y2: person 194.842 448.538 196.785 359 person 256.667 438.983 258.607 359 person 362.734 443.965 364.624 359

Image height is 360,width is 480. @CaoWGG

[CaoWGG]

@facerless check your code about converting onnx heads = OrderedDict([('hm', 80), ('reg', 2), ('wh', 2)])

[facerless]

Thanks for your help! I set heads={'hm': 80, 'wh': 2, 'reg': 2}，lead to this bug! After i set heads = OrderedDict([('hm', 80), ('reg', 2), ('wh', 2)]), the detection result is right!