Open HikariTJU opened 2 years ago
This model can take images of different sizes as input. However, to achieve best performance, it is recommended to resize the image such that both height and width are within the range of [800, 1333], and then pad the image with zeros such that both height and width are divisible by 32.
scatter
op requiring standard shapes and the model presumably passing a broadcasted shape.
scatter
op so that it works with nonstandard shapes driver run MaskRCNN-12.onnx --input-dim @image 3 800 800
and scatter
is set to allow nonstandard shapes
compute_shape
error with certain image sizesHi, I delete the code that check shapes and it works, but fp16 still get error
terminate called after throwing an instance of 'migraphx::version_1::exception'
what(): /code/AMDMIGraphX/src/include/migraphx/op/get_tuple_elem.hpp:56: compute_shape: GET_TUPLE_ELEM: index 0 is out of range 0
[1] 185791 abort (core dumped) migraphx-driver perf --onnx MaskRCNN-12.onnx --input-dim @image 3 800 800 --fp16
The verify command that checks between the reference and GPU versions doesn't pass, so there's probably some bugs. Will have to investigate.
I'm running maskrcnn onnx model from
link
and get the error belowusing MIGraphX Version: 2.3 compiled from latest develop branch and rocm=5.2.1 on mi210