Hi, when I was trying converting a ReID (person re-identification) model into tensorrt, I encounter two bugs when using onnx-tensorrt tool (https://github.com/onnx/onnx-tensorrt), then I use gdb doing some debug, but because TensorRT is not open sourced, I can’t going down further, so I hope get some help here
The first bug is related to split operation, and it happens after this commit (https://github.com/onnx/onnx-tensorrt/commit/2066f534f66320b7ecdf3eccbaf18ff1fdba6287)，according to the commit message, it adds the support for dynamic split, but I found this commit is not compatible with the static split in our model, after this line (https://github.com/onnx/onnx-tensorrt/blob/2066f534f66320b7ecdf3eccbaf18ff1fdba6287/builtin_op_importers.cpp#L1888), the output shape turns out to be (-1, -1, -1, -1), and in the previous version before this commit, the output shape is like (32, 64, 32, 32) or something similar.
The second bug is related to Instance Normalization(I guess), it crashes at this line of code (https://github.com/onnx/onnx-tensorrt/blob/8716c9b32dcc947287f2ede9ef7d563601bb2ee0/main.cpp#L245), when I using gdb backtrace, it seems like the error happens at tensorrt trying to save something like PluginV2Param into a cache, and as far as I know, the only op in our model which needs tensorrt plugin is Instance Normalization.
Here is the model download url(https://www.dropbox.com/s/g2rxzx45vfb7q9h/resnet_50_ibn_a_op10.onnx?dl=0) for test.
And the system and cuda version as follows:
GPU: RTX 2070
System: CentOS 7