[Bug] test_tensorrt.py::test_conv2d_transpose failed

[lp6m]

There is a bug in the code generation for TensorRT that causes the conv2d_transpose test to fail.

run_pytest cython a tests/python/contrib/test_tensorrt.py::test_conv2d_transpose
enabled targets: cuda; cuda -model=unknown -libs=cudnn
pytest marker: gpu
================================================================================================ test session starts ================================================================================================
platform linux -- Python 3.6.9, pytest-6.2.5, py-1.10.0, pluggy-1.0.0
rootdir: /workspace
plugins: forked-1.3.0, xdist-2.3.0, profiling-1.7.0
collected 2 items                                                                                                                                                                                                   

tests/python/contrib/test_tensorrt.py .F 
...
---------------------------------------------------------------------------------------------- Captured stderr call ------------------------------------------------------------------------------------------------
[02:42:32] /workspace/src/runtime/contrib/tensorrt/tensorrt_runtime.cc:300: Finished building TensorRT engine for subgraph tvmgen_default_tensorrt_main_0 with batch size 1
[02:42:36] /workspace/src/runtime/contrib/tensorrt/tensorrt_runtime.cc:300: Finished building TensorRT engine for subgraph tvmgen_default_tensorrt_main_0 with batch size 1
------------------------------------------------------------------------- generated xml file: /workspace/build/pytest-results/a-cython.xml --------------------------------------------------------------------------
============================================================================================== short test summary info ==============================================================================================
FAILED tests/python/contrib/test_tensorrt.py::test_conv2d_transpose[run] - AssertionError: 
=========================================================================================== 1 failed, 1 passed in 32.66s ============================================================================================

This problem has not been detected by CI because CI only runs the compile test and not the run test.

If I set the number of output channels in the test to 1 as shown below, the test passes, so I think the cause is that the weights are not converted correctly during the layout conversion, but I have not yet identified where the bug is.

 def test_conv2d_transpose(run_module):
     def get_graph(
         x_shape=(1, 32, 8, 8),
-        k_shape=(32, 16, 3, 3),
+        k_shape=(32, 1, 3, 3),

Also, the test passes fine with commit ID 92ca782, so it is likely that subsequent commits are causing this problem.

Environment

TVM: Release v0.8
TensorRT: 7.2.3

[masahi]

There were several breaking changes in the conv2d transpose implementation.

• [Topi][Op][PyTorch][Vitas] Fix inconsistent kernel layout conventions for conv2d_transpose #9336
• [Conv2DTransposed] Fix wrong shape check and add new TOPI module to support groups #9465

Probably the necessary change didn't get applied to the TensorRT backend. @AndrewZhaoLuo @Laurawly

[AndrewZhaoLuo]

Yeah this is probably my thing. Some weird thing going with layout transforms yep.

If you change "nn.conv2d_transpose": ["NCHW", "default"], to "nn.conv2d_transpose": ["NCHW", "IOHW"] in the python/tvm/relay/op/contrib/tensorrt.py it works.

So the default should be IOHW but need to look closer

[lp6m]

@AndrewZhaoLuo Thank you for your quick reply, I confirmed this change fix this problem.
d0bb4bd

[AndrewZhaoLuo]

Above PR is ready for review, I just forgot to change the default layout in old PR I think.