How to make sense of trtexec debug infor?

I am having a very hard time converting tf models to trt via. How to know where (which operator) the problem comes from when reading trtexec --verbose?
For example, this excerpt from debug infor:

[03/29/2021-06:53:12] [V] [TRT] Registering layer: strided_slice_1 for ONNX node: strided_slice_1
[03/29/2021-06:53:12] [V] [TRT] Registering tensor: strided_slice_1:0 for ONNX tensor: strided_slice_1:0
[03/29/2021-06:53:12] [V] [TRT] strided_slice_1 [Slice] outputs: [strided_slice_1:0 → (1, 512, 512, 3)],
[03/29/2021-06:53:12] [V] [TRT] Parsing node: strided_slice_1__9 [Squeeze]
[03/29/2021-06:53:12] [V] [TRT] Searching for input: strided_slice_1:0
[03/29/2021-06:53:12] [V] [TRT] Searching for input: const_starts__1673
[03/29/2021-06:53:12] [V] [TRT] strided_slice_1__9 [Squeeze] inputs: [strided_slice_1:0 → (1, 512, 512, 3)], [const_starts__1673 → (1)],
terminate called after throwing an instance of ‘std::out_of_range’
what(): Attribute not found: axes

Does this mean that strided_slice_1__9 [Squeeze] inputs: [strided_slice_1:0 → (1, 512, 512, 3)], [const_starts__1673 → (1)] has problem? How to check value of strided_slice_1__9 [Squeeze] inputs if it is the case?

I have this problem when converting efficientdet-d0 to trt engine via onnx using op13.


what(): Attribute not found: axes

The error message indicates that you should have an axes attribute.


Thank you for quick response. So you mean strided_slice_1__9 should have attribute axes?
In general, for the debug infor returned like above, the last [03/29/2021-06:53:12] [V] [TRT] contains information about the operator having problem, is that right?


Based on the log, TensorRT complains about there is no axes attribute in the Squeeze layer.
So it doesn’t know how to convert strided_slice_1 to the next tensor.

To debug the ONNX model, you can update the model to this website: