Q：
When converting torch to bmodel, the bmnetp tool reports an error
A：
Some models involve dynamic shapes, which need to be set to dynamically compile dyn=true
Q：
A：
The memory is full, the process is killed by the system, and the running memory is recommended to be 12g or more
Q：
ssh connection to the development board, a Connection refused problem occurs
A：
First refer to the Shaolin school connection document to check whether the cable is connected correctly, whether the ip setting on the development machine is correct, and whether the ip address of the ssh connection is correct. If everything is correct according to the documentation and the Shaolin faction has been flashed by itself, it may be because the mirrored ssh service of the flashing machine has not been installed normally, and you need to flash the machine again.
Q：
After the model is converted to bmodel, the accuracy of the model decreases
A：
First determine whether the code logic of the post-processing call part is correct; if it is confirmed to be correct, it may be a problem of misalignment of precision. Some models involve image processing tasks, and the processing accuracy of the function calling library needs to be consistent with the data accuracy of the conversion model. If they are different, there will be a loss of accuracy when converting between them.
Q：
A：
An error was reported that the shared dependency library was not found.
The dependency library was found in the file system. export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:<lib_path> to import the dependency library.
Q：
Can I compile a model whose transformation tensor is a dynamic shape?
A：
Yes, when you need to use the bmnetp tool, specify the parameter [--dyn] as true, as long as the actual running shape is smaller than the specified parameter [--shapes].