Summary of Shaolin School Questions:
Q: 
A: The bmodel is missing, the path of the bmodel is not specified, and the test picture is not in the same file path. Copy it and it will be solved.
Q: Running ./install_lib.sh nntc under the docker container will fail to rm the file, what's the matter?
A: The file has been deleted
Q: After I converted the pytorch model into a .torchscript.pt file with the export tool, this error still occurs. The versions of pytorch are all the same, and the official model can be run without any problem.
A: The version is 1.8 when pytorch trace is required
Q:
A: No configuration environment, execute source envsetup_cmodel.sh
Q: There is no sophon under opt/
A: SDK3..0 and below are in /system, and above are in /opt
Q: In addition to modifying the weight file after transplanting the yolov5 model trained by myself, is the coco class class directly modified into the category in the own model? And how to modify the script when transplanting the own model?
A: You need to find the coco.name or ulits file and modify it into your own class
Q: I would like to ask if the model of yolov5s_640_coco_v6.1_3output_fp32_1b.bmodel has been tested for performance, why I only run 5 frames and 1 second with local video
A: The reasoning speed is too slow, you need to train your own data set, and you can speed it up by adjusting the batch and other methods in the later stage
Q: At present, the recognition accuracy of the pt file trained by the customer is quite high, but the board can hardly be recognized when it is converted into the bmodel file, and only people and chairs can be recognized. The recognition accuracy is very low. What is the problem?
A: Need to do some precision tuning
Q: When onnx is converted to bmodel, this operator does not support it. When training the model, use other operators to implement this operator equivalently. Will the speed be much slower?
A: How much slower the specific speed is, of course, depends on how you use other operators to implement equivalently, but if you can solve unsupported operators by yourself, it is already very good
Q: What is the reason for this error when compiling sail? The built-in version 2.7 of sail does not have the attribute 'sail_resize_type', do you want to reinstall it?
A: New features are not supported on 2.7, you must upgrade to the latest version
Q: Under the image of docker, use this sdk, and then run docker and say that there is no such command
A: The Workspace description is already in the docker container
Q: Will it take up a lot of CPU resources when using bmnetu to convert bmodel? When using this string of codes to convert the bmodel of the int8 model, the CPU resources of the computer will run to 100% and the process will be KILLED
A: There are certain requirements for the performance of the equipment in the development environment. If you are killed, you can try to restart or close other software. It may also be that the memory is not enough.
Q: May I ask if the above cmake compilation problem is ok? Does it mean that libsophon_0.4.1_aarch64 cannot be found in the build, but the build is an empty folder manually created in the tutorial
LIBSOPHON_BASIC_PATH : In the cross-compilation mode, the path of libsophon, if the configuration is incorrect, the compilation will fail. This compilation option does not take effect under pcie mode.
Q: The command I entered is like this, the input shapes of the model are obtained from the figure, so should my shape be defined as [1,2048,1]
A: The input shape has four dimensions, batchsize, channel, and the length and width of the image
Q: I re-entered the shapes, and he will report the following error. Is the step size of the convolution kernel movement must be smaller than 16?
A: The step size needs to be less than 16
Q: After setting the environment variables, how to solve the error that there is no GLIBC 2.29 in system/bin when installing the sail inference library? Do I need to update GLIBC by myself?
A: The version does not match, the development version is inconsistent with the running version, modify it to the correct one