Embed OpenCV SVM as preinitialized static variable - C++

TopAnswers C++

Top Answer

Somebody

Unfortunately, OpenCV does not have a setter for the decision function and the support vectors, which does not allow me to inject my own loader.

After browsing the code, I decided to rip out the specific SVM type implementation I am using. This allowed me to write  a Python script that generates C++ float arrays from the yaml files and operate them directly in my code.

This cut down the load time from  around one second on my phone to basically nothing since it is just pointer passing with my code. Interestingly, the optimizer also improved the prediction runtime of my code over OpenCV's by a factor of 2.

I did not test the slower devices but so far they were about 20x slower, so I'd expect to have cut down the load time from ~20s to <1s.

Particularly if the data is static you can even tune the compression. But yeah, we can't get around standard I/O time. You may be able to hide it in some ways... booting the UI, while loading the data on another thread, stuff like that. But it does kinda scatter latches everywhere in your code checking to see if the backing data has been loaded. Compression may be your best bet, you can still embed that in a source file and not have to go to opening a file on the disk, just feed from a string directly into your decompression library of choice.

Is the data static? You can bake static data into C++ files, which will result in larger executables, but only slightly slower load time. I'm assuming that the data is not static though, in which case no, you must stream in that huge file at load time. Depending on the type of data and the repetition therein you may have some savings by compressing it. (You can determine this by zipping your 100mb .yml file. Make sure there's a significant size difference before attempting to add in compression/decompression.) But there is no "preloading" non-static files into memory. Your limiting factor is your I/O bus not your loading or streaming method.

OpenCV is perfectly fine with loading the yaml at runtime. That is not the issue. My problem stems from the fact that mobile systems incur a huge time penalty in loading/parsing a ~100mb text file for startup. This could be solved by having the data embedded in a format that does not need to be parsed.

If you're asking if you can read a .yml file with C++20 the answer is no. You'd need to use Boost or similar to read the .yml. Generating code always comes with a building cost, as in you'll need to do some build engineering to make sure that the code in place is up to date. If you're writing the build system yourself, and you don't have a dedicated engineer to ensure that the build system is up to date, I'd suggest that you do the .yml reading in C++ and do away with the Python all together. yaml-cpp or this boost library may be good options for you: https://github.com/boost-vault/serialization

Thanks for the pointer but that was not what I was getting at. I currently have the SVM stored as a YAML file. Now I need to somehow transform that into float arrays. I think it will be easy enough to create a code generator with Python, but my question was more along the lines: Can I do this with "vanilla" C++ (20) without an extra code gen step. But I suppose the code generator path would be easiest.

I think the main problem to me is, how to embed the float data such that I can consume them with opencv with as little overhead as possible. I'd assume it should be possible to store the vectors/weights as float arrays that should be trivially loadable by OpenCV but this opens up the issue of embedding data at compiletime with the correct endianess.

@Somebody, re: [your question](#question), I'm afraid that question is far to broad. In general the right response is slurping from a parsed file and embedding directly into the OpenCV implementation. If you'd like to provide some specifics or focus on a question that may be helpful. For example: 1. How do I slurp a file? 2. How do I determine platform endianness? 3. How can I convert hex read from a file into the proper float? 4. Does OpenCV provide a method for external parsing of SVMs? (You'd have better luck looking at OpenCV's documentation than asking that last one here.)

1 Answer