Yes, basiscally it’s something like this.
I though of multiple approaches at the time.
The first step would have been to encode the coefficients as a base64 string, instead of an array of hexadecimal constants, although the gain is maybe not very high.
Then, instead of encoding the coefficients themselves, I thought of compressing the coefficient array first with gzip, lzma, or anything else, and encode the result as the base64 string. At program load time, I would have to decode the string and decompress the data to get back the coefficients.
The problem with these two approaches, it that although they gain a bit of code size, there is also an overhead for the decoder / decompresser code.
My codingame framework code was already very big and I had very little margin to include additional code. So, I went further and instead of compressing and encoding the coefficients only, I did it for the program binary itself:
- I compile all of my framework code with NN coefficient arrays directly as an executable in local, which may already be smaller than the source code itself because of unused code, and other things.
- The resulting binary is then optimally compressed using UPX, which is a nice executable compression tool, which embeds decompression code directly in the executable, in addition to aggressive transformations to achieve minimal executable size.
- I encode the resulting binary in base64, and I include it in another small source code that will decode the base64 to a temporary file, at runtime on CodinGame servers, and then execute the resulting binary.