Model accuracy after conversion to wasm

[chuaziheng]

Hi, wanted to ask if you have experienced a big drop in model accuracy when running inference in browser? The performance dropped quite significantly as compared to running the same NCNN model in CPU. However, I saw your demo and the accuracy still looks good, do you have any tips regarding the model conversion to preserve the accuracy? Thank you :)

[enderminyard]

@chuaziheng how did you convert the model to WebAssembly?

I have tried to compile the ncnn repo for the last two days, first on my computer, then in Docker.

I tried these tutorials:

waittim.github.io/2020/11/15/build-ncnn-wasm/
github.com/nihui/ncnn-webassembly-nanodet
github.com/Tencent/ncnn/wiki/how-to-build#build-for-webassembly

Through trial and error, I produced this Dockerfile for my compilation process: https://pastebin.com/ASN5p8CA

I was able to compile…something? The terminal would show compilation progress until some 100% score was reached but no JS or WebAssembly files were generated in this process.