[Feature request] A more light-weighted model

[fkcptlst]

Maybe add some optional light-weighted models to choose from? After all, Inception is a bit too heavy for small servers. Its memory consumption is too large.

[arnidan]

Hello @fkcptlst, thank you for the great idea!

Could you check docker pull ghcr.io/arnidan/nsfw-api:50_merge-min image on your environment?
It's bundled with quantized model

[arnidan]

I made some tests.
Seems like there is no big difference in RAM usage between the models. It's saving around 50MB, 600MB (default) vs 650MB (min)

[fkcptlst]

Yes, my test results are the same. It turns out that the Inception sometimes even consumes less RAM than MobileNet, which is a bit peculiar.

Thanks for the effort!