big difference in results using saved_model.tflite when compared to nsfwjs.com

[sandeep5193]

I have used saved_model.tflite from v1.1.0 and v1.2.0. image used is 224x224, same jpeg in both testing. here are results:

nsfwjs.com
Neutral - 67.35%
Porn - 31.40%
Drawing - 0.79%
Hentai - 0.26%
Sexy - 0.19%

saved_model.tflite (v1.1.0, on android with tensorflow lite)
porn: 0.8919
sexy: 0.0455
hentai: 0.0355
drawings: 0.0209
neutral: 0.0062

jpeg used is here(safe-for-work):

I am new to this, so correct me if am wrong, I have simply used this tflite without any changes. Am I using the correct version?

[sandeep5193]

while using tflite in android, input RGB is normalised to [0,1] from [0,255].

[sandeep5193]

Tried predict.classify() JavaScript function of this project, with v1.1.0's saved_model.h5. It gave exactly same results as with my android code(using saved_model.tflite).

Wondering if model used in nsfwjs.com is same as v1.1.0 of this repo or something else. @GantMan can you please clarify?

[KevinColemanInc]

The v1.1.0 is much worse than the nswfjs. No idea why.

[ThatHackerDudeFromCyberspace]

The v1.1.0 is much worse than the nswfjs. No idea why.

NSFWJS seems to have higher accuracy on all models, not just v1.1.0 nvm, it seems that nsfwjs is only using v1.0 models

[KevinColemanInc]

If anyone find's this thread in the future, I ended up using Bumble's Private Detector for my NSFW-FLASK service.

Bumble's has higher precision and accuracy. Bumble's model is about 400mb compressed while this is 135Mb

[ThatHackerDudeFromCyberspace]

Wow, thanks!