Different scores from browser and server recognition under the same model

[visionki]

The image I used is this one: https://www.pixiv.net/artworks/80908440 The model uses this: https://github.com/infinitered/nsfwjs/tree/master/example/nsfw_demo/public/model

The result returned on https://nsfwjs.com/ is:

But the result I get using the model on the server is:

{
            "className": "Drawing",
            "probability": 0.8200194835662842
        },
        {
            "className": "Hentai",
            "probability": 0.17648540437221527
        },
        {
            "className": "Porn",
            "probability": 0.002067374065518379
        },
        {
            "className": "Neutral",
            "probability": 0.0013537625782191753
        },
        {
            "className": "Sexy",
            "probability": 0.00007402057235594839
        }

The code used in the backend is the sample code in the README.md。

_model = await nsfw.load('/model/', { size: 299 })

I'm not very familiar with tensorflow,I wonder why I get different results, it seems that the browser recognition is more accurate. How can I get more accurate results in the server ?

[GantMan]

So the code from the readme needs to be update to use TSNode which will encode the images the same.

Try: https://js.tensorflow.org/api_node/3.18.0/

Let me know if that matches

[visionki]

So the code from the readme needs to be update to use TSNode which will encode the images the same.

Try: https://js.tensorflow.org/api_node/3.18.0/

Let me know if that matches

I try to change const image = await convert(req.file.buffer) to const image = await tf.node.decodeJpeg(req.file.buffer) in the sample code, the result obtained is:

        {
            "className": "Drawing",
            "probability": 0.8348554968833923
        },
        {
            "className": "Hentai",
            "probability": 0.16166625916957855
        },
        {
            "className": "Porn",
            "probability": 0.0019792686216533184
        },
        {
            "className": "Neutral",
            "probability": 0.0014262152835726738
        },
        {
            "className": "Sexy",
            "probability": 0.00007269137859111652
        }

There doesn't seem to be a big change in the results , I'm not very familiar with these, not sure if what I've modified is as you said

[erickythierry]

i had the same problem. and I noticed that the detection changes completely if the image has a slight difference in resolution. I used a 500x500 image to test and then I tested the same image but reduced to 400x400. detection dropped dramatically (from ~0.90 to ~0.60).

a very big difference for the same image with minimal size difference

[araufdogan]

Hello @GantMan,

I have the same issue. I think server-side nsfwjs not working (or tutorial not working). I have tested with;

• 3 models
• Your convert function and tf decodeImage
• tfjs-node version 3.18.0

Results are not same with nsfwjs.com. It's not about %1-2 difference, they are totally different. If you want anything, I can test it to help fix the issue.

Best regards.

[GantMan]

This seems like it must be a bug with tfjs-node - I can't think of anything specifically that is causing the issue.

Can you check on tfjs-node for differences?

[dantecsm]

In 2023, this issue still exists. Does anyone have any ideas or solutions?