Is it possible to run NSFW.js on Cloudflare?

[ambrt]

I'm seeking a cloud host for NSFW.js and wonder if Cloudflare Workers can do it.

Can anyone share some experience?

[L4Ph]

I'm not sure how long my motivation will last, but I'm creating something like this now.
https://github.com/L4Ph/nsfwts

Furthermore, in the project I'm contributing to, I've implemented it as follows and confirmed it's working on Remix (Cloudflare Pages).
https://github.com/aipictors/aipictors/blob/fd1adf889a8ea8ffa820f9e9bdde166f3ae43d60/app/routes/(%24lang).debug.nsfw/route.tsx
aipictors/aipictors#560

While the performance might depend on the CPU and memory available, I believe there's a high possibility that it can run on Cloudflare Workers since it's WASM (WebAssembly).

[ambrt]

I might try and see.

I wonder if 128MB ram of worker would be enough.

[ambrt]

There is a project using TensorFlow -UpscalerJS- that has CF worker example.
If that works then its possible that NSFW.js will too.

EDIT:
Tensor in UpscalerJS is made in browser and then sent to CF worker. So it seems that without node environment it won't work.

[ambrt]

I managed to run it on Cloudflare with five caveats:

• models except one that you use have to be removed manually from node modules
• you have to have paid Cloudflare account for this version
• you have to monkey patch 'dist/esm/index.js` to rename all models imports to one that you use only
• it takes ~25 seconds on a live deploy to classify one image
• image has to be smaller otherwise you'll end up with worker OOM error

Code for the app:

import { Hono } from 'hono';
import * as tf from '@tensorflow/tfjs';
import * as nsfwjs from "nsfwjs";
import * as jpeg from "jpeg-js";
const app = new Hono<{ Bindings: CloudflareBindings }>()

const convert = async (img) => {
  // Decoded image in UInt8 Byte array
console.log(jpeg)
  const image = await jpeg.decode(img, { useTArray: true });

  const numChannels = 3;
  const numPixels = image.width * image.height;
  const values = new Int32Array(numPixels * numChannels);

  for (let i = 0; i < numPixels; i++)
    for (let c = 0; c < numChannels; ++c)
      values[i * numChannels + c] = image.data[i * 4 + c];

  return tf.tensor3d(values, [image.height, image.width, numChannels], "int32");
};
await tf.setBackend("cpu");

app.get('/', async (c) => {
let img = await fetch("https://picsum.photos/id/237/200/300")
img =await  img.arrayBuffer();


let model=await nsfwjs.load("MobileNetV2Mid", {type:"graph"});
const image = await convert(img);
const predictions = await model.classify(image);
image.dispose();
return   c.json(predictions);

console.log("ok");
})

export default app

Renaming example is:

(...)
           case 2:
                weightBundles = void 0;
                if (!(modelName === "MobileNetV2")) return [3, 4];
                return [4, import("./model_imports/mobilenet_v2_mid.js")];
            case 3:
                (weightBundles = (_c.sent()).weightBundles);
                return [3, 8];
            case 4:
                if (!(modelName === "MobileNetV2Mid")) return [3, 6];
                return [4, import("./model_imports/mobilenet_v2_mid.js")];
            case 5:
                (weightBundles = (_c.sent()).weightBundles);
                return [3, 8];
            case 6:
                if (!(modelName === "InceptionV3")) return [3, 8];
                return [4, import("./model_imports/mobilenet_v2_mid.js")];

(...)