Node JS - Load Local Model

[benmussi]

Hello,

I am using NSFW JS on a Node JS application - (https://github.com/infinitered/nsfwjs#node-js-app)

Per the instructions, I can properly load the model using:
const model = await nsfw.load();

However, instead of using the default model on the S3 bucket, I would like to store the model locally and so I have saved the model files locally in a folder with a relative path of ./model/

I tried multiple ways of loading the locally stored model without success and could use help, please. Per the instructions, I tried 3 different ways:

const model = await nsfw.load("file://./model/");
==> This resulted in the error "Only HTTP(S) protocols are supported"

const model = await nsfw.load("file://" + __dirname + "/model/");
==> This resulted in the error "Only absolute URLS are supported"

const model = await nsfw.load("file://c/:\\Users\\User\\NSFWJS\\model");
==> This resulted in the error "Only absolute URLS are supported"

Can anyone provide a suggestion? Thanks!

[rizkiramadhanx]

me to

[cdanwards]

@benmussi @rizkiramadhanx taking a look at this and trying to reproduce. Which of the examples are you referencing? Do you have any other relevant information to provide?

[FurkanGozukara]

i have the same problem

i need to load model locally and classify images locally on my pc to batch process

how can i do that? a simple code would be amazing

[georgzoeller]

I managed to make the wasm model work in nodejs. Was a pain.

• tf-backend-wasm
• setWasmPaths (path to model)
• monkeypatch global fetch with a custom url format (wasm://filename} that, when detected, returns the model via file.read, otherwise return the original fetch
• use lovell/sharp to png images, clip the alpha channel and manually convert to tensor3d since we are in node.

You have to return something like this

...
 return {
            ok: true,
            arrayBuffer: async () => buffer,
            json: async () => {
                const text = arrayBufferToString(buffer)
                
                const ret = JSON.parse(text)
                
                return ret
                
            },
            text: async () => buffer.toString(),
        };
        ```

[hjaber]

Works for me on bun and node:

root/
├── public/
│   └── nsfwjs/
│       ├── group1-shard1of2
│       ├── group1-shard2of2
│       └── model.json

import * as tf from "@tensorflow/tfjs-node";
import * as nsfwjs from "nsfwjs";

let nsfwModel: nsfwjs.NSFWJS | null = null;

export async function loadNSFWModel() {
  try {
    tf.enableProdMode();
    nsfwModel = await nsfwjs.load("file://./public/nsfwjs/", { type: "graph" });
  } catch (error) {
    console.error("Failed to load NSFW model:", error);
  }
}