Failed to load model from local (nodejs)

[azurenekowo]

Greetings. As the title have said, I am clueless about how to load a local model. More details below.

Here is my codebase:

global.model = null
async function loadModel() {
    const loadfrom = 'file://' + __dirname.replace(/\\/g, '/') + '/mobilenet_v2_140_224/web_model/'
    console.log(loadfrom)
    const tf = require('@tensorflow/tfjs-node')
    const nsfw = require('nsfwjs')
    model = await nsfw.load(loadfrom, { type: 'graph' })
}
loadModel()

(Ignore the regex, it only changes "" to "/" because I'm on a Windows machine. removing that produces the same error)

Console output:

file://C:/Users/user/Desktop/demo/mobilenet_v2_140_224/web_model/
C:\Users\user\Desktop\demo\node_modules\node-fetch\lib\index.js:1327
                throw new TypeError('Only absolute URLs are supported');
                      ^

TypeError: Only absolute URLs are supported
    at getNodeRequestOptions (C:\Users\user\Desktop\demo\node_modules\node-fetch\lib\index.js:1327:9)
    at C:\Users\user\Desktop\demo\node_modules\node-fetch\lib\index.js:1440:19
    at new Promise (<anonymous>)
    at fetch (C:\Users\user\Desktop\demo\node_modules\node-fetch\lib\index.js:1437:9)
    at HTTPRequest.PlatformNode.fetch (C:\Users\user\Desktop\demo\node_modules\@tensorflow\tfjs-core\dist\tf-core.node.js:7430:16)
    at HTTPRequest.<anonymous> (C:\Users\user\Desktop\demo\node_modules\@tensorflow\tfjs-core\dist\tf-core.node.js:8289:55)
    at step (C:\Users\user\Desktop\demo\node_modules\@tensorflow\tfjs-core\dist\tf-core.node.js:125:27
    at Object.next (C:\Users\user\Desktop\demo\node_modules\@tensorflow\tfjs-core\dist\tf-core.node.js:74:53)
    at C:\Users\user\Desktop\demo\node_modules\@tensorflow\tfjs-core\dist\tf-core.node.js:67:71
    at new Promise (<anonymous>)

It happens with both with and without the file:// prefix.

Interestingly, if I put in a relative path that is the same format with the one in Node.js app demo on this repo readme:
file:// ./mobilenet_v2_140_224/web_model/, it threw out TypeError [ERR_INVALID_URL]: Invalid URL.

Paths I have tried:
file://./mobilenet_v2_140_224/web_model/ (removed space) => TypeError: Only HTTP(S) protocols are supported
file:///mobilenet_v2_140_224/web_model/ => TypeError: Only absolute URLs are supported
./mobilenet_v2_140_224/web_model/ => TypeError: Only absolute URLs are supported
/mobilenet_v2_140_224/web_model/ => TypeError: Only absolute URLs are supported

Both issues I have found related to locally loading the model are still open to this day: #512 #522 #652

Models I have used:
https://github.com/infinitered/nsfwjs/tree/master/example/nsfw_demo/public/quant_mid
https://github.com/GantMan/nsfw_model/releases/tag/1.1.0

Am I doing something wrong or the documentation is obsecure that I couldn't find anything? Calling load() loads the default model which turns out predicts images differently than the models on nsfwjs.com.

Any help would be greatly appreciated.

[cdanwards]

@scarletzumii sorry that you're experiencing this outstanding bug. We're taking some time to look into it!

[huhm]

overide the load function ,and post fetchFunc option to tf.loadLayersModel/tf.loadGraphModel

[sonofmagic]

I meet the same issue. I download the latest nsfw_model from GantMan/nsfw_model's releases page.

And fail to load layer model web_model or web_model_quantized.

Would you provide a example which loads local model ? Thanks.

[huhm]

You can overide the load method like this

import { NSFWJS } from 'nsfwjs';
 const baseUrl = path.join(__dirname, '../xxx/quant_mid/');
const nsfwIns=new NSFWJS(baseUrl, {
    size: 224,
    type: 'graph',
  })
overideLoad(nsfwIns, baseUrl);
function overideLoad(context: NSFWJS, modelBaseUrl: string) {
  context.load = async function nsfwnetOverideLoad() {
    const { size, type } = this.options;
    /**
     * @type String xxx/model.json
     */
    const pathOrIOHandler = this.pathOrIOHandler;
    const loadOptions = {
      onProgress: (fraction: number) => {
        {
          console.log(`ModelLoad onProgress:${(fraction * 100).toFixed(1)}%`);
        }
      },
      fetchFunc(fpath: string) {
        let curPath = fpath;
        //ForWindows
        if (!fs.existsSync(curPath)) {
          curPath = path.resolve(modelBaseUrl, './' + fpath);
        }
        console.log('ModelLoad file: ' + fpath, curPath);
        return import('node-fetch').then(({ Response: fetchResponse }) => {
          return new Promise((resolve, reject) => {
            fs.readFile(curPath, (err, data) => {
              if (err) {
                reject(err);
                return;
              }
              resolve(
                new fetchResponse(data, {
                  //   headers:{'Content-Type':
                  // }
                }),
              );
            });
          });
        });
      },
    };
    if (type === 'graph') {
      this.model = await tf.loadGraphModel(pathOrIOHandler, loadOptions);
    } else {
      // this is a Layers Model
      this.model = await tf.loadLayersModel(pathOrIOHandler, loadOptions);
      this.endpoints = this.model.layers.map((l) => l.name);
    }

    // Warmup the model.
    const result = tf.tidy(() =>
      this.model.predict(tf.zeros([1, size, size, 3])),
    ) as tf.Tensor;
    await result.data();
    result.dispose();
  };
}

[xxaier]

var bin, img, model_fp, nsfw;

import ROOT from './ROOT';

import {
  join
} from 'path';

import {
  NSFWJS
} from 'nsfwjs';

import tf from '@tensorflow/tfjs-node';

await tf.enableProdMode();

await tf.ready();

model_fp = 'file://' + join(ROOT, 'model/quant_mid/model.json');

nsfw = new NSFWJS(0, {
  size: 224
});

nsfw.load = async function() {
  this.model = (await tf.loadGraphModel(model_fp));
};

await nsfw.load();

console.log(nsfw.classify);

import {
  readFileSync
} from 'fs';

img = readFileSync('/Users/z/art/pkg/bot/adult/out/9/0.png');

bin = tf.node.decodeImage(img, 3);

console.log((await nsfw.classify(bin)));

bin.dispose();

[JemiloII]

None of these work without actually sending a request. I don't want to send any requests. Any better way to load this without node-fetch or better yet, maybe use torch instead of tensorflow?

[GantMan]

Is there a good torch for JS solution?

[miaoihan]

And? Is there no one to solve the problem?

[GantMan]

We're nearing a sprint to make NSFWJS work well on mobile.

If we have the time/resources we'll come back and look at this. It would be great if someone who figures it out to post their solution.