Local model import not implemented

[DoormatIka]

I'd like to load a model from disk with this code:

const url = "file:///Users/Thaddea/Documents/vslily/Programming/js/kogachan/data/models/";
const model = await nsfw.load(url, { size: 299, });

but it throws:

TypeError: fetch failed
    at Object.fetch (node:internal/deps/undici/undici:11118:11) {
  cause: Error: not implemented... yet...
      at makeNetworkError (node:internal/deps/undici/undici:9273:35)
      at schemeFetch (node:internal/deps/undici/undici:10603:18)
      at node:internal/deps/undici/undici:10483:26
      at mainFetch (node:internal/deps/undici/undici:10500:11)
      at fetching (node:internal/deps/undici/undici:10457:7)
      at Agent.fetch2 (node:internal/deps/undici/undici:10337:20)
      at Object.fetch (node:internal/deps/undici/undici:11116:28)
      at fetch (node:internal/process/pre_execution:216:25)
      at HTTPRequest.PlatformNode.fetch (C:\Users\Thaddea\Documents\vslily\Programming\js\kogachan\tfjs-core\src\platforms\platform_node.ts:55:27)
      at HTTPRequest.<anonymous> (C:\Users\Thaddea\Documents\vslily\Programming\js\kogachan\tfjs-core\src\io\http.ts:142:43) {
    [cause]: undefined
  }
}

[DoormatIka]

I just had to put "model.json" at the end of the url.

[GantMan]

OH glad you got it!

[holzmaster]

I'm trying to deploy something via docker and getting this error.
The files+model.json are located in /model.

When using Node.js 20, exactly the error from the original post is thrown, since undici (Node's new fetch implementation) doesn't support file-URLs.

However, when passing --no-experimental-fetch to node, this will (supposedly) trigger tensorflow to use node-fetch instead. This almost works, but these two URLs don't work:

file://model/model.json

/app/node_modules/node-fetch/lib/index.js:1331
    throw new TypeError('Only HTTP(S) protocols are supported');

file:///model/model.json
(note the third /)

/app/node_modules/node-fetch/lib/index.js:1327
    throw new TypeError('Only absolute URLs are supported');

The latter URL is created by using Node's url module:

import * as path from "node:path";
import { pathToFileURL } from "node:url";

const modelPath = path.join(config.modelDir, "model.json");
const modelUrl = pathToFileURL(modelPath);
console.log(modelUrl.toString());

Normal paths are not supported, which seems to be a design decision:
/model/model.json

/app/node_modules/node-fetch/lib/index.js:1327
    throw new TypeError('Only absolute URLs are supported');

Any chance to load the files from disk?

Edit:
I also tried using a file io handler:

const model = new NSFWJS(tf.io.fileSystem("/model/model.json"), {
	size: 224,
	type: "graph",
});
await model.load();

Results in:

/app/node_modules/@tensorflow/tfjs-converter/dist/tf-converter.node.js:30397
        this.version = graph.versions.producer + "." + graph.versions.minConsumer;
                                      ^

TypeError: Cannot read properties of undefined (reading 'producer')
    at GraphModel.loadSync (/app/node_modules/@tensorflow/tfjs-converter/dist/tf-converter.node.js:30397:39)
    at /app/node_modules/@tensorflow/tfjs-converter/dist/tf-converter.node.js:30373:72

When omitting type (I suppose that uses a layer model):

/app/node_modules/@tensorflow/tfjs-layers/dist/tf-layers.node.js:258
        var _this = _super.call(this, message) || this;
                           ^

ValueError: Error when checking : expected input_1 to have shape [null,299,299,3] but got array with shape [1,224,224,3].
    at new ValueError (/app/node_modules/@tensorflow/tfjs-layers/dist/tf-layers.node.js:258:28)