Set the weight file location of tf.loadLayersModel

TensorFlow.js version

1.3.2

Describe the problem or feature request

Hello. As I tried to create and use a simple model, it seems that the location of the weight file can be set automatically or only the prefix of the path can be set. In the case of me, I want to set the position of the weight like setting the position of the model, but I wonder if there is a reason not to set it.

The API I am expecting is as follows.

const loadModel = await tf.loadLayersModel(model.default,{weightPath:weight.default});

[Rcuz8]

Tensorflow.js version: 2.0.1

I’m having an issue during this process.

I’m using Firebase Storage, which has a url looking like this:

<APP_URL>/o/directory_1%2Fdirectory_2%2Fdirectory_3%2Fmodel.json?<Query Params>                 => 200 OK
<APP_URL>/o/directory_1%2Fdirectory_2%2Fdirectory_3%2Fgroup1-shard1of1.bin?<Query Params>       => 200 OK
...

This is close to a standard directory-like convention (/ => %2F) but doesn’t conform exactly. When I load the model, the parsed shard URL is:

<APP_URL>/o/group1-shard1of1.bin?<Query Params>

which cuts off the directory structure and yields a 404. The queries on their own, and the model.json query in the program, yield 200 status code responses and the correct results.

I’ve tried loading the model in two ways.

Method 1:

const loadhttp = (url, dir2, dir3) =>
  tf.io.http(url, {
    weightPathPrefix: `directory_1%2F${dir2}%2F${dir3}%2F`,
  });
...

let http = loadhttp(model_url, dir2, dir3);

tf.loadLayersModel(http).then((model) => {
    console.log("Loaded model.");
    console.log(model);
})

An example from the docs specifies the entire path, so I when do that by adjusting the function to:

const loadhttp = (url) =>
  tf.io.http(url, {
    weightPathPrefix: before(url, 'model.json'),
  });

const before = (str, sub) => str.substring(0, str.indexOf(sub))

I still get the same result of:

<APP_URL>/o/group1-shard1of1.bin?<Query Params>          => 404 Not Found

Method 2:

I’ve adjusted the model.json file’s weightsManifest from:

"weightsManifest": [{"paths": [group1-shard1of1.bin"] ...

to

"weightsManifest": [{"paths": ["directory_1%2Fdirectory_2%2Fdirectory_3%2Fgroup1-shard1of1.bin"] ...

(and the JS code to)

const loadhttp = (url) =>
  tf.io.http(url);

And it was surprisingly unsuccessful (same incorrect 404 weights URL). My assumption was this directed the second query, and my only explanation is that this value is being overridden in some way. Although this is the less ideal method, I was fairly certain it would work.

[JulianKlug]

I am experiencing the same issue as described by @Rcuz8 above. Manually replacing “Functional” by “Model” in the model.json file solved the issue for me.