Load model directly from TensorFlow Hub

[AzizZayed]

Description

DJL currently does not support loading a model directly from TFHub like this:

Criteria<int[], Image[]> criteria =
        Criteria.builder()
        ...
        .optModelUrls(“https://tfhub.dev/deepmind/biggan-deep-256/1”)
        ...
        .build();

The above code would be the goal of the enhancement, to be able to load a model using a TFHub link. We can define a customized url like: “tfhub://deepmind/biggan-deep-256/1”, so we can handle them differently
or “djl://tfhub/deepmind/biggan-deep-256/1".

The current way to do it is a bit tricky:

1. Go on the TFHub link, right click on the download button and copy the link address
2. In the terminal, type curl -v <link address>
3. Find the location header and it points to the real download link.

curl -v https://tfhub.dev/deepmind/biggan-deep-256/1?tf-hub-format=compressed

Response:

...
< location: https://storage.googleapis.com/tfhub-modules/deepmind/biggan-deep-256/1.tar.gz
...

And now we can load the model with the given link:

Criteria<int[], Image[]> criteria =
        Criteria.builder()
        ...
        .optModelUrls(“https://storage.googleapis.com/tfhub-modules/deepmind/biggan-deep-256/1.tar.gz")
        ...
        .build();

DJL beginners will deeply appreciate such a feature since it removes another layer of complexity when trying to run inference on a model quickly. I encountered this issue myself.