scripts/vsmlrt.py: add support for esrgan janai models

[hooke007]

Continue from #41 (comment)

edit:
models - 1.zip

[AkarinVS]

there will be more and more waifu2x/esrgan/cugan-like models in the future, and it's impractical to maintaining an up-to-date support of all such models in the python wrapper, at least not in a timely manner. so my questions are:

1. what should be criteria for adding one? and also for including in the official builtin models package?
2. could we design a generic python API (the current vsmlrt.inference is one such design, but obviously it fails to attract usage) to support using those models without having the users waiting for the vsmlrt.py to update?

[hooke007]

and it's impractical to maintaining an up-to-date support of all such models in the python wrapper, at least not in a timely manner.

Agreed, actually this is a draft. I prefer to only keep two or three models of janai (one for speed and one for quality). Comparing the previous esrgan's official model. janai could preserve anime's original art style rather than totally rebuilding(/breaking?) it. Of course, from the perspective of serious encoding, there is still a gap in its level. This is what I think is lacking in the current ESRGAN model.

[AkarinVS]

After extensive discussions, we propose the following policy regarding contributing models into vs-mlrt:

1. we will add a new model package (besides the existing models.7z and ext-models.7z), contrib-models.7z to host community contributed onnx files to our releases.
2. The python wrappers, depending on whether it needs to reuse existing model specific functions in vsmlrt.py, live in either of these files (both files will be included in main release, only the onnx files are packaged separately due to space constraints):
   • if it has to modify the existing code, like in this PR's case, you can modify vsmlrt.py, but please make the model id number starting from 5000, so that the code can know if the corresponding ONNX file is coming from models/ext-models or contrib-models. If you modify the vsmlrt.py, please make sure any code modifications has properly isolated from the rest of the code (to avoid breaking existing models).
   • If it's an entirely new model family that you have to write your own model specific functions (e.g. vsmlrt.Waifu2x), then you can add such code into a new contrib.py file (specifics to be released later after we've done some reorganizations).
3. for a PR to be accepted, the author must agree to maintain the ONNX model files and corresponding python wrappers. If a contributed model is broken for some time and its original author fails to respond to issue reports, we reserve the rights to remove such models.
4. we will add a rolling nightly prerelease so that users can immediately use contributed models without waiting for the next release.

What do you think about this proposed policy?

Thanks.

[hooke007]

I agree with this policy.

[hooke007]

make the model id number starting from 5000

It would be out of tuple's range.

[hooke007]

The only solution I could work out is leave away from tuple. Now it does work properly.

[WolframRhodium]

That's a better solution. LGTM. Please also attach your name.

[hooke007]

Please also attach your name.

Hmm, Where should I add into the code? I added it in Line470