[custom model] Add support for MODNet

[xenova]

This PR is a step towards supporting custom models (even those not supported by transformers).

In this case, the main additional requirement is to support size_divisbility after resizing. Future PRs will add:

• A way to remove the unsupported warning.
• A way to set what type of model it is (encoder-only, encoder-decoder, decoder-only, etc).

Usage:

import { AutoProcessor, RawImage, AutoModel } from '@xenova/transformers';

// Load model and processor
const model = await AutoModel.from_pretrained('Xenova/modnet-onnx', { quantized: false });
const processor = await AutoProcessor.from_pretrained('Xenova/modnet-onnx');

// Load image from URL
const url = 'https://images.pexels.com/photos/5965592/pexels-photo-5965592.jpeg?auto=compress&cs=tinysrgb&w=1024';
const image = await RawImage.fromURL(url);

// Process image
const { pixel_values: input } = await processor(image);

// Predict alpha matte
const { output } = await model({input});

// Save output image
const outputImage = await RawImage.fromTensor(output[0].mul(255).to('uint8')).resize(image.width, image.height);
outputImage.save('output.png');

Input image	Visualized mask

For more information, see the original repo and example colab.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[xenova]

For future reference, the quantized version of the model requires --per_channel False --reduce_range False

[cyio]

online demo: https://huggingface.co/spaces/oakerx/vanilla-js-portrait-matting