Transformers.js  provides browser-based machine learning capabilities that mirror the Hugging Face Transformers Python library. It supports a wide range of pretrained models across multiple domains.

Under the hood, Transformers.js  uses ONNX Runtime for model execution in the browser. You can convert existing PyTorch, TensorFlow, or JAX models to ONNX format using the Optimum library.

import * as transformers from from '@huggingface/transformers';
 
const path = 'xenova/resnet-50';
const options = {
  dtype: 'fp16',
  device: 'webnn-gpu', // 'webnn', 'webnn-cpu', 'webnn-npu'
  session_options: {
      freeDimensionOverrides: {
        batch_size: 1,
        num_channels: 3,
        height: 224,
        width: 224,
      },
      context: {},
  },
};
 
const classifier = await transformers.pipeline('image-classification', path, options);
let [err, output] = await asyncErrorHandling(classifier(imageUrl, { topk: 3 }));

WebNN does not currently support dynamic shapes and requires free_dimension_overrides to be set.

free_dimension_overrides can be set in sample code above or just in config.json of your models as a field within “transformers.js_config”. When free_dimension_overrides is not set, you may experience significant performance degradation.