Spaces:
Running
Running
| // The model part of the WhatDoWeSee Space, free of DOM code so it can be tested outside a browser. | |
| /** Models the page can run. Each loads its files from its own Hub repo at a fixed revision. */ | |
| export const MODELS = [ | |
| { | |
| id: "efficientnetv2_b0-v0.4.0", | |
| name: "EfficientNetV2-B0 · v0.4.0", | |
| repo: "StephanMeijer/WhatDoWeSee-EfficientNetV2-B0", | |
| revision: "v0.4.0", | |
| }, | |
| { | |
| id: "mobilenetv4-v0.3.0", | |
| name: "MobileNetV4 · v0.3.0", | |
| repo: "StephanMeijer/WhatDoWeSee-MobileNetV4", | |
| revision: "v0.3.0", | |
| }, | |
| { | |
| id: "efficientnet_b0-v0.3.0", | |
| name: "EfficientNet-B0 · v0.3.0", | |
| repo: "StephanMeijer/WhatDoWeSee-EfficientNet-B0", | |
| revision: "v0.3.0", | |
| }, | |
| { | |
| id: "mobilenetv4-v0.2", | |
| name: "MobileNetV4 · v0.2", | |
| repo: "StephanMeijer/WhatDoWeSee-MobileNetV4", | |
| revision: "v0.2", | |
| }, | |
| { | |
| id: "efficientnet_b0-v0.2", | |
| name: "EfficientNet-B0 · v0.2", | |
| repo: "StephanMeijer/WhatDoWeSee-EfficientNet-B0", | |
| revision: "v0.2", | |
| }, | |
| { | |
| id: "efficientnet_b0-v0.1", | |
| name: "EfficientNet-B0 · v0.1", | |
| repo: "StephanMeijer/WhatDoWeSee-EfficientNet-B0", | |
| revision: "v0.1", | |
| }, | |
| ]; | |
| export function fileUrl(model, file) { | |
| return `https://huggingface.co/${model.repo}/resolve/${model.revision}/${file}`; | |
| } | |
| /** | |
| * RGBA pixels already resized to size × size (no crop) to the model input: RGB scaled to [0, 1], | |
| * normalised with the config's mean and std, NCHW float32. | |
| */ | |
| export function toTensorData(rgba, size, mean, std) { | |
| const plane = size * size; | |
| const out = new Float32Array(3 * plane); | |
| for (let i = 0; i < plane; i++) { | |
| for (let c = 0; c < 3; c++) { | |
| out[c * plane + i] = (rgba[i * 4 + c] / 255 - mean[c]) / std[c]; | |
| } | |
| } | |
| return out; | |
| } | |
| /** Every label with its probability and threshold, the present ones first, highest first. */ | |
| export function tags(logits, config) { | |
| return config.labels | |
| .map((id, i) => { | |
| const p = 1 / (1 + Math.exp(-logits[i])); | |
| const threshold = config.thresholds[id] ?? 0.5; | |
| return { id, name: config.names[i], definition: config.definitions[id], p, threshold, present: p >= threshold }; | |
| }) | |
| .sort((a, b) => Number(b.present) - Number(a.present) || b.p - a.p); | |
| } | |
| /** Run the model on prepared pixels; `ort` is onnxruntime-web (or -node in tests). */ | |
| export async function run(ort, session, config, rgba) { | |
| const size = config.input.shape[3]; | |
| const data = toTensorData(rgba, size, config.input.mean, config.input.std); | |
| const output = await session.run({ [config.input.name]: new ort.Tensor("float32", data, [1, 3, size, size]) }); | |
| return tags(output[config.output.name].data, config); | |
| } | |