Add tool for model comparison (experimental)

User can select two model and their meta is fetched from GitHub. Features, accuracy figures and speed benchmarks are displayed in a table, with an additional chart comparing the accuracy scores if available. Main use case: demonstrating and visualising trade-offs between larger and smaller models of the same type.
This commit is contained in:
ines 2017-10-30 14:09:43 +01:00
parent fb2710211b
commit 1eb1ed0c7c
4 changed files with 249 additions and 1 deletions

View File

@ -46,6 +46,7 @@ if IS_PAGE
- NavHighlighter = "new NavHighlighter('data-section', 'data-nav');"
- GitHubEmbed = "new GitHubEmbed('" + SOCIAL.github + "', 'data-gh-embed');"
- ModelLoader = "new ModelLoader('" + MODELS_REPO + "'," + JSON.stringify(CURRENT_MODELS) + "," + JSON.stringify(MODEL_LICENSES) + "," + JSON.stringify(MODEL_BENCHMARKS) + ");"
- ModelComparer = "new ModelComparer('" + MODELS_REPO + "'," + JSON.stringify(MODEL_LICENSES) + "," + JSON.stringify(MODEL_BENCHMARKS) + "," + JSON.stringify(LANGUAGES) + "," + JSON.stringify(MODEL_META) + "," + JSON.stringify(default_models || false) + ");"
//- Browsers with JS module support.
Will be ignored otherwise.
@ -64,6 +65,9 @@ script(type="module")
if HAS_MODELS
| import { ModelLoader } from '/assets/js/models.js';
!=ModelLoader
if compare_models
| import { ModelComparer } from '/assets/js/models.js';
!=ModelComparer
//- Browsers with no JS module support.
Won't be fetched or interpreted otherwise.
@ -78,3 +82,5 @@ script(nomodule)
!=GitHubEmbed
if HAS_MODELS
!=ModeLoader
if compare_models
!=ModelComparer

View File

@ -158,3 +158,152 @@ export class ModelLoader {
}
}
export class ModelComparer {
/**
* Compare to model meta files and render chart and comparison table.
* @param {string} repo - Path tp GitHub repository containing releases.
* @param {Object} licenses - License IDs mapped to URLs.
* @param {Object} benchmarkKeys - Objects of available keys by type, e.g.
* 'parser', 'ner', 'speed', mapped to labels.
* @param {Object} languages - Available languages, ID mapped to name.
* @param {Object} defaultModels - Models to compare on load, 'model1' and
* 'model2' mapped to model names.
*/
constructor(repo, licenses = {}, benchmarkKeys = {}, languages = {}, labels = {}, defaultModels) {
this.url = `https://raw.githubusercontent.com/${repo}/master`;
this.repo = `https://github.com/${repo}`;
this.tpl = new Templater('compare');
this.benchKeys = benchmarkKeys;
this.licenses = licenses;
this.languages = languages;
this.labels = labels;
this.models = {};
this.colors = CHART_COLORS;
this.defaultModels = defaultModels;
this.fetchCompat()
.then(compat => this.init(compat))
.catch(this.showError.bind(this))
}
init(compat) {
this.compat = compat;
const selectA = this.tpl.get('model1');
const selectB = this.tpl.get('model2');
selectA.addEventListener('change', this.onSelect.bind(this));
selectB.addEventListener('change', this.onSelect.bind(this));
this.chart = new Chart('chart_compare_accuracy', { type: 'bar',
options: { responsive: true, scales: {
yAxes: [{ label: 'Accuracy', ticks: { min: 70 }}],
xAxes: [{ barPercentage: 0.75 }]
}}
});
if (this.defaultModels) {
selectA.value = this.defaultModels.model1;
selectB.value = this.defaultModels.model2;
this.getModels(this.defaultModels);
}
}
fetchCompat() {
return new Promise((resolve, reject) =>
fetch(`${this.url}/compatibility.json`)
.then(res => handleResponse(res))
.then(json => json.ok ? resolve(json.spacy) : reject()))
}
fetchModel(name) {
const version = getLatestVersion(name, this.compat);
const modelName = `${name}-${version}`;
return new Promise((resolve, reject) => {
// resolve immediately if model already loaded, e.g. in this.models
if (this.models[name]) resolve(this.models[name]);
else fetch(`${this.url}/meta/${modelName}.json`)
.then(res => handleResponse(res))
.then(json => json.ok ? resolve(this.saveModel(name, json)) : reject())
})
}
/**
* "Save" meta to this.models so it only has to be fetched from GitHub once.
* @param {string} name - The model name.
* @param {Object} data - The model meta data.
*/
saveModel(name, data) {
this.models[name] = data;
return data;
}
showError() {
this.tpl.get('result').style.display = 'none';
this.tpl.get('error').style.display = 'block';
}
onSelect(ev) {
const modelId = ev.target.value;
const otherId = (ev.target.id == 'model1') ? 'model2' : 'model1';
const otherVal = this.tpl.get(otherId);
const otherModel = otherVal.options[otherVal.selectedIndex].value;
if (otherModel != '') this.getModels({
[ev.target.id]: modelId,
[otherId]: otherModel
})
}
getModels({ model1, model2 }) {
this.tpl.get('result').setAttribute('data-loading', '');
this.fetchModel(model1)
.then(data1 => this.fetchModel(model2)
.then(data2 => this.render({ model1: data1, model2: data2 })))
.catch(this.showError.bind(this))
}
/**
* Render two models, and populate the chart and table. Currently quite hacky :(
* @param {Object} models - The models to render.
* @param {Object} models.model1 - The first model (via first <select>).
* @param {Object} models.model2 - The second model (via second <select>).
*/
render({ model1, model2 }) {
const accKeys = Object.assign({}, this.benchKeys.parser, this.benchKeys.ner);
const allKeys = [...Object.keys(model1.accuracy || []), ...Object.keys(model2.accuracy || [])];
const metaKeys = Object.keys(accKeys).filter(k => allKeys.includes(k));
const labels = metaKeys.map(key => accKeys[key]);
const datasets = [model1, model2]
.map(({ lang, name, version, accuracy = {} }, i) => ({
label: `${lang}_${name}-${version}`,
backgroundColor: this.colors[`model${i + 1}`],
data: metaKeys.map(key => (accuracy[key] || 0).toFixed(2))
}));
this.chart.data = { labels, datasets };
this.chart.update();
[model1, model2].forEach((model, i) => this.renderTable(metaKeys, i + 1, model));
this.tpl.get('result').removeAttribute('data-loading');
}
renderTable(metaKeys, i, { lang, name, version, size, description,
notes, author, url, license, sources, vectors, pipeline, accuracy = {},
speed = {}}) {
const type = name.split('_')[0]; // extract type from model name
const genre = name.split('_')[1]; // extract genre from model name
this.tpl.fill(`table-head${i}`, `${lang}_${name}`);
this.tpl.get(`link${i}`).setAttribute('href', `/models/${lang}#${lang}_${name}`);
this.tpl.fill(`download${i}`, `spacy download ${lang}_${name}\n`);
this.tpl.fill(`lang${i}`, this.languages[lang] || lang);
this.tpl.fill(`type${i}`, this.labels[type] || type);
this.tpl.fill(`genre${i}`, this.labels[genre] || genre);
this.tpl.fill(`version${i}`, formats.version(version), true);
this.tpl.fill(`size${i}`, size);
this.tpl.fill(`desc${i}`, description || 'n/a');
this.tpl.fill(`pipeline${i}`, formats.pipeline(pipeline), true);
this.tpl.fill(`vectors${i}`, formats.vectors(vectors));
this.tpl.fill(`sources${i}`, formats.sources(sources));
this.tpl.fill(`author${i}`, formats.author(author, url), true);
this.tpl.fill(`license${i}`, formats.license(license, this.licenses[license]), true);
// check if model accuracy or speed includes one of the pre-set keys
for (let key of [...metaKeys, ...Object.keys(this.benchKeys.speed)]) {
if (accuracy[key]) this.tpl.fill(`${key}${i}`, accuracy[key].toFixed(2))
else if (speed[key]) this.tpl.fill(`${key}${i}`, convertNumber(Math.round(speed[key])))
else this.tpl.fill(`${key}${i}`, 'n/a')
}
}
}

View File

@ -1,7 +1,8 @@
{
"sidebar": {
"Models": {
"Overview": "./"
"Overview": "./",
"Comparison": "comparison"
},
"Language models": {
@ -26,6 +27,17 @@
}
},
"comparison": {
"title": "Model Comparison",
"teaser": "Compare spaCy's statistical models and their accuracy.",
"tag": "experimental",
"compare_models": true,
"default_models": {
"model1": "en_core_web_sm",
"model2": "en_core_web_lg"
}
},
"MODELS": {
"en": ["en_core_web_sm", "en_core_web_lg", "en_vectors_web_lg"],
"de": ["de_dep_news_sm"],

View File

@ -0,0 +1,81 @@
//- 💫 DOCS > MODELS > COMPARISON
include ../_includes/_mixins
p
| This experimental tool helps you compare spaCy's statistical models
| by features, accuracy and speed. This can be especially useful to get an
| idea of the trade-offs between larger and smaller models of the same
| type. For example, #[code lg] models tend to be more accurate than
| the corresponding #[code sm] versions but they're often significantly
| larger in file size and memory usage.
- TPL = "compare"
+grid.o-box
for i in [1, 2]
+grid-col("half", "no-gutter")
label.u-heading.u-text-label.u-text-center.u-color-theme(for="model#{i}") Model #{i}
.o-field.o-grid.o-grid--vcenter.u-padding-small
select.o-field__select.u-text-small(id="model#{i}" data-tpl=TPL data-tpl-key="model#{i}")
option(selected="" disabled="" value="") Select model...
for models, _ in MODELS
for model in models
option(value=model)=model
div(data-tpl=TPL data-tpl-key="error" style="display: none")
+infobox
| Unable to load model details and accuracy figures from GitHub to
| compare the models. For details of the individual models, see the
| overview of the
| #[+a(gh("spacy-models") + "/releases") latest model releases].
div(data-tpl=TPL data-tpl-key="result")
+chart("compare_accuracy", 350)
+aside-code("Download", "text")(style="display: none")
for i in [1, 2]
span(data-tpl=TPL data-tpl-key="download#{i}")
+table.o-block-small(data-tpl=TPL data-tpl-key="table")
+row("head")
+head-cell
for i in [1, 2]
+head-cell(style="width: 40%")
a(data-tpl=TPL data-tpl-key="link#{i}")
code(data-tpl=TPL data-tpl-key="table-head#{i}" style="text-transform: initial; font-weight: normal")
for label, id in {lang: "Language", type: "Type", genre: "Genre"}
+row
+cell #[+label=label]
for i in [1, 2]
+cell(data-tpl=TPL data-tpl-key="#{id}#{i}") n/a
for label in ["Version", "Size", "Pipeline", "Vectors", "Sources", "Author", "License"]
- var field = label.toLowerCase()
+row
+cell.u-nowrap
+label=label
if MODEL_META[field]
| #[+help(MODEL_META[field]).u-color-subtle]
for i in [1, 2]
+cell
span(data-tpl=TPL data-tpl-key=field + i) #[em n/a]
+row
+cell #[+label Description]
for i in [1, 2]
+cell.u-text-tiny(data-tpl=TPL data-tpl-key="desc#{i}") n/a
for benchmark, _ in MODEL_BENCHMARKS
- var counter = 0
for label, field in benchmark
+row((counter == 0) ? "divider" : null)
+cell.u-nowrap
+label=label
if MODEL_META[field]
| #[+help(MODEL_META[field]).u-color-subtle]
for i in [1, 2]
+cell
span(data-tpl=TPL data-tpl-key=field + i) n/a
- counter++