@epfml/discojs 3.0.1-p20241025115642.0 → 3.0.1-p20241028120035.0

package/dist/models/gpt/index.js

@@ -2,21 +2,27 @@
  * this code is taken from gpt-tfjs with modifications from @peacefulotter and @lukemovement
  **/
 import createDebug from "debug";
-import { List } from 'immutable';
-import * as tf from '@tensorflow/tfjs';
-import { WeightsContainer } from '../../index.js';
+import { List, Range } from "immutable";
+import * as tf from "@tensorflow/tfjs";
+import { WeightsContainer } from "../../index.js";
 import { Model, EpochLogs } from "../index.js";
-import { GPTForCausalLM } from './model.js';
-import { DEFAULT_CONFIG } from './config.js';
-import evaluate from './evaluate.js';
+import { GPTModel } from "./model.js";
+import { DEFAULT_CONFIG } from "./config.js";
+import evaluate from "./evaluate.js";
 const debug = createDebug("discojs:models:gpt");
 export class GPT extends Model {
     model;
+    #blockSize;
     #maxBatchCount;
+    #vocabSize;
     constructor(partialConfig, layersModel) {
         super();
-        this.model = new GPTForCausalLM(partialConfig, layersModel);
+        const model = new GPTModel(partialConfig, layersModel);
+        model.compile();
+        this.model = model;
+        this.#blockSize = partialConfig?.blockSize ?? DEFAULT_CONFIG.blockSize;
         this.#maxBatchCount = partialConfig?.maxIter ?? DEFAULT_CONFIG.maxIter;
+        this.#vocabSize = partialConfig?.vocabSize ?? DEFAULT_CONFIG.vocabSize;
     }
     /**
      * The GPT train methods wraps the model.fitDataset call in a for loop to act as a generator (of logs)
@@ -27,26 +33,20 @@ export class GPT extends Model {
      * @param epochs the number of passes of the training dataset
      * @param tracker
      */
-    async *train(trainingData, validationData) {
-        this.model.compile();
-        const batches = await trainingData.iterator(); // tf.LazyIterator isn't an AsyncGenerator
+    async *train(trainingDataset, validationDataset) {
         let batchesLogs = List();
-        for (let batchNumber = 0; batchNumber < this.#maxBatchCount; batchNumber++) {
-            const iteration = await batches.next();
-            if (iteration.done)
-                break;
-            const batch = iteration.value;
+        for await (const [batch, _] of trainingDataset.zip(Range(0, this.#maxBatchCount))) {
             const batchLogs = await this.#runBatch(batch);
-            tf.dispose(batch);
             yield batchLogs;
             batchesLogs = batchesLogs.push(batchLogs);
         }
-        const validation = validationData && (await this.#evaluate(validationData));
+        const validation = validationDataset && (await this.#evaluate(validationDataset));
         return new EpochLogs(batchesLogs, validation);
     }
     async #runBatch(batch) {
+        const tfBatch = this.#batchToTF(batch);
         let logs;
-        await this.model.fitDataset(tf.data.array([batch]), {
+        await this.model.fitDataset(tf.data.array([tfBatch]), {
             epochs: 1,
             verbose: 0, // don't pollute
             callbacks: {
@@ -55,6 +55,7 @@ export class GPT extends Model {
                 },
             },
         });
+        tf.dispose(tfBatch);
         if (logs === undefined)
             throw new Error("batch didn't gave any logs");
         const { loss, acc: accuracy } = logs;
@@ -67,38 +68,56 @@ export class GPT extends Model {
         };
     }
     async #evaluate(dataset) {
-        const evaluation = await evaluate(this.model, dataset.map((t) => {
-            switch (t) {
-                case null:
-                case undefined:
-                    throw new Error("nullish value in dataset");
-                default:
-                    // TODO unsafe cast
-                    return t;
-            }
-        }), this.config.maxEvalBatches);
+        const evaluation = await evaluate(this.model, tf.data.generator(async function* () {
+            yield* dataset.map((batch) => this.#batchToTF(batch));
+        }.bind(this)), this.config.maxEvalBatches);
         return {
             accuracy: evaluation.val_acc,
             loss: evaluation.val_loss,
         };
     }
-    predict(input) {
-        const ret = this.model.predict(input);
-        if (Array.isArray(ret)) {
-            throw new Error("prediction yield many Tensors but should have only returned one");
-        }
-        return Promise.resolve(ret);
+    #batchToTF(batch) {
+        return tf.tidy(() => ({
+            xs: tf.stack(batch.map(([line]) => tf.tensor1d(line.toArray(), "int32")).toArray()), // cast as stack doesn't type
+            ys: tf.stack(batch
+                .map(([line, next]) => tf.oneHot(line.shift().push(next).toArray(), this.#vocabSize))
+                .toArray()), // cast as oneHot/stack doesn't type
+        }));
     }
-    async generate(input, tokenizer, newTokens = 10) {
-        const { input_ids: tokens } = await tokenizer(input, { return_tensor: false });
-        const generationConfig = {
-            maxNewTokens: newTokens,
+    async predict(batch, options) {
+        const config = {
             temperature: 1.0,
-            doSample: false
+            doSample: false,
+            ...options,
         };
-        const predictedTokens = await this.model.generate(tokens, generationConfig);
-        const generatedWords = tokenizer.decode(predictedTokens[0]);
-        return generatedWords;
+        return List(await Promise.all(batch.map((tokens) => this.#predictSingle(tokens, config))));
+    }
+    async #predictSingle(tokens, config) {
+        // slice input tokens if longer than context length
+        tokens = tokens.slice(-this.#blockSize);
+        const input = tf.tidy(() => tf.tensor1d(tokens.toArray(), "int32").expandDims(0));
+        const logits = tf.tidy(() => {
+            const output = this.model.predict(input);
+            if (Array.isArray(output))
+                throw new Error("The model outputs too multiple values");
+            if (output.rank !== 3)
+                throw new Error("The model outputs wrong shape");
+            return output.squeeze([0]);
+        });
+        input.dispose();
+        const probs = tf.tidy(() => logits
+            .slice([logits.shape[0] - 1])
+            .squeeze([0])
+            .div(config.temperature)
+            .softmax());
+        logits.dispose();
+        const next = tf.tidy(() => config.doSample
+            ? tf.multinomial(probs, 1).squeeze([0])
+            : probs.argMax());
+        probs.dispose();
+        const ret = await next.array();
+        next.dispose();
+        return ret;
     }
     get config() {
         return this.model.getGPTConfig;
@@ -117,7 +136,7 @@ export class GPT extends Model {
     serialize() {
         return {
             weights: this.weights,
-            config: this.config
+            config: this.config,
         };
     }
     extract() {

package/dist/models/gpt/model.d.ts

@@ -14,25 +14,11 @@ export declare abstract class Dataset<T> {
  * GPTModel extends tf.LayersModel and overrides tfjs' default training loop
  *
  */
-declare class GPTModel extends tf.LayersModel {
+export declare class GPTModel extends tf.LayersModel {
     protected readonly config: Required<GPTConfig>;
     constructor(partialConfig?: GPTConfig, layersModel?: tf.LayersModel);
     get getGPTConfig(): Required<GPTConfig>;
     compile(): void;
     fitDataset<T>(dataset: Dataset<T>, trainingArgs: tf.ModelFitDatasetArgs<T>): Promise<tf.History>;
 }
-interface GenerateConfig {
-    maxNewTokens: number;
-    temperature: number;
-    doSample: boolean;
-}
-/**
- * GPTForCausalLM stands for GPT model for Causal Language Modeling. Causal because it only looks at past tokens and not future ones
- * This class extends GPTModel and adds supports for text generation
- *
- */
-export declare class GPTForCausalLM extends GPTModel {
-    generate(idxRaw: tf.TensorLike, conf: GenerateConfig): Promise<number[][]>;
-    private generateOnce;
-}
 export {};

package/dist/models/gpt/model.js

@@ -9,7 +9,7 @@ const debug = createDebug("discojs:models:gpt");
  * GPTModel extends tf.LayersModel and overrides tfjs' default training loop
  *
  */
-class GPTModel extends tf.LayersModel {
+export class GPTModel extends tf.LayersModel {
     config;
     constructor(partialConfig, layersModel) {
         // Fill missing config parameters with default values
@@ -124,77 +124,3 @@ class GPTModel extends tf.LayersModel {
         return new tf.History();
     }
 }
-const defaultGenerateConfig = {
-    maxNewTokens: 20,
-    temperature: 1.0,
-    doSample: false
-};
-function prepareIdx(idx) {
-    return tf.tidy(() => {
-        let ret;
-        if (idx instanceof tf.Tensor) {
-            ret = idx.clone();
-        }
-        else {
-            ret = tf.tensor(idx);
-        }
-        if (ret.dtype !== 'int32') {
-            ret = ret.toInt();
-        }
-        switch (ret.shape.length) {
-            case 1:
-                return ret.expandDims(0);
-            case 2:
-                return ret;
-            default:
-                throw new Error('unexpected shape');
-        }
-    });
-}
-/**
- * GPTForCausalLM stands for GPT model for Causal Language Modeling. Causal because it only looks at past tokens and not future ones
- * This class extends GPTModel and adds supports for text generation
- *
- */
-export class GPTForCausalLM extends GPTModel {
-    async generate(idxRaw, conf) {
-        const config = Object.assign({}, defaultGenerateConfig, conf);
-        let idx = prepareIdx(idxRaw);
-        for (let step = 0; step < config.maxNewTokens; step++) {
-            const idxNext = this.generateOnce(this, idx, config);
-            const idxNew = idx.concat(idxNext, 1);
-            tf.dispose(idx);
-            idx = idxNew;
-            tf.dispose(idxNext);
-        }
-        const idxArr = await idx.array();
-        tf.dispose(idx);
-        return idxArr;
-    }
-    generateOnce(model, idx, config) {
-        const idxNext = tf.tidy(() => {
-            // slice input tokens if longer than context length
-            const blockSize = this.config.blockSize;
-            idx = idx.shape[1] <= blockSize
-                ? idx : idx.slice([0, idx.shape[1] - blockSize]);
-            const output = model.predict(idx);
-            if (Array.isArray(output))
-                throw new Error('The model outputs too multiple values');
-            if (output.shape.length !== 3)
-                throw new Error('The model outputs wrong shape');
-            const logits = output;
-            const logitsScaled = logits
-                .slice([0, idx.shape[1] - 1, 0])
-                .reshape([logits.shape[0], logits.shape[2]])
-                .div(tf.scalar(config.temperature));
-            const probs = logitsScaled.softmax(-1);
-            if (config.doSample) {
-                return tf.multinomial(probs, 1);
-            }
-            else {
-                return probs.argMax(-1).expandDims(1);
-            }
-        });
-        return idxNext;
-    }
-}

package/dist/models/model.d.ts

@@ -1,14 +1,11 @@
-import type tf from "@tensorflow/tfjs";
-import type { WeightsContainer } from "../index.js";
+import type { Batched, Dataset, DataFormat, DataType, WeightsContainer } from "../index.js";
 import type { BatchLogs, EpochLogs } from "./logs.js";
-export type Prediction = tf.Tensor;
-export type Sample = tf.Tensor;
 /**
  * Trainable predictor
  *
  * Allow for various implementation of models (various train function, tensor-library, ...)
  **/
-export declare abstract class Model implements Disposable {
+export declare abstract class Model<D extends DataType> implements Disposable {
     /** Return training state */
     abstract get weights(): WeightsContainer;
     /** Set training state */
@@ -16,15 +13,13 @@ export declare abstract class Model implements Disposable {
     /**
      * Improve predictor
      *
-     * @param trainingData dataset to optimize for
-     * @param validationData dataset to measure how well it is training
-     * @param epochs number of pass over the training dataset
-     * @param tracker watch the various steps
-     * @yields on every epoch, training can be stop by `return`ing it
+     * @param trainingDataset dataset to optimize for
+     * @param validationDataset dataset to measure how well it is training
+     * @yields on every epoch, training can be stop by `return`ing or `throw`ing it
      */
-    abstract train(trainingData: tf.data.Dataset<tf.TensorContainer>, validationData?: tf.data.Dataset<tf.TensorContainer>): AsyncGenerator<BatchLogs, EpochLogs>;
+    abstract train(trainingDataset: Dataset<Batched<DataFormat.ModelEncoded[D]>>, validationDataset?: Dataset<Batched<DataFormat.ModelEncoded[D]>>): AsyncGenerator<BatchLogs, EpochLogs>;
     /** Predict likely values */
-    abstract predict(input: Sample): Promise<Prediction>;
+    abstract predict(batch: Batched<DataFormat.ModelEncoded[D][0]>): Promise<Batched<DataFormat.ModelEncoded[D][1]>>;
     /**
      * This method is automatically called to cleanup the memory occupied by the model
      * when leaving the definition scope if the instance has been defined with the `using` keyword.

package/dist/models/tfjs.d.ts

@@ -1,21 +1,22 @@
 import * as tf from '@tensorflow/tfjs';
-import { WeightsContainer } from '../index.js';
+import { Batched, Dataset, DataFormat, DataType, WeightsContainer } from "../index.js";
 import { BatchLogs } from './index.js';
 import { Model } from './index.js';
-import { Prediction, Sample } from './model.js';
 import { EpochLogs } from './logs.js';
+type Serialized<D extends DataType> = [D, tf.io.ModelArtifacts];
 /** TensorFlow JavaScript model with standard training */
-export declare class TFJS extends Model {
+export declare class TFJS<D extends "image" | "tabular"> extends Model<D> {
     #private;
+    readonly datatype: D;
     private readonly model;
     /** Wrap the given trainable model */
-    constructor(model: tf.LayersModel);
+    constructor(datatype: D, model: tf.LayersModel);
     get weights(): WeightsContainer;
     set weights(ws: WeightsContainer);
-    train(trainingData: tf.data.Dataset<tf.TensorContainer>, validationData?: tf.data.Dataset<tf.TensorContainer>): AsyncGenerator<BatchLogs, EpochLogs>;
-    predict(input: Sample): Promise<Prediction>;
-    static deserialize(raw: tf.io.ModelArtifacts): Promise<Model>;
-    serialize(): Promise<tf.io.ModelArtifacts>;
+    train(trainingDataset: Dataset<Batched<DataFormat.ModelEncoded[D]>>, validationDataset?: Dataset<Batched<DataFormat.ModelEncoded[D]>>): AsyncGenerator<BatchLogs, EpochLogs>;
+    predict(batch: Batched<DataFormat.ModelEncoded[D][0]>): Promise<Batched<DataFormat.ModelEncoded[D][1]>>;
+    static deserialize<D extends "image" | "tabular">([datatype, artifacts,]: Serialized<D>): Promise<TFJS<D>>;
+    serialize(): Promise<Serialized<D>>;
     [Symbol.dispose](): void;
     /**
      * extract wrapped model
@@ -24,3 +25,4 @@ export declare class TFJS extends Model {
      */
     extract(): tf.LayersModel;
 }
+export {};

package/dist/models/tfjs.js

@@ -1,18 +1,22 @@
-import { List, Map } from 'immutable';
+import { List, Map, Range } from "immutable";
 import * as tf from '@tensorflow/tfjs';
-import { WeightsContainer } from '../index.js';
+import { WeightsContainer, } from "../index.js";
 import { Model } from './index.js';
 import { EpochLogs } from './logs.js';
 /** TensorFlow JavaScript model with standard training */
 export class TFJS extends Model {
+    datatype;
     model;
     /** Wrap the given trainable model */
-    constructor(model) {
+    constructor(datatype, model) {
         super();
+        this.datatype = datatype;
         this.model = model;
         if (model.loss === undefined) {
             throw new Error('TFJS models need to be compiled to be used');
         }
+        if (model.outputs.length !== 1)
+            throw new Error("only support single output model");
     }
     get weights() {
         return new WeightsContainer(this.model.weights.map((w) => w.read()));
@@ -20,57 +24,43 @@ export class TFJS extends Model {
     set weights(ws) {
         this.model.setWeights(ws.weights);
     }
-    async *train(trainingData, validationData) {
-        const batches = await trainingData.iterator(); // tf.LazyIterator isn't an AsyncGenerator
+    async *train(trainingDataset, validationDataset) {
         let batchesLogs = List();
-        for (let batchNumber = 0; true; batchNumber++) {
-            const iteration = await batches.next();
-            if (iteration.done)
-                break;
-            const batch = iteration.value;
+        for await (const [batch, batchNumber] of trainingDataset.zip(Range())) {
             const batchLogs = {
                 batch: batchNumber,
                 ...(await this.#runBatch(batch)),
             };
-            tf.dispose(batch);
             yield batchLogs;
             batchesLogs = batchesLogs.push(batchLogs);
         }
-        const validation = validationData && (await this.#evaluate(validationData));
+        const validation = validationDataset && (await this.#evaluate(validationDataset));
         return new EpochLogs(batchesLogs, validation);
     }
     async #runBatch(batch) {
-        let logs;
-        await this.model.fitDataset(tf.data.array([batch]), {
+        const { xs, ys } = this.#batchToTF(batch);
+        const { history } = await this.model.fit(xs, ys, {
             epochs: 1,
             verbose: 0, // don't pollute
-            callbacks: {
-                onEpochEnd: (_, cur) => {
-                    logs = cur;
-                },
-            },
         });
-        if (logs === undefined)
-            throw new Error("batch didn't gave any logs");
-        const { loss, acc: accuracy } = logs;
-        if (loss === undefined || isNaN(loss))
-            throw new Error("training loss is undefined or NaN");
+        const { loss: losses, acc: accuracies } = history;
+        if (losses === undefined ||
+            accuracies === undefined ||
+            typeof losses[0] !== "number" ||
+            typeof accuracies[0] !== "number" ||
+            isNaN(losses[0]) ||
+            isNaN(accuracies[0]))
+            throw new Error("training loss or accuracy is undefined or NaN");
         return {
-            accuracy,
-            loss,
+            accuracy: accuracies[0],
+            loss: losses[0],
             memoryUsage: tf.memory().numBytes / 1024 / 1024 / 1024,
         };
     }
     async #evaluate(dataset) {
-        const evaluation = await this.model.evaluateDataset(dataset.map((t) => {
-            switch (t) {
-                case null:
-                case undefined:
-                    throw new Error("nullish value in dataset");
-                default:
-                    return t;
-            }
-        }));
+        const evaluation = await this.model.evaluateDataset(tf.data.generator(async function* () {
+            yield* dataset.map((batch) => this.#batchToTF(batch));
+        }.bind(this)));
         const metricToValue = Map(List(this.model.metricsNames).zip(Array.isArray(evaluation)
             ? List(await Promise.all(evaluation.map((t) => t.data())))
             : List.of(await evaluation.data()))).map((values) => {
@@ -87,16 +77,39 @@ export class TFJS extends Model {
             throw new Error("some needed metrics are missing");
         return { accuracy, loss };
     }
-    predict(input) {
-        const ret = this.model.predict(input);
-        if (Array.isArray(ret)) {
-            throw new Error('prediction yield many Tensors but should have only returned one');
+    async predict(batch) {
+        async function cleanupPredicted(y) {
+            if (y.shape[0] === 1) {
+                // Binary classification
+                const threshold = tf.scalar(0.5);
+                const binaryTensor = y.greaterEqual(threshold);
+                const binaryArray = await binaryTensor.data();
+                tf.dispose([y, binaryTensor, threshold]);
+                return binaryArray[0];
+            }
+            // Multi-class classification
+            const indexTensor = y.argMax();
+            const indexArray = await indexTensor.data();
+            tf.dispose([y, indexTensor]);
+            return indexArray[0];
+            // Multi-label classification is not supported
         }
-        return Promise.resolve(ret);
+        const xs = this.#batchWithoutLabelToTF(batch);
+        const prediction = this.model.predict(xs);
+        if (Array.isArray(prediction))
+            throw new Error("prediction yield many Tensors but should have only returned one");
+        tf.dispose(xs);
+        if (prediction.rank !== 2)
+            throw new Error("unexpected batched prediction shape");
+        const ret = List(await Promise.all(tf.unstack(prediction).map((y) => cleanupPredicted(
+        // cast as unstack reduce by one the rank
+        y))));
+        prediction.dispose();
+        return ret;
     }
-    static async deserialize(raw) {
-        return new this(await tf.loadLayersModel({
-            load: () => Promise.resolve(raw)
+    static async deserialize([datatype, artifacts,]) {
+        return new this(datatype, await tf.loadLayersModel({
+            load: () => Promise.resolve(artifacts),
         }));
     }
     async serialize() {
@@ -115,7 +128,7 @@ export class TFJS extends Model {
         }, {
             includeOptimizer: true // keep model compiled
         });
-        return await ret;
+        return [this.datatype, await ret];
     }
     [Symbol.dispose]() {
         this.model.dispose();
@@ -128,4 +141,53 @@ export class TFJS extends Model {
     extract() {
         return this.model;
     }
+    #batchToTF(batch) {
+        const outputSize = tf.util.sizeFromShape(this.model.outputShape.map((dim) => {
+            if (Array.isArray(dim))
+                throw new Error("TODO support multiple outputs");
+            return dim ?? 1;
+        }));
+        switch (this.datatype) {
+            case "image": {
+                // cast as typescript doesn't reduce generic type
+                const b = batch;
+                return tf.tidy(() => ({
+                    xs: tf.stack(b
+                        .map(([image]) => tf.tensor3d(image.data, [image.width, image.height, 3], "float32"))
+                        .toArray()),
+                    ys: tf.stack(b
+                        .map(([_, label]) => tf.oneHot(label, outputSize, 1, 0, "int32"))
+                        .toArray()),
+                }));
+            }
+            case "tabular": {
+                // cast as typescript doesn't reduce generic type
+                const b = batch;
+                return tf.tidy(() => ({
+                    xs: tf.stack(b.map(([inputs, _]) => tf.tensor1d(inputs.toArray())).toArray()),
+                    ys: tf.stack(b.map(([_, output]) => tf.tensor1d([output])).toArray()),
+                }));
+            }
+        }
+        const _ = this.datatype;
+        throw new Error("should never happen");
+    }
+    #batchWithoutLabelToTF(batch) {
+        switch (this.datatype) {
+            case "image": {
+                // cast as typescript doesn't reduce generic type
+                const b = batch;
+                return tf.tidy(() => tf.stack(b
+                    .map((image) => tf.tensor3d(image.data, [image.width, image.height, 3], "float32"))
+                    .toArray()));
+            }
+            case "tabular": {
+                // cast as typescript doesn't reduce generic type
+                const b = batch;
+                return tf.tidy(() => tf.stack(b.map((inputs) => tf.tensor1d(inputs.toArray())).toArray()));
+            }
+        }
+        const _ = this.datatype;
+        throw new Error("should never happen");
+    }
 }

package/dist/models/tokenizer.d.ts

@@ -11,4 +11,4 @@ import { PreTrainedTokenizer } from '@xenova/transformers';
  * @param task the task object specifying which tokenizer to use
  * @returns an initialized tokenizer object
  */
-export declare function getTaskTokenizer(task: Task): Promise<PreTrainedTokenizer>;
+export declare function getTaskTokenizer(task: Task<'text'>): Promise<PreTrainedTokenizer>;

package/dist/privacy.js

@@ -5,7 +5,7 @@ async function frobeniusNorm(weights) {
         .reduce((a, b) => a.add(b))
         .data();
     if (squared.length !== 1)
-        throw new Error("unexcepted weights shape");
+        throw new Error("unexpected weights shape");
     return Math.sqrt(squared[0]);
 }
 /** Scramble weights */

package/dist/processing/image.d.ts

@@ -0,0 +1,18 @@
+import { Image } from "../index.js";
+/** Image where intensity is represented in the range 0..1 */
+export declare class NormalizedImage<D extends 1 | 3 | 4 = 1 | 3 | 4, W extends number = number, H extends number = number> {
+    readonly data: Readonly<Float32Array>;
+    readonly width: W;
+    readonly height: H;
+    readonly depth: D;
+    private constructor();
+    static from<D extends 1 | 3 | 4 = 1 | 3 | 4, W extends number = number, H extends number = number>(image: Image<D, W, H>): NormalizedImage<D, W, H>;
+}
+/** Remove the alpha channel of an image */
+export declare function removeAlpha<W extends number, H extends number>(image: Image<4, W, H>): Image<3, W, H>;
+export declare function removeAlpha<D extends 1 | 3, W extends number, H extends number>(image: Image<D | 4, W, H>): Image<D, W, H>;
+/** Convert monochrome images to multicolor */
+export declare function expandToMulticolor<W extends number, H extends number>(image: Image<1, W, H>): Image<3, W, H>;
+export declare function expandToMulticolor<D extends 3 | 4, W extends number, H extends number>(image: Image<1 | D, W, H>): Image<D, W, H>;
+export declare function resize<D extends 1 | 3 | 4, W extends number, H extends number>(width: W, height: H, image: Image<D, number, number>): Image<4, W, H>;
+export declare function normalize<D extends 1 | 3 | 4, W extends number, H extends number>(image: Image<D, W, H>): NormalizedImage<D, W, H>;