@weavelogic/knowledge-graph-agent 0.6.0 → 0.7.1

package/dist/node_modules/@xenova/transformers/src/utils/image.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"image.js","sources":["../../../../../../node_modules/@xenova/transformers/src/utils/image.js"],"sourcesContent":["\n/**\n * @file Helper module for image processing. \n * \n * These functions and classes are only used internally, \n * meaning an end-user shouldn't need to access anything here.\n * \n * @module utils/image\n */\n\nimport { getFile } from './hub.js';\nimport { env } from '../env.js';\nimport { Tensor } from './tensor.js';\n\n// Will be empty (or not used) if running in browser or web-worker\nimport sharp from 'sharp';\n\nconst BROWSER_ENV = typeof self !== 'undefined';\nconst WEBWORKER_ENV = BROWSER_ENV && self.constructor.name === 'DedicatedWorkerGlobalScope';\n\nlet createCanvasFunction;\nlet ImageDataClass;\nlet loadImageFunction;\nif (BROWSER_ENV) {\n    // Running in browser or web-worker\n    createCanvasFunction = (/** @type {number} */ width, /** @type {number} */ height) => {\n        if (!self.OffscreenCanvas) {\n            throw new Error('OffscreenCanvas not supported by this browser.');\n        }\n        return new self.OffscreenCanvas(width, height)\n    };\n    loadImageFunction = self.createImageBitmap;\n    ImageDataClass = self.ImageData;\n\n} else if (sharp) {\n    // Running in Node.js, electron, or other non-browser environment\n\n    loadImageFunction = async (/**@type {sharp.Sharp}*/img) => {\n        const metadata = await img.metadata();\n        const rawChannels = metadata.channels;\n\n        let { data, info } = await img.rotate().raw().toBuffer({ resolveWithObject: true });\n\n        const newImage = new RawImage(new Uint8ClampedArray(data), info.width, info.height, info.channels);\n        if (rawChannels !== undefined && rawChannels !== info.channels) {\n            // Make sure the new image has the same number of channels as the input image.\n            // This is necessary for grayscale images.\n            newImage.convert(rawChannels);\n        }\n        return newImage;\n    }\n\n} else {\n    throw new Error('Unable to load image processing library.');\n}\n\n\n// Defined here: https://github.com/python-pillow/Pillow/blob/a405e8406b83f8bfb8916e93971edc7407b8b1ff/src/libImaging/Imaging.h#L262-L268\nconst RESAMPLING_MAPPING = {\n    0: 'nearest',\n    1: 'lanczos',\n    2: 'bilinear',\n    3: 'bicubic',\n    4: 'box',\n    5: 'hamming',\n}\n\n/**\n * Mapping from file extensions to MIME types.\n */\nconst CONTENT_TYPE_MAP = new Map([\n    ['png', 'image/png'],\n    ['jpg', 'image/jpeg'],\n    ['jpeg', 'image/jpeg'],\n    ['gif', 'image/gif'],\n]);\n\nexport class RawImage {\n\n    /**\n     * Create a new `RawImage` object.\n     * @param {Uint8ClampedArray|Uint8Array} data The pixel data.\n     * @param {number} width The width of the image.\n     * @param {number} height The height of the image.\n     * @param {1|2|3|4} channels The number of channels.\n     */\n    constructor(data, width, height, channels) {\n        this.data = data;\n        this.width = width;\n        this.height = height;\n        this.channels = channels;\n    }\n\n    /** \n     * Returns the size of the image (width, height).\n     * @returns {[number, number]} The size of the image (width, height).\n     */\n    get size() {\n        return [this.width, this.height];\n    }\n\n    /**\n     * Helper method for reading an image from a variety of input types.\n     * @param {RawImage|string|URL} input \n     * @returns The image object.\n     * \n     * **Example:** Read image from a URL.\n     * ```javascript\n     * let image = await RawImage.read('https://huggingface.co/datasets/Xenova/transformers.js-docs/resolve/main/football-match.jpg');\n     * // RawImage {\n     * //   \"data\": Uint8ClampedArray [ 25, 25, 25, 19, 19, 19, ... ],\n     * //   \"width\": 800,\n     * //   \"height\": 533,\n     * //   \"channels\": 3\n     * // }\n     * ```\n     */\n    static async read(input) {\n        if (input instanceof RawImage) {\n            return input;\n        } else if (typeof input === 'string' || input instanceof URL) {\n            return await this.fromURL(input);\n        } else {\n            throw new Error(`Unsupported input type: ${typeof input}`);\n        }\n    }\n\n\n    /**\n     * Read an image from a URL or file path.\n     * @param {string|URL} url The URL or file path to read the image from.\n     * @returns {Promise<RawImage>} The image object.\n     */\n    static async fromURL(url) {\n        let response = await getFile(url);\n        if (response.status !== 200) {\n            throw new Error(`Unable to read image from \"${url}\" (${response.status} ${response.statusText})`);\n        }\n        let blob = await response.blob();\n        return this.fromBlob(blob);\n    }\n\n    /**\n     * Helper method to create a new Image from a blob.\n     * @param {Blob} blob The blob to read the image from.\n     * @returns {Promise<RawImage>} The image object.\n     */\n    static async fromBlob(blob) {\n        if (BROWSER_ENV) {\n            // Running in environment with canvas\n            let img = await loadImageFunction(blob);\n\n            const ctx = createCanvasFunction(img.width, img.height).getContext('2d');\n\n            // Draw image to context\n            ctx.drawImage(img, 0, 0);\n\n            return new this(ctx.getImageData(0, 0, img.width, img.height).data, img.width, img.height, 4);\n\n        } else {\n            // Use sharp.js to read (and possible resize) the image.\n            let img = sharp(await blob.arrayBuffer());\n\n            return await loadImageFunction(img);\n        }\n    }\n\n    /**\n     * Helper method to create a new Image from a tensor\n     * @param {Tensor} tensor \n     */\n    static fromTensor(tensor, channel_format = 'CHW') {\n        if (tensor.dims.length !== 3) {\n            throw new Error(`Tensor should have 3 dimensions, but has ${tensor.dims.length} dimensions.`);\n        }\n\n        if (channel_format === 'CHW') {\n            tensor = tensor.transpose(1, 2, 0);\n        } else if (channel_format === 'HWC') {\n            // Do nothing\n        } else {\n            throw new Error(`Unsupported channel format: ${channel_format}`);\n        }\n        if (!(tensor.data instanceof Uint8ClampedArray || tensor.data instanceof Uint8Array)) {\n            throw new Error(`Unsupported tensor type: ${tensor.type}`);\n        }\n        switch (tensor.dims[2]) {\n            case 1:\n            case 2:\n            case 3:\n            case 4:\n                return new RawImage(tensor.data, tensor.dims[1], tensor.dims[0], tensor.dims[2]);\n            default:\n                throw new Error(`Unsupported number of channels: ${tensor.dims[2]}`);\n        }\n    }\n\n    /**\n     * Convert the image to grayscale format.\n     * @returns {RawImage} `this` to support chaining.\n     */\n    grayscale() {\n        if (this.channels === 1) {\n            return this;\n        }\n\n        let newData = new Uint8ClampedArray(this.width * this.height * 1);\n        switch (this.channels) {\n            case 3: // rgb to grayscale\n            case 4: // rgba to grayscale\n                for (let i = 0, offset = 0; i < this.data.length; i += this.channels) {\n                    const red = this.data[i];\n                    const green = this.data[i + 1];\n                    const blue = this.data[i + 2];\n\n                    newData[offset++] = Math.round(0.2989 * red + 0.5870 * green + 0.1140 * blue);\n                }\n                break;\n            default:\n                throw new Error(`Conversion failed due to unsupported number of channels: ${this.channels}`);\n        }\n        return this._update(newData, this.width, this.height, 1);\n    }\n\n    /**\n     * Convert the image to RGB format.\n     * @returns {RawImage} `this` to support chaining.\n     */\n    rgb() {\n        if (this.channels === 3) {\n            return this;\n        }\n\n        let newData = new Uint8ClampedArray(this.width * this.height * 3);\n\n        switch (this.channels) {\n            case 1: // grayscale to rgb\n                for (let i = 0, offset = 0; i < this.data.length; ++i) {\n                    newData[offset++] = this.data[i];\n                    newData[offset++] = this.data[i];\n                    newData[offset++] = this.data[i];\n                }\n                break;\n            case 4: // rgba to rgb\n                for (let i = 0, offset = 0; i < this.data.length; i += 4) {\n                    newData[offset++] = this.data[i];\n                    newData[offset++] = this.data[i + 1];\n                    newData[offset++] = this.data[i + 2];\n                }\n                break;\n            default:\n                throw new Error(`Conversion failed due to unsupported number of channels: ${this.channels}`);\n        }\n        return this._update(newData, this.width, this.height, 3);\n\n    }\n\n    /**\n     * Convert the image to RGBA format.\n     * @returns {RawImage} `this` to support chaining.\n     */\n    rgba() {\n        if (this.channels === 4) {\n            return this;\n        }\n\n        let newData = new Uint8ClampedArray(this.width * this.height * 4);\n\n        switch (this.channels) {\n            case 1: // grayscale to rgba\n                for (let i = 0, offset = 0; i < this.data.length; ++i) {\n                    newData[offset++] = this.data[i];\n                    newData[offset++] = this.data[i];\n                    newData[offset++] = this.data[i];\n                    newData[offset++] = 255;\n                }\n                break;\n            case 3: // rgb to rgba\n                for (let i = 0, offset = 0; i < this.data.length; i += 3) {\n                    newData[offset++] = this.data[i];\n                    newData[offset++] = this.data[i + 1];\n                    newData[offset++] = this.data[i + 2];\n                    newData[offset++] = 255;\n                }\n                break;\n            default:\n                throw new Error(`Conversion failed due to unsupported number of channels: ${this.channels}`);\n        }\n\n        return this._update(newData, this.width, this.height, 4);\n    }\n\n    /**\n     * Resize the image to the given dimensions. This method uses the canvas API to perform the resizing.\n     * @param {number} width The width of the new image.\n     * @param {number} height The height of the new image.\n     * @param {Object} options Additional options for resizing.\n     * @param {0|1|2|3|4|5|string} [options.resample] The resampling method to use.\n     * @returns {Promise<RawImage>} `this` to support chaining.\n     */\n    async resize(width, height, {\n        resample = 2,\n    } = {}) {\n\n        // Ensure resample method is a string\n        let resampleMethod = RESAMPLING_MAPPING[resample] ?? resample;\n\n        if (BROWSER_ENV) {\n            // TODO use `resample` in browser environment\n\n            // Store number of channels before resizing\n            let numChannels = this.channels;\n\n            // Create canvas object for this image\n            let canvas = this.toCanvas();\n\n            // Actually perform resizing using the canvas API\n            const ctx = createCanvasFunction(width, height).getContext('2d');\n\n            // Draw image to context, resizing in the process\n            ctx.drawImage(canvas, 0, 0, width, height);\n\n            // Create image from the resized data\n            let resizedImage = new RawImage(ctx.getImageData(0, 0, width, height).data, width, height, 4);\n\n            // Convert back so that image has the same number of channels as before\n            return resizedImage.convert(numChannels);\n\n        } else {\n            // Create sharp image from raw data, and resize\n            let img = this.toSharp();\n\n            switch (resampleMethod) {\n                case 'box':\n                case 'hamming':\n                    if (resampleMethod === 'box' || resampleMethod === 'hamming') {\n                        console.warn(`Resampling method ${resampleMethod} is not yet supported. Using bilinear instead.`);\n                        resampleMethod = 'bilinear';\n                    }\n\n                case 'nearest':\n                case 'bilinear':\n                case 'bicubic':\n                    // Perform resizing using affine transform. \n                    // This matches how the python Pillow library does it.\n                    img = img.affine([width / this.width, 0, 0, height / this.height], {\n                        interpolator: resampleMethod\n                    });\n                    break;\n\n                case 'lanczos':\n                    // https://github.com/python-pillow/Pillow/discussions/5519\n                    // https://github.com/lovell/sharp/blob/main/docs/api-resize.md\n                    img = img.resize({\n                        width, height,\n                        fit: 'fill',\n                        kernel: 'lanczos3', // PIL Lanczos uses a kernel size of 3 \n                    });\n                    break;\n\n                default:\n                    throw new Error(`Resampling method ${resampleMethod} is not supported.`);\n            }\n\n            return await loadImageFunction(img);\n        }\n\n    }\n\n    async pad([left, right, top, bottom]) {\n        left = Math.max(left, 0);\n        right = Math.max(right, 0);\n        top = Math.max(top, 0);\n        bottom = Math.max(bottom, 0);\n\n        if (left === 0 && right === 0 && top === 0 && bottom === 0) {\n            // No padding needed\n            return this;\n        }\n\n        if (BROWSER_ENV) {\n            // Store number of channels before padding\n            let numChannels = this.channels;\n\n            // Create canvas object for this image\n            let canvas = this.toCanvas();\n\n            let newWidth = this.width + left + right;\n            let newHeight = this.height + top + bottom;\n\n            // Create a new canvas of the desired size.\n            const ctx = createCanvasFunction(newWidth, newHeight).getContext('2d');\n\n            // Draw image to context, padding in the process\n            ctx.drawImage(canvas,\n                0, 0, this.width, this.height,\n                left, top, newWidth, newHeight\n            );\n\n            // Create image from the padded data\n            let paddedImage = new RawImage(\n                ctx.getImageData(0, 0, newWidth, newHeight).data,\n                newWidth, newHeight, 4);\n\n            // Convert back so that image has the same number of channels as before\n            return paddedImage.convert(numChannels);\n\n        } else {\n            let img = this.toSharp().extend({ left, right, top, bottom });\n            return await loadImageFunction(img);\n        }\n    }\n\n    async crop([x_min, y_min, x_max, y_max]) {\n        // Ensure crop bounds are within the image\n        x_min = Math.max(x_min, 0);\n        y_min = Math.max(y_min, 0);\n        x_max = Math.min(x_max, this.width - 1);\n        y_max = Math.min(y_max, this.height - 1);\n\n        // Do nothing if the crop is the entire image\n        if (x_min === 0 && y_min === 0 && x_max === this.width - 1 && y_max === this.height - 1) {\n            return this;\n        }\n\n        const crop_width = x_max - x_min + 1;\n        const crop_height = y_max - y_min + 1;\n\n        if (BROWSER_ENV) {\n            // Store number of channels before resizing\n            const numChannels = this.channels;\n\n            // Create canvas object for this image\n            const canvas = this.toCanvas();\n\n            // Create a new canvas of the desired size. This is needed since if the \n            // image is too small, we need to pad it with black pixels.\n            const ctx = createCanvasFunction(crop_width, crop_height).getContext('2d');\n\n            // Draw image to context, cropping in the process\n            ctx.drawImage(canvas,\n                x_min, y_min, crop_width, crop_height,\n                0, 0, crop_width, crop_height\n            );\n\n            // Create image from the resized data\n            const resizedImage = new RawImage(ctx.getImageData(0, 0, crop_width, crop_height).data, crop_width, crop_height, 4);\n\n            // Convert back so that image has the same number of channels as before\n            return resizedImage.convert(numChannels);\n\n        } else {\n            // Create sharp image from raw data\n            const img = this.toSharp().extract({\n                left: x_min,\n                top: y_min,\n                width: crop_width,\n                height: crop_height,\n            });\n\n            return await loadImageFunction(img);\n        }\n\n    }\n\n    async center_crop(crop_width, crop_height) {\n        // If the image is already the desired size, return it\n        if (this.width === crop_width && this.height === crop_height) {\n            return this;\n        }\n\n        // Determine bounds of the image in the new canvas\n        let width_offset = (this.width - crop_width) / 2;\n        let height_offset = (this.height - crop_height) / 2;\n\n\n        if (BROWSER_ENV) {\n            // Store number of channels before resizing\n            let numChannels = this.channels;\n\n            // Create canvas object for this image\n            let canvas = this.toCanvas();\n\n            // Create a new canvas of the desired size. This is needed since if the \n            // image is too small, we need to pad it with black pixels.\n            const ctx = createCanvasFunction(crop_width, crop_height).getContext('2d');\n\n            let sourceX = 0;\n            let sourceY = 0;\n            let destX = 0;\n            let destY = 0;\n\n            if (width_offset >= 0) {\n                sourceX = width_offset;\n            } else {\n                destX = -width_offset;\n            }\n\n            if (height_offset >= 0) {\n                sourceY = height_offset;\n            } else {\n                destY = -height_offset;\n            }\n\n            // Draw image to context, cropping in the process\n            ctx.drawImage(canvas,\n                sourceX, sourceY, crop_width, crop_height,\n                destX, destY, crop_width, crop_height\n            );\n\n            // Create image from the resized data\n            let resizedImage = new RawImage(ctx.getImageData(0, 0, crop_width, crop_height).data, crop_width, crop_height, 4);\n\n            // Convert back so that image has the same number of channels as before\n            return resizedImage.convert(numChannels);\n\n        } else {\n            // Create sharp image from raw data\n            let img = this.toSharp();\n\n            if (width_offset >= 0 && height_offset >= 0) {\n                // Cropped image lies entirely within the original image\n                img = img.extract({\n                    left: Math.floor(width_offset),\n                    top: Math.floor(height_offset),\n                    width: crop_width,\n                    height: crop_height,\n                })\n            } else if (width_offset <= 0 && height_offset <= 0) {\n                // Cropped image lies entirely outside the original image,\n                // so we add padding\n                let top = Math.floor(-height_offset);\n                let left = Math.floor(-width_offset);\n                img = img.extend({\n                    top: top,\n                    left: left,\n\n                    // Ensures the resulting image has the desired dimensions\n                    right: crop_width - this.width - left,\n                    bottom: crop_height - this.height - top,\n                });\n            } else {\n                // Cropped image lies partially outside the original image.\n                // We first pad, then crop.\n\n                let y_padding = [0, 0];\n                let y_extract = 0;\n                if (height_offset < 0) {\n                    y_padding[0] = Math.floor(-height_offset);\n                    y_padding[1] = crop_height - this.height - y_padding[0];\n                } else {\n                    y_extract = Math.floor(height_offset);\n                }\n\n                let x_padding = [0, 0];\n                let x_extract = 0;\n                if (width_offset < 0) {\n                    x_padding[0] = Math.floor(-width_offset);\n                    x_padding[1] = crop_width - this.width - x_padding[0];\n                } else {\n                    x_extract = Math.floor(width_offset);\n                }\n\n                img = img.extend({\n                    top: y_padding[0],\n                    bottom: y_padding[1],\n                    left: x_padding[0],\n                    right: x_padding[1],\n                }).extract({\n                    left: x_extract,\n                    top: y_extract,\n                    width: crop_width,\n                    height: crop_height,\n                })\n            }\n\n            return await loadImageFunction(img);\n        }\n    }\n\n    async toBlob(type = 'image/png', quality = 1) {\n        if (!BROWSER_ENV) {\n            throw new Error('toBlob() is only supported in browser environments.')\n        }\n\n        const canvas = this.toCanvas();\n        return await canvas.convertToBlob({ type, quality });\n    }\n\n    toTensor(channel_format = 'CHW') {\n        let tensor = new Tensor(\n            'uint8',\n            new Uint8Array(this.data),\n            [this.height, this.width, this.channels]\n        );\n\n        if (channel_format === 'HWC') {\n            // Do nothing\n        } else if (channel_format === 'CHW') { // hwc -> chw\n            tensor = tensor.permute(2, 0, 1);\n        } else {\n            throw new Error(`Unsupported channel format: ${channel_format}`);\n        }\n        return tensor;\n    }\n\n    toCanvas() {\n        if (!BROWSER_ENV) {\n            throw new Error('toCanvas() is only supported in browser environments.')\n        }\n\n        // Clone, and convert data to RGBA before drawing to canvas.\n        // This is because the canvas API only supports RGBA\n        let cloned = this.clone().rgba();\n\n        // Create canvas object for the cloned image\n        let clonedCanvas = createCanvasFunction(cloned.width, cloned.height);\n\n        // Draw image to context\n        let data = new ImageDataClass(cloned.data, cloned.width, cloned.height);\n        clonedCanvas.getContext('2d').putImageData(data, 0, 0);\n\n        return clonedCanvas;\n    }\n\n    /**\n     * Helper method to update the image data.\n     * @param {Uint8ClampedArray} data The new image data.\n     * @param {number} width The new width of the image.\n     * @param {number} height The new height of the image.\n     * @param {1|2|3|4|null} [channels] The new number of channels of the image.\n     * @private\n     */\n    _update(data, width, height, channels = null) {\n        this.data = data;\n        this.width = width;\n        this.height = height;\n        if (channels !== null) {\n            this.channels = channels;\n        }\n        return this;\n    }\n\n    /**\n     * Clone the image\n     * @returns {RawImage} The cloned image\n     */\n    clone() {\n        return new RawImage(this.data.slice(), this.width, this.height, this.channels);\n    }\n\n    /**\n     * Helper method for converting image to have a certain number of channels\n     * @param {number} numChannels The number of channels. Must be 1, 3, or 4.\n     * @returns {RawImage} `this` to support chaining.\n     */\n    convert(numChannels) {\n        if (this.channels === numChannels) return this; // Already correct number of channels\n\n        switch (numChannels) {\n            case 1:\n                this.grayscale();\n                break;\n            case 3:\n                this.rgb();\n                break;\n            case 4:\n                this.rgba();\n                break;\n            default:\n                throw new Error(`Conversion failed due to unsupported number of channels: ${this.channels}`);\n        }\n        return this;\n    }\n\n    /**\n     * Save the image to the given path.\n     * @param {string} path The path to save the image to.\n     */\n    async save(path) {\n\n        if (BROWSER_ENV) {\n            if (WEBWORKER_ENV) {\n                throw new Error('Unable to save an image from a Web Worker.')\n            }\n\n            const extension = path.split('.').pop().toLowerCase();\n            const mime = CONTENT_TYPE_MAP.get(extension) ?? 'image/png';\n\n            // Convert image to Blob\n            const blob = await this.toBlob(mime);\n\n            // Convert the canvas content to a data URL\n            const dataURL = URL.createObjectURL(blob);\n\n            // Create an anchor element with the data URL as the href attribute\n            const downloadLink = document.createElement('a');\n            downloadLink.href = dataURL;\n\n            // Set the download attribute to specify the desired filename for the downloaded image\n            downloadLink.download = path;\n\n            // Trigger the download\n            downloadLink.click();\n\n            // Clean up: remove the anchor element from the DOM\n            downloadLink.remove();\n\n        } else if (!env.useFS) {\n            throw new Error('Unable to save the image because filesystem is disabled in this environment.')\n\n        } else {\n            const img = this.toSharp();\n            return await img.toFile(path);\n        }\n    }\n\n    toSharp() {\n        if (BROWSER_ENV) {\n            throw new Error('toSharp() is only supported in server-side environments.')\n        }\n\n        return sharp(this.data, {\n            raw: {\n                width: this.width,\n                height: this.height,\n                channels: this.channels\n            }\n        });\n    }\n}\n"],"names":[],"mappings":";;;;;AAiBA,MAAM,cAAc,OAAO,SAAS;AAMpC,IAAI,YAAa;AAAA,SAWN,MAAO;AAAA,KAkBX;AACH,QAAM,IAAI,MAAM,0CAA0C;AAC9D;","x_google_ignoreList":[0]}

package/dist/node_modules/@xenova/transformers/src/utils/maths.js

@@ -0,0 +1,89 @@
+function permute_data(array, dims, axes) {
+  const shape = new Array(axes.length);
+  const stride = new Array(axes.length);
+  for (let i = axes.length - 1, s = 1; i >= 0; --i) {
+    stride[i] = s;
+    shape[i] = dims[axes[i]];
+    s *= shape[i];
+  }
+  const invStride = axes.map((_, i) => stride[axes.indexOf(i)]);
+  const permutedData = new array.constructor(array.length);
+  for (let i = 0; i < array.length; ++i) {
+    let newIndex = 0;
+    for (let j = dims.length - 1, k = i; j >= 0; --j) {
+      newIndex += k % dims[j] * invStride[j];
+      k = Math.floor(k / dims[j]);
+    }
+    permutedData[newIndex] = array[i];
+  }
+  return [permutedData, shape];
+}
+function softmax(arr) {
+  const maxVal = max(arr)[0];
+  const exps = arr.map((x) => Math.exp(x - maxVal));
+  const sumExps = exps.reduce((acc, val) => acc + val, 0);
+  const softmaxArr = exps.map((x) => x / sumExps);
+  return (
+    /** @type {T} */
+    softmaxArr
+  );
+}
+function log_softmax(arr) {
+  const softmaxArr = softmax(arr);
+  const logSoftmaxArr = softmaxArr.map((x) => Math.log(x));
+  return (
+    /** @type {T} */
+    logSoftmaxArr
+  );
+}
+function getTopItems(items, top_k = 0) {
+  items = Array.from(items).map((x, i) => [i, x]).sort((a, b) => b[1] - a[1]);
+  if (top_k !== null && top_k > 0) {
+    items = items.slice(0, top_k);
+  }
+  return items;
+}
+function max(arr) {
+  if (arr.length === 0) throw Error("Array must not be empty");
+  let max2 = arr[0];
+  let indexOfMax = 0;
+  for (let i = 1; i < arr.length; ++i) {
+    if (arr[i] > max2) {
+      max2 = arr[i];
+      indexOfMax = i;
+    }
+  }
+  return [Number(max2), indexOfMax];
+}
+function medianFilter(data, windowSize) {
+  if (windowSize % 2 === 0 || windowSize <= 0) {
+    throw new Error("Window size must be a positive odd number");
+  }
+  const outputArray = new data.constructor(data.length);
+  const buffer = new data.constructor(windowSize);
+  const halfWindowSize = Math.floor(windowSize / 2);
+  for (let i = 0; i < data.length; ++i) {
+    let valuesIndex = 0;
+    for (let j = -halfWindowSize; j <= halfWindowSize; ++j) {
+      let index = i + j;
+      if (index < 0) {
+        index = Math.abs(index);
+      } else if (index >= data.length) {
+        index = 2 * (data.length - 1) - index;
+      }
+      buffer[valuesIndex++] = data[index];
+    }
+    buffer.sort();
+    outputArray[i] = buffer[halfWindowSize];
+  }
+  return outputArray;
+}
+export {
+  getTopItems,
+  log_softmax,
+  max,
+  medianFilter,
+  permute_data,
+  softmax
+};
+//# sourceMappingURL=maths.js.map

package/dist/node_modules/@xenova/transformers/src/utils/maths.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"maths.js","sources":["../../../../../../node_modules/@xenova/transformers/src/utils/maths.js"],"sourcesContent":["\n/**\n * @file Helper module for mathematical processing. \n * \n * These functions and classes are only used internally, \n * meaning an end-user shouldn't need to access anything here.\n * \n * @module utils/maths\n */\n\n/**\n * @typedef {Int8Array | Uint8Array | Uint8ClampedArray | Int16Array | Uint16Array | Int32Array | Uint32Array | Float32Array | Float64Array} TypedArray\n * @typedef {BigInt64Array | BigUint64Array} BigTypedArray\n * @typedef {TypedArray | BigTypedArray} AnyTypedArray\n */\n\n/**\n * @param {TypedArray} input\n */\nexport function interpolate_data(input, [in_channels, in_height, in_width], [out_height, out_width], mode = 'bilinear', align_corners = false) {\n    // TODO use mode and align_corners\n\n    // Output image dimensions\n    const x_scale = out_width / in_width;\n    const y_scale = out_height / in_height;\n\n    // Output image\n    // @ts-ignore\n    const out_img = new input.constructor(out_height * out_width * in_channels);\n\n    // Pre-calculate strides\n    const inStride = in_height * in_width;\n    const outStride = out_height * out_width;\n\n    for (let i = 0; i < out_height; ++i) {\n        for (let j = 0; j < out_width; ++j) {\n            // Calculate output offset\n            const outOffset = i * out_width + j;\n\n            // Calculate input pixel coordinates\n            const x = (j + 0.5) / x_scale - 0.5;\n            const y = (i + 0.5) / y_scale - 0.5;\n\n            // Calculate the four nearest input pixels\n            // We also check if the input pixel coordinates are within the image bounds\n            let x1 = Math.floor(x);\n            let y1 = Math.floor(y);\n            const x2 = Math.min(x1 + 1, in_width - 1);\n            const y2 = Math.min(y1 + 1, in_height - 1);\n\n            x1 = Math.max(x1, 0);\n            y1 = Math.max(y1, 0);\n\n\n            // Calculate the fractional distances between the input pixel and the four nearest pixels\n            const s = x - x1;\n            const t = y - y1;\n\n            // Perform bilinear interpolation\n            const w1 = (1 - s) * (1 - t);\n            const w2 = s * (1 - t);\n            const w3 = (1 - s) * t;\n            const w4 = s * t;\n\n            // Calculate the four nearest input pixel indices\n            const yStride = y1 * in_width;\n            const xStride = y2 * in_width;\n            const idx1 = yStride + x1;\n            const idx2 = yStride + x2;\n            const idx3 = xStride + x1;\n            const idx4 = xStride + x2;\n\n            for (let k = 0; k < in_channels; ++k) {\n                // Calculate channel offset\n                const cOffset = k * inStride;\n\n                out_img[k * outStride + outOffset] =\n                    w1 * input[cOffset + idx1] +\n                    w2 * input[cOffset + idx2] +\n                    w3 * input[cOffset + idx3] +\n                    w4 * input[cOffset + idx4];\n            }\n        }\n    }\n\n    return out_img;\n}\n\n\n/**\n * Helper method to permute a `AnyTypedArray` directly\n * @template {AnyTypedArray} T \n * @param {T} array \n * @param {number[]} dims \n * @param {number[]} axes \n * @returns {[T, number[]]} The permuted array and the new shape.\n */\nexport function permute_data(array, dims, axes) {\n    // Calculate the new shape of the permuted array\n    // and the stride of the original array\n    const shape = new Array(axes.length);\n    const stride = new Array(axes.length);\n\n    for (let i = axes.length - 1, s = 1; i >= 0; --i) {\n        stride[i] = s;\n        shape[i] = dims[axes[i]];\n        s *= shape[i];\n    }\n\n    // Precompute inverse mapping of stride\n    const invStride = axes.map((_, i) => stride[axes.indexOf(i)]);\n\n    // Create the permuted array with the new shape\n    // @ts-ignore\n    const permutedData = new array.constructor(array.length);\n\n    // Permute the original array to the new array\n    for (let i = 0; i < array.length; ++i) {\n        let newIndex = 0;\n        for (let j = dims.length - 1, k = i; j >= 0; --j) {\n            newIndex += (k % dims[j]) * invStride[j];\n            k = Math.floor(k / dims[j]);\n        }\n        permutedData[newIndex] = array[i];\n    }\n\n    return [permutedData, shape];\n}\n\n\n/**\n * Compute the softmax of an array of numbers.\n * @template {TypedArray|number[]} T\n * @param {T} arr The array of numbers to compute the softmax of.\n * @returns {T} The softmax array.\n */\nexport function softmax(arr) {\n    // Compute the maximum value in the array\n    const maxVal = max(arr)[0];\n\n    // Compute the exponentials of the array values\n    const exps = arr.map(x => Math.exp(x - maxVal));\n\n    // Compute the sum of the exponentials\n    // @ts-ignore\n    const sumExps = exps.reduce((acc, val) => acc + val, 0);\n\n    // Compute the softmax values\n    const softmaxArr = exps.map(x => x / sumExps);\n\n    return /** @type {T} */(softmaxArr);\n}\n\n/**\n * Calculates the logarithm of the softmax function for the input array.\n * @template {TypedArray|number[]} T\n * @param {T} arr The input array to calculate the log_softmax function for.\n * @returns {T} The resulting log_softmax array.\n */\nexport function log_softmax(arr) {\n    // Compute the softmax values\n    const softmaxArr = softmax(arr);\n\n    // Apply log formula to each element\n    const logSoftmaxArr = softmaxArr.map(x => Math.log(x));\n\n    return /** @type {T} */(logSoftmaxArr);\n}\n\n/**\n * Calculates the dot product of two arrays.\n * @param {number[]} arr1 The first array.\n * @param {number[]} arr2 The second array.\n * @returns {number} The dot product of arr1 and arr2.\n */\nexport function dot(arr1, arr2) {\n    let result = 0;\n    for (let i = 0; i < arr1.length; ++i) {\n        result += arr1[i] * arr2[i];\n    }\n    return result;\n}\n\n\n/**\n * Get the top k items from an iterable, sorted by descending order\n * @param {any[]|TypedArray} items The items to be sorted\n * @param {number|null} [top_k=0] The number of top items to return (default: 0 = return all)\n * @returns {[number, any][]} The top k items, sorted by descending order\n */\nexport function getTopItems(items, top_k = 0) {\n    // if top == 0, return all\n\n    items = Array.from(items)\n        .map((x, i) => [i, x])            // Get indices ([index, score])\n        .sort((a, b) => b[1] - a[1])      // Sort by log probabilities\n\n    if (top_k !== null && top_k > 0) {\n        items = items.slice(0, top_k);    // Get top k items\n    }\n\n    return items\n}\n\n/**\n * Computes the cosine similarity between two arrays.\n *\n * @param {number[]} arr1 The first array.\n * @param {number[]} arr2 The second array.\n * @returns {number} The cosine similarity between the two arrays.\n */\nexport function cos_sim(arr1, arr2) {\n    // Calculate dot product of the two arrays\n    const dotProduct = dot(arr1, arr2);\n\n    // Calculate the magnitude of the first array\n    const magnitudeA = magnitude(arr1);\n\n    // Calculate the magnitude of the second array\n    const magnitudeB = magnitude(arr2);\n\n    // Calculate the cosine similarity\n    const cosineSimilarity = dotProduct / (magnitudeA * magnitudeB);\n\n    return cosineSimilarity;\n}\n\n/**\n * Calculates the magnitude of a given array.\n * @param {number[]} arr The array to calculate the magnitude of.\n * @returns {number} The magnitude of the array.\n */\nexport function magnitude(arr) {\n    return Math.sqrt(arr.reduce((acc, val) => acc + val * val, 0));\n}\n\n\n/**\n * Returns the value and index of the minimum element in an array.\n * @param {number[]|TypedArray} arr array of numbers.\n * @returns {number[]} the value and index of the minimum element, of the form: [valueOfMin, indexOfMin]\n * @throws {Error} If array is empty.\n */\nexport function min(arr) {\n    if (arr.length === 0) throw Error('Array must not be empty');\n    let min = arr[0];\n    let indexOfMin = 0;\n    for (let i = 1; i < arr.length; ++i) {\n        if (arr[i] < min) {\n            min = arr[i];\n            indexOfMin = i;\n        }\n    }\n    return [min, indexOfMin];\n}\n\n\n/**\n * Returns the value and index of the maximum element in an array.\n * @param {number[]|AnyTypedArray} arr array of numbers.\n * @returns {[number, number]} the value and index of the maximum element, of the form: [valueOfMax, indexOfMax]\n * @throws {Error} If array is empty.\n */\nexport function max(arr) {\n    if (arr.length === 0) throw Error('Array must not be empty');\n    let max = arr[0];\n    let indexOfMax = 0;\n    for (let i = 1; i < arr.length; ++i) {\n        if (arr[i] > max) {\n            max = arr[i];\n            indexOfMax = i;\n        }\n    }\n    return [Number(max), indexOfMax];\n}\n\nfunction isPowerOfTwo(number) {\n    // Check if the number is greater than 0 and has only one bit set to 1\n    return (number > 0) && ((number & (number - 1)) === 0);\n}\n\n/**\n * Implementation of Radix-4 FFT.\n * \n * P2FFT class provides functionality for performing Fast Fourier Transform on arrays\n * which are a power of two in length.\n * Code adapted from https://www.npmjs.com/package/fft.js\n */\nclass P2FFT {\n    /**\n     * @param {number} size The size of the input array. Must be a power of two larger than 1.\n     * @throws {Error} FFT size must be a power of two larger than 1.\n     */\n    constructor(size) {\n        this.size = size | 0; // convert to a 32-bit signed integer\n        if (this.size <= 1 || !isPowerOfTwo(this.size))\n            throw new Error('FFT size must be a power of two larger than 1');\n\n        this._csize = size << 1;\n\n        this.table = new Float64Array(this.size * 2);\n        for (let i = 0; i < this.table.length; i += 2) {\n            const angle = Math.PI * i / this.size;\n            this.table[i] = Math.cos(angle);\n            this.table[i + 1] = -Math.sin(angle);\n        }\n\n        // Find size's power of two\n        let power = 0;\n        for (let t = 1; this.size > t; t <<= 1)\n            ++power;\n\n        // Calculate initial step's width:\n        //   * If we are full radix-4, it is 2x smaller to give inital len=8\n        //   * Otherwise it is the same as `power` to give len=4\n        this._width = power % 2 === 0 ? power - 1 : power;\n\n        // Pre-compute bit-reversal patterns\n        this._bitrev = new Int32Array(1 << this._width);\n        for (let j = 0; j < this._bitrev.length; ++j) {\n            this._bitrev[j] = 0;\n            for (let shift = 0; shift < this._width; shift += 2) {\n                const revShift = this._width - shift - 2;\n                this._bitrev[j] |= ((j >>> shift) & 3) << revShift;\n            }\n        }\n    }\n\n    /**\n     * Create a complex number array with size `2 * size`\n     *\n     * @returns {Float64Array} A complex number array with size `2 * size`\n     */\n    createComplexArray() {\n        return new Float64Array(this._csize);\n    }\n\n    /**\n     * Converts a complex number representation stored in a Float64Array to an array of real numbers.\n     * \n     * @param {Float64Array} complex The complex number representation to be converted.\n     * @param {number[]} [storage] An optional array to store the result in.\n     * @returns {number[]} An array of real numbers representing the input complex number representation.\n     */\n    fromComplexArray(complex, storage) {\n        const res = storage || new Array(complex.length >>> 1);\n        for (let i = 0; i < complex.length; i += 2)\n            res[i >>> 1] = complex[i];\n        return res;\n    }\n\n    /**\n     * Convert a real-valued input array to a complex-valued output array.\n     * @param {Float64Array} input The real-valued input array.\n     * @param {Float64Array} [storage] Optional buffer to store the output array.\n     * @returns {Float64Array} The complex-valued output array.\n     */\n    toComplexArray(input, storage) {\n        const res = storage || this.createComplexArray();\n        for (let i = 0; i < res.length; i += 2) {\n            res[i] = input[i >>> 1];\n            res[i + 1] = 0;\n        }\n        return res;\n    }\n\n    /**\n     * Performs a Fast Fourier Transform (FFT) on the given input data and stores the result in the output buffer.\n     * \n     * @param {Float64Array} out The output buffer to store the result.\n     * @param {Float64Array} data The input data to transform.\n     * \n     * @throws {Error} Input and output buffers must be different.\n     * \n     * @returns {void}\n     */\n    transform(out, data) {\n        if (out === data)\n            throw new Error('Input and output buffers must be different');\n\n        this._transform4(out, data, 1 /* DONE */);\n    }\n\n    /**\n     * Performs a real-valued forward FFT on the given input buffer and stores the result in the given output buffer.\n     * The input buffer must contain real values only, while the output buffer will contain complex values. The input and\n     * output buffers must be different.\n     *\n     * @param {Float64Array} out The output buffer.\n     * @param {Float64Array} data The input buffer containing real values.\n     *\n     * @throws {Error} If the input and output buffers are the same.\n     */\n    realTransform(out, data) {\n        if (out === data)\n            throw new Error('Input and output buffers must be different');\n\n        this._realTransform4(out, data, 1 /* DONE */);\n    }\n\n    /**\n     * Performs an inverse FFT transformation on the given `data` array, and stores the result in `out`.\n     * The `out` array must be a different buffer than the `data` array. The `out` array will contain the\n     * result of the transformation. The `data` array will not be modified.\n     * \n     * @param {Float64Array} out The output buffer for the transformed data.\n     * @param {Float64Array} data The input data to transform.\n     * @throws {Error} If `out` and `data` refer to the same buffer.\n     * @returns {void}\n     */\n    inverseTransform(out, data) {\n        if (out === data)\n            throw new Error('Input and output buffers must be different');\n\n        this._transform4(out, data, -1 /* DONE */);\n        for (let i = 0; i < out.length; ++i)\n            out[i] /= this.size;\n    }\n\n    /**\n     * Performs a radix-4 implementation of a discrete Fourier transform on a given set of data.\n     *\n     * @param {Float64Array} out The output buffer for the transformed data.\n     * @param {Float64Array} data The input buffer of data to be transformed.\n     * @param {number} inv A scaling factor to apply to the transform.\n     * @returns {void}\n     */\n    _transform4(out, data, inv) {\n        // radix-4 implementation\n\n        const size = this._csize;\n\n        // Initial step (permute and transform)\n        const width = this._width;\n        let step = 1 << width;\n        let len = (size / step) << 1;\n\n        let outOff;\n        let t;\n        const bitrev = this._bitrev;\n        if (len === 4) {\n            for (outOff = 0, t = 0; outOff < size; outOff += len, ++t) {\n                const off = bitrev[t];\n                this._singleTransform2(data, out, outOff, off, step);\n            }\n        } else {\n            // len === 8\n            for (outOff = 0, t = 0; outOff < size; outOff += len, ++t) {\n                const off = bitrev[t];\n                this._singleTransform4(data, out, outOff, off, step, inv);\n            }\n        }\n\n        // Loop through steps in decreasing order\n        const table = this.table;\n        for (step >>= 2; step >= 2; step >>= 2) {\n            len = (size / step) << 1;\n            const quarterLen = len >>> 2;\n\n            // Loop through offsets in the data\n            for (outOff = 0; outOff < size; outOff += len) {\n                // Full case\n                const limit = outOff + quarterLen - 1;\n                for (let i = outOff, k = 0; i < limit; i += 2, k += step) {\n                    const A = i;\n                    const B = A + quarterLen;\n                    const C = B + quarterLen;\n                    const D = C + quarterLen;\n\n                    // Original values\n                    const Ar = out[A];\n                    const Ai = out[A + 1];\n                    const Br = out[B];\n                    const Bi = out[B + 1];\n                    const Cr = out[C];\n                    const Ci = out[C + 1];\n                    const Dr = out[D];\n                    const Di = out[D + 1];\n\n                    const tableBr = table[k];\n                    const tableBi = inv * table[k + 1];\n                    const MBr = Br * tableBr - Bi * tableBi;\n                    const MBi = Br * tableBi + Bi * tableBr;\n\n                    const tableCr = table[2 * k];\n                    const tableCi = inv * table[2 * k + 1];\n                    const MCr = Cr * tableCr - Ci * tableCi;\n                    const MCi = Cr * tableCi + Ci * tableCr;\n\n                    const tableDr = table[3 * k];\n                    const tableDi = inv * table[3 * k + 1];\n                    const MDr = Dr * tableDr - Di * tableDi;\n                    const MDi = Dr * tableDi + Di * tableDr;\n\n                    // Pre-Final values\n                    const T0r = Ar + MCr;\n                    const T0i = Ai + MCi;\n                    const T1r = Ar - MCr;\n                    const T1i = Ai - MCi;\n                    const T2r = MBr + MDr;\n                    const T2i = MBi + MDi;\n                    const T3r = inv * (MBr - MDr);\n                    const T3i = inv * (MBi - MDi);\n\n                    // Final values\n                    out[A] = T0r + T2r;\n                    out[A + 1] = T0i + T2i;\n                    out[B] = T1r + T3i;\n                    out[B + 1] = T1i - T3r;\n                    out[C] = T0r - T2r;\n                    out[C + 1] = T0i - T2i;\n                    out[D] = T1r - T3i;\n                    out[D + 1] = T1i + T3r;\n                }\n            }\n        }\n    }\n\n    /**\n     * Performs a radix-2 implementation of a discrete Fourier transform on a given set of data.\n     *\n     * @param {Float64Array} data The input buffer of data to be transformed.\n     * @param {Float64Array} out The output buffer for the transformed data.\n     * @param {number} outOff The offset at which to write the output data.\n     * @param {number} off The offset at which to begin reading the input data.\n     * @param {number} step The step size for indexing the input data.\n     * @returns {void}\n     */\n    _singleTransform2(data, out, outOff, off, step) {\n        // radix-2 implementation\n        // NOTE: Only called for len=4\n\n        const evenR = data[off];\n        const evenI = data[off + 1];\n        const oddR = data[off + step];\n        const oddI = data[off + step + 1];\n\n        out[outOff] = evenR + oddR;\n        out[outOff + 1] = evenI + oddI;\n        out[outOff + 2] = evenR - oddR;\n        out[outOff + 3] = evenI - oddI;\n    }\n\n    /**\n     * Performs radix-4 transformation on input data of length 8\n     *\n     * @param {Float64Array} data Input data array of length 8\n     * @param {Float64Array} out Output data array of length 8\n     * @param {number} outOff Index of output array to start writing from\n     * @param {number} off Index of input array to start reading from\n     * @param {number} step Step size between elements in input array\n     * @param {number} inv Scaling factor for inverse transform\n     * \n     * @returns {void}\n     */\n    _singleTransform4(data, out, outOff, off, step, inv) {\n        // radix-4\n        // NOTE: Only called for len=8\n        const step2 = step * 2;\n        const step3 = step * 3;\n\n        // Original values\n        const Ar = data[off];\n        const Ai = data[off + 1];\n        const Br = data[off + step];\n        const Bi = data[off + step + 1];\n        const Cr = data[off + step2];\n        const Ci = data[off + step2 + 1];\n        const Dr = data[off + step3];\n        const Di = data[off + step3 + 1];\n\n        // Pre-Final values\n        const T0r = Ar + Cr;\n        const T0i = Ai + Ci;\n        const T1r = Ar - Cr;\n        const T1i = Ai - Ci;\n        const T2r = Br + Dr;\n        const T2i = Bi + Di;\n        const T3r = inv * (Br - Dr);\n        const T3i = inv * (Bi - Di);\n\n        // Final values\n        out[outOff] = T0r + T2r;\n        out[outOff + 1] = T0i + T2i;\n        out[outOff + 2] = T1r + T3i;\n        out[outOff + 3] = T1i - T3r;\n        out[outOff + 4] = T0r - T2r;\n        out[outOff + 5] = T0i - T2i;\n        out[outOff + 6] = T1r - T3i;\n        out[outOff + 7] = T1i + T3r;\n    }\n\n    /**\n     * Real input radix-4 implementation\n     * @param {Float64Array} out Output array for the transformed data\n     * @param {Float64Array} data Input array of real data to be transformed\n     * @param {number} inv The scale factor used to normalize the inverse transform\n     */\n    _realTransform4(out, data, inv) {\n        // Real input radix-4 implementation\n        const size = this._csize;\n\n        // Initial step (permute and transform)\n        const width = this._width;\n        let step = 1 << width;\n        let len = (size / step) << 1;\n\n        let outOff;\n        let t;\n        const bitrev = this._bitrev;\n        if (len === 4) {\n            for (outOff = 0, t = 0; outOff < size; outOff += len, ++t) {\n                const off = bitrev[t];\n                this._singleRealTransform2(data, out, outOff, off >>> 1, step >>> 1);\n            }\n        } else {\n            // len === 8\n            for (outOff = 0, t = 0; outOff < size; outOff += len, ++t) {\n                const off = bitrev[t];\n                this._singleRealTransform4(data, out, outOff, off >>> 1, step >>> 1, inv);\n            }\n        }\n\n        // Loop through steps in decreasing order\n        const table = this.table;\n        for (step >>= 2; step >= 2; step >>= 2) {\n            len = (size / step) << 1;\n            const halfLen = len >>> 1;\n            const quarterLen = halfLen >>> 1;\n            const hquarterLen = quarterLen >>> 1;\n\n            // Loop through offsets in the data\n            for (outOff = 0; outOff < size; outOff += len) {\n                for (let i = 0, k = 0; i <= hquarterLen; i += 2, k += step) {\n                    const A = outOff + i;\n                    const B = A + quarterLen;\n                    const C = B + quarterLen;\n                    const D = C + quarterLen;\n\n                    // Original values\n                    const Ar = out[A];\n                    const Ai = out[A + 1];\n                    const Br = out[B];\n                    const Bi = out[B + 1];\n                    const Cr = out[C];\n                    const Ci = out[C + 1];\n                    const Dr = out[D];\n                    const Di = out[D + 1];\n\n                    // Middle values\n                    const MAr = Ar;\n                    const MAi = Ai;\n\n                    const tableBr = table[k];\n                    const tableBi = inv * table[k + 1];\n                    const MBr = Br * tableBr - Bi * tableBi;\n                    const MBi = Br * tableBi + Bi * tableBr;\n\n                    const tableCr = table[2 * k];\n                    const tableCi = inv * table[2 * k + 1];\n                    const MCr = Cr * tableCr - Ci * tableCi;\n                    const MCi = Cr * tableCi + Ci * tableCr;\n\n                    const tableDr = table[3 * k];\n                    const tableDi = inv * table[3 * k + 1];\n                    const MDr = Dr * tableDr - Di * tableDi;\n                    const MDi = Dr * tableDi + Di * tableDr;\n\n                    // Pre-Final values\n                    const T0r = MAr + MCr;\n                    const T0i = MAi + MCi;\n                    const T1r = MAr - MCr;\n                    const T1i = MAi - MCi;\n                    const T2r = MBr + MDr;\n                    const T2i = MBi + MDi;\n                    const T3r = inv * (MBr - MDr);\n                    const T3i = inv * (MBi - MDi);\n\n                    // Final values\n                    out[A] = T0r + T2r;\n                    out[A + 1] = T0i + T2i;\n                    out[B] = T1r + T3i;\n                    out[B + 1] = T1i - T3r;\n\n                    // Output final middle point\n                    if (i === 0) {\n                        out[C] = T0r - T2r;\n                        out[C + 1] = T0i - T2i;\n                        continue;\n                    }\n\n                    // Do not overwrite ourselves\n                    if (i === hquarterLen)\n                        continue;\n\n                    const SA = outOff + quarterLen - i;\n                    const SB = outOff + halfLen - i;\n\n                    out[SA] = T1r - inv * T3i;\n                    out[SA + 1] = -T1i - inv * T3r;\n                    out[SB] = T0r - inv * T2r;\n                    out[SB + 1] = -T0i + inv * T2i;\n                }\n            }\n        }\n\n        // Complete the spectrum by adding its mirrored negative frequency components.\n        const half = size >>> 1;\n        for (let i = 2; i < half; i += 2) {\n            out[size - i] = out[i];\n            out[size - i + 1] = -out[i + 1];\n        }\n    }\n\n    /**\n     * Performs a single real input radix-2 transformation on the provided data\n     * \n     * @param {Float64Array} data The input data array\n     * @param {Float64Array} out The output data array\n     * @param {number} outOff The output offset\n     * @param {number} off The input offset\n     * @param {number} step The step\n     * \n     * @returns {void}\n     */\n    _singleRealTransform2(data, out, outOff, off, step) {\n        // radix-2 implementation\n        // NOTE: Only called for len=4\n\n        const evenR = data[off];\n        const oddR = data[off + step];\n\n        out[outOff] = evenR + oddR;\n        out[outOff + 1] = 0;\n        out[outOff + 2] = evenR - oddR;\n        out[outOff + 3] = 0;\n    }\n\n    /**\n     * Computes a single real-valued transform using radix-4 algorithm.\n     * This method is only called for len=8.\n     *\n     * @param {Float64Array} data The input data array.\n     * @param {Float64Array} out The output data array.\n     * @param {number} outOff The offset into the output array.\n     * @param {number} off The offset into the input array.\n     * @param {number} step The step size for the input array.\n     * @param {number} inv The value of inverse.\n     */\n    _singleRealTransform4(data, out, outOff, off, step, inv) {\n        // radix-4\n        // NOTE: Only called for len=8\n        const step2 = step * 2;\n        const step3 = step * 3;\n\n        // Original values\n        const Ar = data[off];\n        const Br = data[off + step];\n        const Cr = data[off + step2];\n        const Dr = data[off + step3];\n\n        // Pre-Final values\n        const T0r = Ar + Cr;\n        const T1r = Ar - Cr;\n        const T2r = Br + Dr;\n        const T3r = inv * (Br - Dr);\n\n        // Final values\n        out[outOff] = T0r + T2r;\n        out[outOff + 1] = 0;\n        out[outOff + 2] = T1r;\n        out[outOff + 3] = -T3r;\n        out[outOff + 4] = T0r - T2r;\n        out[outOff + 5] = 0;\n        out[outOff + 6] = T1r;\n        out[outOff + 7] = T3r;\n    }\n}\n\n/**\n * NP2FFT class provides functionality for performing Fast Fourier Transform on arrays\n * which are not a power of two in length. In such cases, the chirp-z transform is used.\n * \n * For more information, see: https://math.stackexchange.com/questions/77118/non-power-of-2-ffts/77156#77156\n */\nclass NP2FFT {\n\n    /**\n     * Constructs a new NP2FFT object.\n     * @param {number} fft_length The length of the FFT\n     */\n    constructor(fft_length) {\n        // Helper variables\n        const a = 2 * (fft_length - 1);\n        const b = 2 * (2 * fft_length - 1);\n        const nextP2 = 2 ** (Math.ceil(Math.log2(b)))\n        this.bufferSize = nextP2;\n        this._a = a;\n\n        // Define buffers\n        // Compute chirp for transform\n        const chirp = new Float64Array(b);\n        const ichirp = new Float64Array(nextP2);\n        this._chirpBuffer = new Float64Array(nextP2);\n        this._buffer1 = new Float64Array(nextP2);\n        this._buffer2 = new Float64Array(nextP2);\n        this._outBuffer1 = new Float64Array(nextP2);\n        this._outBuffer2 = new Float64Array(nextP2);\n\n        // Compute complex exponentiation\n        const theta = -2 * Math.PI / fft_length;\n        const baseR = Math.cos(theta);\n        const baseI = Math.sin(theta);\n\n        // Precompute helper for chirp-z transform\n        for (let i = 0; i < b >> 1; ++i) {\n            // Compute complex power:\n            const e = (i + 1 - fft_length) ** 2 / 2.0;\n\n            // Compute the modulus and argument of the result\n            const result_mod = Math.sqrt(baseR ** 2 + baseI ** 2) ** e;\n            const result_arg = e * Math.atan2(baseI, baseR);\n\n            // Convert the result back to rectangular form\n            // and assign to chirp and ichirp\n            const i2 = 2 * i;\n            chirp[i2] = result_mod * Math.cos(result_arg);\n            chirp[i2 + 1] = result_mod * Math.sin(result_arg);\n\n            // conjugate\n            ichirp[i2] = chirp[i2];\n            ichirp[i2 + 1] = - chirp[i2 + 1];\n        }\n        this._slicedChirpBuffer = chirp.subarray(a, b);\n\n        // create object to perform Fast Fourier Transforms\n        // with `nextP2` complex numbers\n        this._f = new P2FFT(nextP2 >> 1);\n        this._f.transform(this._chirpBuffer, ichirp);\n    }\n\n    _transform(output, input, real) {\n        const ib1 = this._buffer1;\n        const ib2 = this._buffer2;\n        const ob2 = this._outBuffer1;\n        const ob3 = this._outBuffer2;\n        const cb = this._chirpBuffer;\n        const sb = this._slicedChirpBuffer;\n        const a = this._a;\n\n        if (real) {\n            // Real multiplication\n            for (let j = 0; j < sb.length; j += 2) {\n                const j2 = j + 1\n                const j3 = j >> 1;\n\n                const a_real = input[j3];\n                ib1[j] = a_real * sb[j];\n                ib1[j2] = a_real * sb[j2];\n            }\n        } else {\n            // Complex multiplication\n            for (let j = 0; j < sb.length; j += 2) {\n                const j2 = j + 1\n                ib1[j] = input[j] * sb[j] - input[j2] * sb[j2];\n                ib1[j2] = input[j] * sb[j2] + input[j2] * sb[j];\n            }\n        }\n        this._f.transform(ob2, ib1);\n\n        for (let j = 0; j < cb.length; j += 2) {\n            const j2 = j + 1;\n\n            ib2[j] = ob2[j] * cb[j] - ob2[j2] * cb[j2];\n            ib2[j2] = ob2[j] * cb[j2] + ob2[j2] * cb[j];\n        }\n        this._f.inverseTransform(ob3, ib2);\n\n        for (let j = 0; j < ob3.length; j += 2) {\n            const a_real = ob3[j + a];\n            const a_imag = ob3[j + a + 1];\n            const b_real = sb[j];\n            const b_imag = sb[j + 1];\n\n            output[j] = a_real * b_real - a_imag * b_imag;\n            output[j + 1] = a_real * b_imag + a_imag * b_real;\n        }\n    }\n\n    transform(output, input) {\n        this._transform(output, input, false);\n    }\n\n    realTransform(output, input) {\n        this._transform(output, input, true);\n    }\n}\n\nexport class FFT {\n    constructor(fft_length) {\n        this.fft_length = fft_length;\n        this.isPowerOfTwo = isPowerOfTwo(fft_length);\n        if (this.isPowerOfTwo) {\n            this.fft = new P2FFT(fft_length);\n            this.outputBufferSize = 2 * fft_length;\n        } else {\n            this.fft = new NP2FFT(fft_length);\n            this.outputBufferSize = this.fft.bufferSize;\n        }\n    }\n\n    realTransform(out, input) {\n        this.fft.realTransform(out, input);\n    }\n\n    transform(out, input) {\n        this.fft.transform(out, input);\n    }\n}\n\n\n/**\n * Performs median filter on the provided data. Padding is done by mirroring the data.\n * @param {AnyTypedArray} data The input array\n * @param {number} windowSize The window size\n */\nexport function medianFilter(data, windowSize) {\n\n    if (windowSize % 2 === 0 || windowSize <= 0) {\n        throw new Error('Window size must be a positive odd number');\n    }\n\n    // @ts-ignore\n    const outputArray = new data.constructor(data.length);\n\n    // @ts-ignore\n    const buffer = new data.constructor(windowSize); // Reusable array for storing values\n\n    const halfWindowSize = Math.floor(windowSize / 2);\n\n    for (let i = 0; i < data.length; ++i) {\n        let valuesIndex = 0;\n\n        for (let j = -halfWindowSize; j <= halfWindowSize; ++j) {\n            let index = i + j;\n            if (index < 0) {\n                index = Math.abs(index);\n            } else if (index >= data.length) {\n                index = 2 * (data.length - 1) - index;\n            }\n\n            buffer[valuesIndex++] = data[index];\n        }\n\n        buffer.sort();\n        outputArray[i] = buffer[halfWindowSize];\n    }\n\n    return outputArray;\n}\n\n/**\n * Helper function to round a number to a given number of decimals\n * @param {number} num The number to round\n * @param {number} decimals The number of decimals\n * @returns {number} The rounded number\n */\nexport function round(num, decimals) {\n    const pow = Math.pow(10, decimals);\n    return Math.round(num * pow) / pow;\n}\n\n/**\n * Helper function to round a number to the nearest integer, with ties rounded to the nearest even number.\n * Also known as \"bankers' rounding\". This is the default rounding mode in python. For example:\n * 1.5 rounds to 2 and 2.5 rounds to 2.\n * \n * @param {number} x The number to round\n * @returns {number} The rounded number\n */\nexport function bankers_round(x) {\n    const r = Math.round(x);\n    const br = Math.abs(x) % 1 === 0.5 ? (r % 2 === 0 ? r : r - 1) : r;\n    return br;\n}\n"],"names":["max"],"mappings":"AAiGO,SAAS,aAAa,OAAO,MAAM,MAAM;AAG5C,QAAM,QAAQ,IAAI,MAAM,KAAK,MAAM;AACnC,QAAM,SAAS,IAAI,MAAM,KAAK,MAAM;AAEpC,WAAS,IAAI,KAAK,SAAS,GAAG,IAAI,GAAG,KAAK,GAAG,EAAE,GAAG;AAC9C,WAAO,CAAC,IAAI;AACZ,UAAM,CAAC,IAAI,KAAK,KAAK,CAAC,CAAC;AACvB,SAAK,MAAM,CAAC;AAAA,EAChB;AAGA,QAAM,YAAY,KAAK,IAAI,CAAC,GAAG,MAAM,OAAO,KAAK,QAAQ,CAAC,CAAC,CAAC;AAI5D,QAAM,eAAe,IAAI,MAAM,YAAY,MAAM,MAAM;AAGvD,WAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,EAAE,GAAG;AACnC,QAAI,WAAW;AACf,aAAS,IAAI,KAAK,SAAS,GAAG,IAAI,GAAG,KAAK,GAAG,EAAE,GAAG;AAC9C,kBAAa,IAAI,KAAK,CAAC,IAAK,UAAU,CAAC;AACvC,UAAI,KAAK,MAAM,IAAI,KAAK,CAAC,CAAC;AAAA,IAC9B;AACA,iBAAa,QAAQ,IAAI,MAAM,CAAC;AAAA,EACpC;AAEA,SAAO,CAAC,cAAc,KAAK;AAC/B;AASO,SAAS,QAAQ,KAAK;AAEzB,QAAM,SAAS,IAAI,GAAG,EAAE,CAAC;AAGzB,QAAM,OAAO,IAAI,IAAI,OAAK,KAAK,IAAI,IAAI,MAAM,CAAC;AAI9C,QAAM,UAAU,KAAK,OAAO,CAAC,KAAK,QAAQ,MAAM,KAAK,CAAC;AAGtD,QAAM,aAAa,KAAK,IAAI,OAAK,IAAI,OAAO;AAE5C;AAAA;AAAA,IAAwB;AAAA;AAC5B;AAQO,SAAS,YAAY,KAAK;AAE7B,QAAM,aAAa,QAAQ,GAAG;AAG9B,QAAM,gBAAgB,WAAW,IAAI,OAAK,KAAK,IAAI,CAAC,CAAC;AAErD;AAAA;AAAA,IAAwB;AAAA;AAC5B;AAuBO,SAAS,YAAY,OAAO,QAAQ,GAAG;AAG1C,UAAQ,MAAM,KAAK,KAAK,EACnB,IAAI,CAAC,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC,EACpB,KAAK,CAAC,GAAG,MAAM,EAAE,CAAC,IAAI,EAAE,CAAC,CAAC;AAE/B,MAAI,UAAU,QAAQ,QAAQ,GAAG;AAC7B,YAAQ,MAAM,MAAM,GAAG,KAAK;AAAA,EAChC;AAEA,SAAO;AACX;AA6DO,SAAS,IAAI,KAAK;AACrB,MAAI,IAAI,WAAW,EAAG,OAAM,MAAM,yBAAyB;AAC3D,MAAIA,OAAM,IAAI,CAAC;AACf,MAAI,aAAa;AACjB,WAAS,IAAI,GAAG,IAAI,IAAI,QAAQ,EAAE,GAAG;AACjC,QAAI,IAAI,CAAC,IAAIA,MAAK;AACd,MAAAA,OAAM,IAAI,CAAC;AACX,mBAAa;AAAA,IACjB;AAAA,EACJ;AACA,SAAO,CAAC,OAAOA,IAAG,GAAG,UAAU;AACnC;AA4oBO,SAAS,aAAa,MAAM,YAAY;AAE3C,MAAI,aAAa,MAAM,KAAK,cAAc,GAAG;AACzC,UAAM,IAAI,MAAM,2CAA2C;AAAA,EAC/D;AAGA,QAAM,cAAc,IAAI,KAAK,YAAY,KAAK,MAAM;AAGpD,QAAM,SAAS,IAAI,KAAK,YAAY,UAAU;AAE9C,QAAM,iBAAiB,KAAK,MAAM,aAAa,CAAC;AAEhD,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,EAAE,GAAG;AAClC,QAAI,cAAc;AAElB,aAAS,IAAI,CAAC,gBAAgB,KAAK,gBAAgB,EAAE,GAAG;AACpD,UAAI,QAAQ,IAAI;AAChB,UAAI,QAAQ,GAAG;AACX,gBAAQ,KAAK,IAAI,KAAK;AAAA,MAC1B,WAAW,SAAS,KAAK,QAAQ;AAC7B,gBAAQ,KAAK,KAAK,SAAS,KAAK;AAAA,MACpC;AAEA,aAAO,aAAa,IAAI,KAAK,KAAK;AAAA,IACtC;AAEA,WAAO,KAAI;AACX,gBAAY,CAAC,IAAI,OAAO,cAAc;AAAA,EAC1C;AAEA,SAAO;AACX;","x_google_ignoreList":[0]}