@hoff97/tensor-js/dist/lib/autograd/variable.d.ts

PyTorch like deep learning inferrence library

import { Tensor } from '../library';
import { CPUTensor } from '../tensor/cpu/tensor';
import { TensorValues, Activation, PadMode, DType } from '../types';
import { BackwardOp, VariableI } from './types';
import { Backend } from '../util/convert';
import REGL from 'regl';
export interface VariableOptions<DTpe extends DType> {
    /**
     * The gradient can optionally be specified
     */
    grad?: Tensor<DTpe>;
    /**
     * Backward edge of this variable
     *
     * You most likely do not want to use this
     */
    backEdge?: BackwardOp<DTpe>;
    /**
     * When set to true, gradients will not be tracked for this
     * variable. Useful for data that is passed into a model.
     */
    noGrad?: boolean;
}
/**
 * Tensor that also has a gradient associated to it
 * When noGrad is false, a dynamic computation graph on
 * this variable will be build.
 *
 * Once backward on a scalar variable (eg. a variable with shape [1])
 * is called, the gradients for all variables will be computed
 */
export declare class Variable<DTpe extends DType = 'float32'> extends Tensor<DTpe> implements VariableI<DTpe> {
    value: Tensor<DTpe>;
    grad?: Tensor<DTpe>;
    backEdge?: BackwardOp<DTpe>;
    noGrad: boolean;
    /**
     * Creates a variable whose value is the specified value
     */
    constructor(value: Tensor<DTpe>, options?: VariableOptions<DTpe>);
    static create<DTpe extends DType>(shape: ReadonlyArray<number>, values: number[], backend: Backend, options?: VariableOptions<DTpe>, dtype?: DTpe): Variable<DTpe>;
    /**
     * Creates a GPU variable from texture data (eg. Image/Video element)
     */
    static fromData(data: REGL.TextureImageData, options?: VariableOptions<'float32'>): Variable<'float32'>;
    cast<DTpe2 extends DType>(dtype: DTpe2): Tensor<DTpe2>;
    /**
     * Performs a backward pass and returns wether the grad is needed or can be deleted
     */
    backward(grad?: Tensor<DTpe>): boolean;
    isLeaf(): boolean;
    constantLike(value: number): Tensor<DTpe>;
    singleConstant(value: number): Tensor<DTpe>;
    getValues(): Promise<TensorValues[DTpe]>;
    getShape(): readonly number[];
    delete(): void;
    protected reshape_impl(shape: readonly number[], copy: boolean): Tensor<DTpe>;
    exp(): Tensor<DTpe>;
    log(): Tensor<DTpe>;
    sqrt(): Tensor<DTpe>;
    abs(): Tensor<DTpe>;
    sin(): Tensor<DTpe>;
    cos(): Tensor<DTpe>;
    tan(): Tensor<DTpe>;
    asin(): Tensor<DTpe>;
    acos(): Tensor<DTpe>;
    atan(): Tensor<DTpe>;
    sinh(): Tensor<DTpe>;
    cosh(): Tensor<DTpe>;
    tanh(): Tensor<DTpe>;
    asinh(): Tensor<DTpe>;
    acosh(): Tensor<DTpe>;
    atanh(): Tensor<DTpe>;
    sigmoid(): Tensor<DTpe>;
    hardSigmoid(alpha: number, beta: number): Tensor<DTpe>;
    sign(): Tensor<DTpe>;
    negate(): Tensor<DTpe>;
    addMultiplyScalar(factor: number, add: number): Tensor<DTpe>;
    powerScalar(power: number, factor: number): Tensor<DTpe>;
    setValues(values: Tensor<DTpe>, starts: number[]): Tensor<DTpe>;
    matMul(tensor: Tensor<DTpe>): Tensor<DTpe>;
    concat(tensor: Tensor<DTpe>, axis: number): Tensor<DTpe>;
    clip(min?: number, max?: number): Tensor<DTpe>;
    clipBackward(grad: Tensor<DTpe>, min?: number, max?: number): Tensor<DTpe>;
    repeat(repeats: number[]): Tensor<DTpe>;
    expand(shape: readonly number[]): Tensor<DTpe>;
    copy(): Tensor<DTpe>;
    gather(axis: number, indices: CPUTensor<'uint32'>): Tensor<DTpe>;
    floor(): Tensor<DTpe>;
    ceil(): Tensor<DTpe>;
    round(): Tensor<DTpe>;
    upsample(scales: number[]): Tensor<DTpe>;
    normalize(mean: Tensor<DTpe>, variance: Tensor<DTpe>, epsilon: number, scale: Tensor<DTpe>, bias: Tensor<DTpe>): Tensor<DTpe>;
    add_impl(th: Tensor<DTpe>, tensor: Tensor<DTpe>, resultShape: readonly number[], alpha: number, beta: number): Tensor<DTpe>;
    subtract_impl(th: Tensor<DTpe>, tensor: Tensor<DTpe>, resultShape: readonly number[], alpha: number, beta: number): Tensor<DTpe>;
    multiply_impl(th: Tensor<DTpe>, tensor: Tensor<DTpe>, resultShape: readonly number[], alpha: number): Tensor<DTpe>;
    divide_impl(th: Tensor<DTpe>, tensor: Tensor<DTpe>, resultShape: readonly number[], alpha: number): Tensor<DTpe>;
    power_impl(th: Tensor<DTpe>, tensor: Tensor<DTpe>, resultShape: readonly number[]): Tensor<DTpe>;
    gemm_impl(b: Tensor<DTpe>, aTranspose: boolean, bTranspose: boolean, alpha: number, beta: number, C?: Tensor<DTpe>): Tensor<DTpe>;
    protected sum_impl(axes: number[], keepDims: boolean): Tensor<DTpe>;
    protected sumSquare_impl(axes: number[], keepDims: boolean): Tensor<DTpe>;
    protected product_impl(axes: number[], keepDims: boolean): Tensor<DTpe>;
    protected max_impl(axes: number[], keepDims: boolean): Tensor<DTpe>;
    protected min_impl(axes: number[], keepDims: boolean): Tensor<DTpe>;
    protected reduceMean_impl(axes: number[], keepDims: boolean): Tensor<DTpe>;
    protected reduceMeanSquare_impl(axes: number[], keepDims: boolean): Tensor<DTpe>;
    protected reduceLogSum_impl(axes: number[], keepDims: boolean): Tensor<DTpe>;
    protected reduceLogSumExp_impl(axes: number[], keepDims: boolean): Tensor<DTpe>;
    protected conv_impl(kernel: Tensor<DTpe>, dilations: number[], group: number, pads: number[], strides: number[], activation: Activation, bias?: Tensor<DTpe>): Tensor<DTpe>;
    protected convTranspose_impl(kernel: Tensor<DTpe>, dilations: number[], group: number, pads: number[], strides: number[]): Tensor<DTpe>;
    protected pad_impl(pads: number[], mode: PadMode, value: number): Tensor<DTpe>;
    protected averagePool_impl(kernelShape: number[], pads: number[], strides: number[], includePad: boolean): Tensor<DTpe>;
    protected transpose_impl(permutation: number[]): Tensor<DTpe>;
    protected slice_impl(starts: number[], ends: number[], axes: number[], steps: number[]): Tensor<DTpe>;
}