onnxruntime/js/web/lib/onnxjs/backends/webgl/ops/batch-normalization.ts

// Copyright (c) Microsoft Corporation. All rights reserved.
// Licensed under the MIT License.

import { AttributeWithCacheKey, createAttributeWithCacheKey } from '../../../attribute-with-cache-key';
import { Graph } from '../../../graph';
import { OperatorImplementation, OperatorInitialization } from '../../../operators';
import { Tensor } from '../../../tensor';
import { getGlsl } from '../glsl-source';
import { WebGLInferenceHandler } from '../inference-handler';
import { ProgramInfo, TextureType } from '../types';

export interface BatchNormalizationAttributes extends AttributeWithCacheKey {
  epsilon: number;
  momentum: number;
  spatial: number;
}

const batchNormalizationProgramMetadata = {
  name: 'BatchNormalization',
  inputNames: ['A', 'Scale', 'B', 'Mean', 'Variance'],
  inputTypes: [
    TextureType.unpacked,
    TextureType.unpacked,
    TextureType.unpacked,
    TextureType.unpacked,
    TextureType.unpacked,
  ],
};

export const batchNormalization: OperatorImplementation<BatchNormalizationAttributes> = (
  inferenceHandler: WebGLInferenceHandler,
  inputs: Tensor[],
  attributes: BatchNormalizationAttributes,
): Tensor[] => {
  validateInputs(inputs);
  const output = inferenceHandler.run(
    {
      ...batchNormalizationProgramMetadata,
      cacheHint: attributes.cacheKey,
      get: () => createBatchNormalizationProgramInfo(inferenceHandler, inputs, attributes),
    },
    inputs,
  );
  return [output];
};

export const parseBatchNormalizationAttributes: OperatorInitialization<BatchNormalizationAttributes> = (
  node: Graph.Node,
): BatchNormalizationAttributes => {
  const epsilon = node.attributes.getFloat('epsilon', 1e-5);
  const momentum = node.attributes.getFloat('momentum', 0.9);
  const spatial = node.attributes.getInt('spatial', 1);
  return createAttributeWithCacheKey({ epsilon, momentum, spatial });
};

const createBatchNormalizationProgramInfo = (
  inferenceHandler: WebGLInferenceHandler,
  inputs: Tensor[],
  attributes: BatchNormalizationAttributes,
): ProgramInfo => {
  const glsl = getGlsl(inferenceHandler.session.backend.glContext.version);
  const rank = inputs[0].dims.length;
  const [scaleWidth, scaleHeight] = inferenceHandler.calculateTextureWidthAndHeight(
    inputs[1].dims,
    TextureType.unpacked,
  );
  const shaderSource = `
  float process(int[${rank}] indices) {
    vec2 position = offsetToCoords(indices[1], ${scaleWidth}, ${scaleHeight});
    float scale = getColorAsFloat(${glsl.texture2D}(Scale, position));
    float mean = getColorAsFloat(${glsl.texture2D}(Mean, position));
    float variance = getColorAsFloat(${glsl.texture2D}(Variance, position));
    float b = getColorAsFloat(${glsl.texture2D}(B, position));

    return scale * ( (_A(indices) - mean) / sqrt(variance + float(${attributes.epsilon})) ) + b;
  }`;
  return {
    ...batchNormalizationProgramMetadata,
    output: { dims: inputs[0].dims, type: inputs[0].type, textureType: TextureType.unpacked },
    shaderSource,
  };
};

const validateInputs = (inputs: Tensor[]): void => {
  if (!inputs || inputs.length !== 5) {
    throw new Error('BatchNormalization requires 5 inputs.');
  }

  const X = inputs[0];
  const scale = inputs[1];
  const B = inputs[2];
  const mean = inputs[3];
  const var_ = inputs[4];

  // input should atleast have three dimensions - N,C,dim1,...,dimn
  // other inputs can have only one dimensions
  if (
    X.dims.length < 3 ||
    scale.dims.length !== 1 ||
    B.dims.length !== 1 ||
    mean.dims.length !== 1 ||
    var_.dims.length !== 1
  ) {
    throw new Error('invalid input shape.');
  }
  if (
    scale.dims[0] !== X.dims[1] ||
    B.dims[0] !== X.dims[1] ||
    mean.dims[0] !== X.dims[1] ||
    var_.dims[0] !== X.dims[1]
  ) {
    throw new Error('invalid input shape.');
  }
  if (
    (X.type !== 'float32' && X.type !== 'float64') ||
    (scale.type !== 'float32' && scale.type !== 'float64') ||
    (B.type !== 'float32' && B.type !== 'float64') ||
    (mean.type !== 'float32' && mean.type !== 'float64') ||
    (var_.type !== 'float32' && var_.type !== 'float64')
  ) {
    throw new Error('invalid input tensor types.');
  }
};