@vladmandic/face-api/src/xception/TinyXception.ts

FaceAPI: AI-powered Face Detection & Rotation Tracking, Face Description & Recognition, Age & Gender & Emotion Prediction for Browser and NodeJS using TensorFlow/JS

import * as tf from '../../dist/tfjs.esm';

import { ConvParams, depthwiseSeparableConv } from '../common/index';
import { NetInput, TNetInput, toNetInput } from '../dom/index';
import { NeuralNetwork } from '../NeuralNetwork';
import { normalize } from '../ops/index';
import { range } from '../utils/index';
import { extractParams } from './extractParams';
import { extractParamsFromWeightMap } from './extractParamsFromWeightMap';
import { MainBlockParams, ReductionBlockParams, TinyXceptionParams } from './types';

function conv(x: tf.Tensor4D, params: ConvParams, stride: [number, number]): tf.Tensor4D {
  return tf.add(tf.conv2d(x, params.filters, stride, 'same'), params.bias);
}

function reductionBlock(x: tf.Tensor4D, params: ReductionBlockParams, isActivateInput = true): tf.Tensor4D {
  let out = isActivateInput ? tf.relu(x) : x;
  out = depthwiseSeparableConv(out, params.separable_conv0, [1, 1]);
  out = depthwiseSeparableConv(tf.relu(out), params.separable_conv1, [1, 1]);
  out = tf.maxPool(out, [3, 3], [2, 2], 'same');
  out = tf.add(out, conv(x, params.expansion_conv, [2, 2]));
  return out;
}

function mainBlock(x: tf.Tensor4D, params: MainBlockParams): tf.Tensor4D {
  let out = depthwiseSeparableConv(tf.relu(x), params.separable_conv0, [1, 1]);
  out = depthwiseSeparableConv(tf.relu(out), params.separable_conv1, [1, 1]);
  out = depthwiseSeparableConv(tf.relu(out), params.separable_conv2, [1, 1]);
  out = tf.add(out, x);
  return out;
}

export class TinyXception extends NeuralNetwork<TinyXceptionParams> {
  private _numMainBlocks: number

  constructor(numMainBlocks: number) {
    super('TinyXception');
    this._numMainBlocks = numMainBlocks;
  }

  public forwardInput(input: NetInput): tf.Tensor4D {
    const { params } = this;
    if (!params) {
      throw new Error('TinyXception - load model before inference');
    }
    return tf.tidy(() => {
      const batchTensor = tf.cast(input.toBatchTensor(112, true), 'float32');
      const meanRgb = [122.782, 117.001, 104.298];
      const normalized = normalize(batchTensor, meanRgb).div(255) as tf.Tensor4D;
      let out = tf.relu(conv(normalized, params.entry_flow.conv_in, [2, 2]));
      out = reductionBlock(out, params.entry_flow.reduction_block_0, false);
      out = reductionBlock(out, params.entry_flow.reduction_block_1);
      range(this._numMainBlocks, 0, 1).forEach((idx) => {
        out = mainBlock(out, params.middle_flow[`main_block_${idx}`]);
      });
      out = reductionBlock(out, params.exit_flow.reduction_block);
      out = tf.relu(depthwiseSeparableConv(out, params.exit_flow.separable_conv, [1, 1]));
      return out;
    });
  }

  public async forward(input: TNetInput): Promise<tf.Tensor4D> {
    return this.forwardInput(await toNetInput(input));
  }

  protected getDefaultModelName(): string {
    return 'tiny_xception_model';
  }

  protected extractParamsFromWeightMap(weightMap: tf.NamedTensorMap) {
    return extractParamsFromWeightMap(weightMap, this._numMainBlocks);
  }

  protected extractParams(weights: Float32Array) {
    return extractParams(weights, this._numMainBlocks);
  }
}