gx

chenyc

2025-06-12 7b72ac13a83764a662159d4a49b7fffb90476ecb

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
import * as tf from '../../dist/tfjs.esm';
 
import { NetInput, TNetInput, toNetInput } from '../dom/index';
import { NeuralNetwork } from '../NeuralNetwork';
import { normalize } from '../ops/index';
import { denseBlock4 } from './denseBlock';
import { extractParams } from './extractParams';
import { extractParamsFromWeightMap } from './extractParamsFromWeightMap';
import { FaceFeatureExtractorParams, IFaceFeatureExtractor } from './types';
 
export class FaceFeatureExtractor extends NeuralNetwork<FaceFeatureExtractorParams> implements IFaceFeatureExtractor<FaceFeatureExtractorParams> {
  constructor() {
    super('FaceFeatureExtractor');
  }
 
  public forwardInput(input: NetInput): tf.Tensor4D {
    const { params } = this;
 
    if (!params) {
      throw new Error('FaceFeatureExtractor - 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 = denseBlock4(normalized, params.dense0, true);
      out = denseBlock4(out, params.dense1);
      out = denseBlock4(out, params.dense2);
      out = denseBlock4(out, params.dense3);
      out = tf.avgPool(out, [7, 7], [2, 2], 'valid');
 
      return out;
    });
  }
 
  public async forward(input: TNetInput): Promise<tf.Tensor4D> {
    return this.forwardInput(await toNetInput(input));
  }
 
  protected getDefaultModelName(): string {
    return 'face_feature_extractor_model';
  }
 
  protected extractParamsFromWeightMap(weightMap: tf.NamedTensorMap) {
    return extractParamsFromWeightMap(weightMap);
  }
 
  protected extractParams(weights: Float32Array) {
    return extractParams(weights);
  }
}