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
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
/**
 * @license
 * Copyright 2018 Google Inc. All Rights Reserved.
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 * http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 * =============================================================================
 */
 
import {ENGINE} from '../engine';
import {dispose, tidy} from '../globals';
import {zerosLike} from '../ops/ops';
import {ConfigDict, registerClass, Serializable, SerializableConstructor} from '../serialization';
import {NamedTensor, NamedVariableMap} from '../tensor_types';
import {Optimizer, OptimizerVariable} from './optimizer';
 
/** @doclink Optimizer */
export class AdadeltaOptimizer extends Optimizer {
  /** @nocollapse */
  static className = 'Adadelta';  // Name matters for Python compatibility.
  private accumulatedGrads: OptimizerVariable[] = [];
  private accumulatedUpdates: OptimizerVariable[] = [];
 
  constructor(
      protected learningRate: number, protected rho: number,
      protected epsilon: number = null) {
    super();
 
    if (epsilon == null) {
      this.epsilon = ENGINE.backend.epsilon();
    }
  }
 
  applyGradients(variableGradients: NamedVariableMap|NamedTensor[]) {
    const variableNames = Array.isArray(variableGradients) ?
        variableGradients.map(item => item.name) :
        Object.keys(variableGradients);
 
    variableNames.forEach((name, i) => {
      const value = ENGINE.registeredVariables[name];
      const trainable = false;
      if (this.accumulatedGrads[i] == null) {
        this.accumulatedGrads[i] = {
          originalName: `${name}/accum_grad`,
          variable: tidy(() => zerosLike(value).variable(trainable))
        };
      }
      if (this.accumulatedUpdates[i] == null) {
        this.accumulatedUpdates[i] = {
          originalName: `${name}/accum_var`,
          variable: tidy(() => zerosLike(value).variable(trainable))
        };
      }
 
      const gradient = Array.isArray(variableGradients) ?
          variableGradients[i].tensor :
          variableGradients[name];
      if (gradient == null) {
        return;
      }
 
      const accumulatedGrad = this.accumulatedGrads[i].variable;
      const accumulatedUpdate = this.accumulatedUpdates[i].variable;
 
      tidy(() => {
        const newAccumulatedGrad = accumulatedGrad.mul(this.rho).add(
            gradient.square().mul(1 - this.rho));
 
        const updates = accumulatedUpdate.add(this.epsilon)
                            .sqrt()
                            .div(accumulatedGrad.add(this.epsilon).sqrt())
                            .mul(gradient);
 
        const newAccumulatedUpdate = accumulatedUpdate.mul(this.rho).add(
            updates.square().mul(1 - this.rho));
 
        accumulatedGrad.assign(newAccumulatedGrad);
        accumulatedUpdate.assign(newAccumulatedUpdate);
 
        const newValue = updates.mul(-this.learningRate).add(value);
        value.assign(newValue);
      });
    });
    this.incrementIterations();
  }
 
  dispose(): void {
    if (this.accumulatedUpdates != null) {
      dispose(this.accumulatedGrads.map(v => v.variable));
      dispose(this.accumulatedUpdates.map(v => v.variable));
    }
  }
 
  async getWeights(): Promise<NamedTensor[]> {
    // Order matters for Python compatibility.
    const variables: OptimizerVariable[] =
        [...this.accumulatedGrads, ...this.accumulatedUpdates];
    return [await this.saveIterations()].concat(
        variables.map(v => ({name: v.originalName, tensor: v.variable})));
  }
 
  async setWeights(weightValues: NamedTensor[]): Promise<void> {
    weightValues = await this.extractIterations(weightValues);
    const variableCount = weightValues.length / 2;
    const trainable = false;
    this.accumulatedGrads =
        weightValues.slice(0, variableCount).map(v => ({
                                                   originalName: v.name,
                                                   variable: v.tensor.variable(
                                                       trainable)
                                                 }));
    this.accumulatedUpdates =
        weightValues.slice(variableCount, variableCount * 2)
            .map(v => ({
                   originalName: v.name,
                   variable: v.tensor.variable(trainable)
                 }));
  }
 
  getConfig(): ConfigDict {
    return {
      'learningRate': this.learningRate,
      'rho': this.rho,
      'epsilon': this.epsilon
    };
  }
 
  /** @nocollapse */
  static fromConfig<T extends Serializable>(
      cls: SerializableConstructor<T>, config: ConfigDict): T {
    return new cls(config['learningRate'], config['rho'], config['epsilon']);
  }
}
registerClass(AdadeltaOptimizer);