Tensorflow.js tf.customGrad()用法及代碼示例

Tensorflow.js是Google開發的開源庫，用於在瀏覽器或節點環境中運行機器學習模型和深度學習神經網絡。

tf.customGrad()函數用於返回指定自定義函數“f”的梯度。在這裏，自定義函數給出{value：Tensor，gradFunc：(dy，saved)→Tensor []}，其中gradFunc給出輸入函數f關於其輸入的自定義梯度。

用法:

tf.customGrad(f)

參數：此函數接受如下所示的參數：

f:它是指定的自定義函數。

返回值：此函數返回指定自定義函數“f”的梯度

範例1：

Javascript


// Importing the tensorflow.js library
import * as tf from "@tensorflow/tfjs"
  
// Initializing a custom function f
const f = (a, save) => {
    
   // Saving a for its availablity later for the gradient
   save([a]);
    
   // Overriding gradient of a^2
   return {
     value:a.square(),
       
     // Here "saved.a" pointing to "a" which
     // have been saved above
     gradFunc:(dy, saved) => [dy.mul(saved[0].abs())]
   };
}
  
// Calling the .customGrad() function 
// over the above specified custom function f
const customOp = tf.customGrad(f);
  
// Initializing a 1D Tensor of some values
const a = tf.tensor1d([0, -1, 1, 2]);
  
// Getting the gradient of above function
// f for the above specified Tensor values
const da = tf.grad(a => customOp(a));
  
// Printing the custom function "f" for the
// above specified Tensor "a"
console.log(`f(a):`);
customOp(a).print();
  
// Printing the gradient of the function "f" for the
// above specified Tensor "a"
console.log(`f'(a):`);
da(a).print();

輸出：

f(a):
Tensor
   [0, 1, 1, 4]
f'(a):
Tensor
   [0, 1, 1, 2]

範例2：

Javascript


// Importing the tensorflow.js library
import * as tf from "@tensorflow/tfjs"
  
// Calling the .customGrad() function 
// with the custom function "f" as 
// it's parameter
const customOp = tf.customGrad(
    
// Initializing a custom function f
(a, save) => {
    
   // Saving a for its availablity later for the gradient
   save([a]);
    
   // Overriding gradient of a^3
   return {
     value:a.pow(tf.scalar(3, 'int32')),
       
     // Here "saved.a" pointing to "a" which
     // have been saved above
     gradFunc:(dy, saved) => [dy.mul(saved[0].abs())]
   };
}
);
  
// Initializing a 1D Tensor of some values
const a = tf.tensor1d([0, -1, 2, -2, 0.3]);
  
// Getting the gradient of above function
// f for the above specified Tensor values
const da = tf.grad(a => customOp(a));
  
// Printing the custom function "f" for the
// above specified Tensor "a"
console.log(`f(a):`);
customOp(a).print();
  
// Printing the gradient of the function "f" for the
// above specified Tensor "a"
console.log(`f'(a):`);
da(a).print();

輸出：

f(a):
Tensor
   [0, -1, 8, -8, 0.027]
f'(a):
Tensor
   [0, 1, 2, 2, 0.3]

參考：https://js.tensorflow.org/api/latest/#customGrad

相關用法

注：本文由純淨天空篩選整理自Kanchan_Ray大神的英文原創作品 Tensorflow.js tf.customGrad() Function。非經特殊聲明，原始代碼版權歸原作者所有，本譯文未經允許或授權，請勿轉載或複製。