已移植到 TensorFlow.js 的预训练 Tensorflow 模型的权重是否可以在运行时通过模型对象访问？

Question

如何在调试期间访问模型的权重？

当我在调试器中执行期间检查 model.model.weights['dense_3/bias'][0] 时，实际权重不存在。但是，当我 console.log 打印权重的表达式时。似乎有某种延迟执行正在进行？

我在下方创建了一个基于有毒分类器 medium article 的代码片段，展示了如何访问特定层的权重对象。

const threshold = 0.9;

// Which toxicity labels to return.
const labelsToInclude = ['identity_attack', 'insult', 'threat'];

toxicity.load(threshold, labelsToInclude).then(model => {
    // Now you can use the `model` object to label sentences. 
    model.classify(['you suck']).then(predictions => {
    console.log("Specific weights: "+ model.model.weights['dense_3/bias'][0])
      document.getElementById("predictions").innerHTML =  JSON.stringify(predictions, null, 2);
    });
});

<!DOCTYPE html>
<html lang="en-us">
<head>
  <meta charset="UTF-8">
  <title>Activity 1: Basic HTML Bio</title>
  <script src="https://cdn.jsdelivr.net/npm/@tensorflow/tfjs@1.0"></script>
<script src="https://cdn.jsdelivr.net/npm/@tensorflow-models/toxicity"></script>
</head>

<body>
<div id="predictions">
Will be populated by prebuilt toxicity model
</div>
</body>

</html>

Answer 1

张量数组中的每一层。可以通过遍历数组来访问层的权重。

const t = model.model.weights['dense_3/bias'][0] // t is a tensor
t.print() // will display the tensor in the console
// to add value to the weight
t.add(tf.scalar(0.5))

console.log(model.model.weights['dense_3/bias'][0]) 将显示一个对象而不是张量的值。原因是张量在 TypeScript 中是 class，它在 js 中被转换为 Function 类型的对象。这就是为什么 console.log(model.model.weights['dense_3/bias'][0]) 将打印一个对象，其键是 class 张量的属性。需要调用 print 方法来查看张量的基础值

const threshold = 0.9;

// Which toxicity labels to return.
const labelsToInclude = ['identity_attack', 'insult', 'threat'];

toxicity.load(threshold, labelsToInclude).then(model => {
    // print weights
    model.model.weights['dense_3/bias'][0].print()
    // continue processing
});

<!DOCTYPE html>
<html lang="en-us">
<head>
  <meta charset="UTF-8">
  <title>Activity 1: Basic HTML Bio</title>
  <script src="https://cdn.jsdelivr.net/npm/@tensorflow/tfjs@1.0"></script>
<script src="https://cdn.jsdelivr.net/npm/@tensorflow-models/toxicity"></script>
</head>

<body>
</body>

</html>

如果想获取cpu上的tensor值，用dom元素的innerHTML显示，可以考虑使用data or dataSync

Answer 2

从另一个我能够推导出如何访问权重。

对于每一层都有一个 data 承诺将允许访问权重。

const threshold = 0.9;

// Which toxicity labels to return.
const labelsToInclude = ['identity_attack', 'insult', 'threat'];

toxicity.load(threshold, labelsToInclude).then(model => {
    // Now you can use the `model` object to label sentences. 
    model.classify(['you suck']).then(predictions => {
      model.model.weights['dense_3/bias'][0].data().then(
        function(value) {
        document.getElementById("specific_weights").innerHTML = JSON.stringify(value);
        });
      
      document.getElementById("predictions").innerHTML =  JSON.stringify(predictions, null, 2);
    });
});

<!DOCTYPE html>
<html lang="en-us">
<head>
  <meta charset="UTF-8">
  <title>Activity 1: Basic HTML Bio</title>
  <script src="https://cdn.jsdelivr.net/npm/@tensorflow/tfjs@1.0"></script>
<script src="https://cdn.jsdelivr.net/npm/@tensorflow-models/toxicity"></script>
</head>

<body>
<div id="predictions">
Will be populated by prebuilt toxicity model
</div>
<div id="specific_weights">
Will contain weights for specific layer
</div>

</body>

</html>

已移植到 TensorFlow.js 的预训练 Tensorflow 模型的权重是否可以在运行时通过模型对象访问？

Are the weights of a pre-trained Tensorflow model that has been ported to TensorFlow.js accessible at runtime through the model object?

javascript

tensorflow

tensorflow.js