I would like to get the data from a 2D tensor with tensorflow.js. I tried to use the data() method like this:
const X = tf.tensor2d([[1, 2, 3, 4], [2, 2, 5, 3]]);
X.data().then(X => console.log(X)};
But the result is a flatten 1D array:
Float32Array(8) [1, 2, 3, 4, 2, 2, 5, 3]
Is there a way to keep the shape of the array?
You can use arraySync method on the tensor object. It returns the array as the same shape it was before synchronously.
const X = tf.tensor2d([[1, 2, 3, 4], [2, 2, 5, 3]]);
console.log(X.arraySync())<script src="https://cdn.jsdelivr.net/npm/@tensorflow/[email protected]/dist/tf.min.js"></script>Data in the Tensor is always stored flattened as types 1 dimensional array, for speed.
The example you gave will not work, because 2nd parameter to tensor2d is shape. To make it work you either need to wrap it another array:
const x = tf.tensor2d([[1, 2, 3, 4], [2, 2, 5, 3]]); //shape inferred as [2, 4]
or you could explicitly provide shape:
const x = tf.tensor2d([1, 2, 3, 4, 2, 2, 5, 3], [2, 4]); // shape explicitly passed
as you suggested though, when you inspect data you will always get 1 dimensional array, regardless of original shape
await x.data() // Float32Array(8) [1, 2, 3, 4, 2, 2, 5, 3]
x.shape // [2, 4]
if however you print() your tensor, shape is taken into account and it will appear as
Tensor
[[1, 2, 3, 4],
[2, 2, 5, 3]]
I use a function to show the 2D tensor in a webpage
async function myTensorTable(myDiv, myOutTensor, myCols, myTitle){
document.getElementById(myDiv).innerHTML += myTitle + '<br>'
const myOutput = await myOutTensor.data()
myTemp = '<table border=3><tr>'
for (myCount = 0; myCount <= myOutTensor.size - 1; myCount++){
myTemp += '<td>'+ myOutput[myCount] + '</td>'
if (myCount % myCols == myCols-1){
myTemp += '</tr><tr>'
}
}
myTemp += '</tr></table>'
document.getElementById(myDiv).innerHTML += myTemp + '<br>'
}
examples of usage at
