Tfjs: Not able to run model on HTMLImageElement

[ritikrishu]

TensorFlow.js version

0.15.1

Browser version

Chrome 71.0.3578.98

Describe the problem or feature request

when I try to run model.detect on a remote image. It says,

Could not write to TensorArray index 0,
          because the value dtype is int32, but TensorArray dtype is float32.

However, if I serve the image from a localhost server, it works.

Code to reproduce the bug / link to feature request

async function init(){
      const model = await cocoSsd.load('mobilenet_v2');
      let img1 = new Image()
      // to bypass cross origin issue
      img1.src = `https://i.imgur.com/xFWR5e4.jpg?${Date.now()}`
      img1.crossOrigin = ""
      img1.onload = async function(){
        img1 = tf.image.resizeBilinear(tf.fromPixels(img1), [400, 400]);
                const result = await model.detect(img1);
      }
}

https://jsfiddle.net/rszv6a70/28/

[ritikrishu]

Also does not works with local file system images, unless I serve them via localhost as well.

[tafsiri]

I can confirm that I can reproduce the bug and it seems to be a regression between 0.14.1 and 0.14.2

@ritikrishu You can change the script tag that loads tfjs in your program to <script src="https://cdn.jsdelivr.net/npm/@tensorflow/[email protected]/dist/tf.min.js"> as a workaround in the meantime.

cc @pyu10055

[ritikrishu]

@tafsiri thanks, bro! you saved my day ❤️

[dsmilkov]

This is fixed in https://github.com/tensorflow/tfjs-core/pull/1577 as part of 1.0