CODEWITHSUNDEEP
javascriptreactjsreact-nativetensorflowtensorflow.js
yudhiesh
yudhiesh

Reputation: 6809

Fail to crop image with error: failed to compile fragment shader Tensorflow.js

I am trying to crop the faces from BlazeFaceModel based on the bounding boxes but I am getting the following error. It throws a huge error in my console with a lot of what looks like C++ code. I have not seen this before and am not sure what to do to fix it.

    const tensorDims = { height: 224, width: 224, depth: 3 };
    const returnTensors = true;
    const faces = await bfModel
      .estimateFaces(tensor, returnTensors)
      .catch(e => console.log(e));
    const tensorReshaped = tensor.reshape([1, 224, 224, 3]);
    const scale = {
      height: styles.camera.height / tensorDims.height,
      width: styles.camera.width / tensorDims.width
    };

    // Faces is an array of objects
    if (!isEmpty(faces)) {
      setModelFaces({ faces });
      faces.map((face, i) => {
        const { topLeft, bottomRight } = face;
        const boxes = tf.concat([topLeft, bottomRight]).reshape([-1, 4]);
        const width =
          (bottomRight.dataSync()[0] - topLeft.dataSync()[0]) * scale.width;
        const height =
          (bottomRight.dataSync()[1] - topLeft.dataSync()[1]) * scale.height;
        const crop = tf.image.cropAndResize(
          tensorReshaped,
          boxes,
          [0],
          [height, width]
        );
        console.log(crop);
      });
    }

Boxes

Tensor {
  "dataId": Object {},
  "dtype": "float32",
  "id": 63582,
  "isDisposedInternal": false,
  "kept": false,
  "rankType": "2",
  "scopeId": 119566,
  "shape": Array [
    1,
    4,
  ],
  "size": 4,
  "strides": Array [
    4,
  ],
}

Upvotes: 2

Views: 525

Answers (1)

edkeveked
edkeveked

Reputation: 18381

The issue is coming from width and height that are scaled leading to float values. width and height are used for the shape of the crop tensor and should therefore be integers. You can either use Math.Round or Math.floor to get an integer value

Upvotes: 2

Related Questions