tfjs 2.8.1 change in behavior in tf.reshape causes errors

upon upgrade from 2.7.0 to 2.8.1, my existing app fails with:

Uncaught (in promise) Error: The implicit shape can't be a fractional number. Got 266 / 3
    at Object.inferFromImplicitShape (util_base.ts:328)
    at Object.reshape76 [as kernelFunc] (Reshape.ts:35)
    at kernelFunc3 (engine.ts:590)
    at engine.ts:660
    at Engine.scopedRun (engine.ts:453)
    at Engine.runKernelFunc (engine.ts:657)
    at Engine.runKernel (engine.ts:522)
    at reshape_ (reshape.ts:60)
    at reshape__op (operation.ts:51)

this may be desired behavior, but it breaks several existing models published on tfhub!

e.g, i’m using blazeface where reshape is failing on results of model inference, so not like that can be changed easily:

const [contours, confidence, contourCoords] = this.meshDetector.predict(face);
const coordsReshaped = tf.reshape(contourCoords, [-1, 3]);
let rawCoords = coordsReshaped.arraySync();

[lina128]

Thanks for the bug report, we will release the fix in 2.8.2 next week.

[google-ml-butler[bot]]

Are you satisfied with the resolution of your issue? Yes No