tfjs 2.8.0 is broken and introduces regressions in tf.image.cropAndResize

[vladmandic]

As subject line says, TFJS 2.8.0 unfortunately seems like a broken version.

First, default parameter in cropAndResize was unintentionally removed (and apparently just re-added via #4407) which causes quite a lot of errors in existing apps

Second, even when specifying resize method as bilinear, cropAndResize FAILS on WebGL baclend:

Uncaught (in promise) Error: Failed to compile fragment shader.
    at createFragmentShader (webgl_util.ts:103)
    at GPGPUContext.createProgram (gpgpu_context.ts:280)
    at compileProgram (gpgpu_math.ts:93)
    at backend_webgl.ts:858
    at MathBackendWebGL.getAndSaveBinary (backend_webgl.ts:902)
    at MathBackendWebGL.runWebGLProgram (backend_webgl.ts:857)
    at Object.cropAndResize3 [as kernelFunc] (CropAndResize.ts:35)
    at kernelFunc3 (engine.ts:590)
    at engine.ts:660
    at Engine.scopedRun (engine.ts:453)

With WebGL code dump highlighting error in

204 setOutput(float(undefined));

IMO, this is is blocking bug - once fixed, TFJS packages should be republished!
(and I'm not sure how this passed any level of testing?)

Btw, same code confirmed working after downgrade to tfjs 2.7.0.

[vladmandic]

cc @annxingyuan @lina128 for visibility - sorry for spam, but this is bad

[lina128]

cc @tafsiri

[annxingyuan]

Hi @vladmandic - thanks for reporting this issue - just wanted to let you know that you can avoid the fragment shader compile error by passing both method and extrapolationValue into cropAndResize, e.g.: tf.image.cropAndResize(image, boxes, boxInd, [5,5], 'bilinear', 0 /* extrapolation value */) // note the last argument

[vladmandic]

@annxingyuan thanks for that - but I have cropAndResize all over my code, I'll just wait until this is fixed and tfjs is republished - in the meantime i'm sticking with tfjs 2.7.0.

[lina128]

Hi @vladmandic, we are going to release the fix in 2.8.1 very soon.

[mgol]

Since 2.8.0 is tagged as latest, the whole package looks deprecated now, see:
https://www.npmjs.com/package/@tensorflow/tfjs
I'd recommend to tag 2.7.0 as latest until you have a fixed version to remove the warning from the package npm page:

npm dist-tag add @tensorflow/tfjs@2.7.0 latest

[mattsoulanille]

Thanks for the tip @mgol! We've tagged 2.7.0 as the latest npm version for all packages that were in the 2.8.0 release.

[lina128]

2.8.1 is released. It should fix the bug.

[google-ml-butler]

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

[vladmandic]

tfjs 2.8.1 fixes this issue, but exposes another issue in #4429