Tfjs: tf.ClipByValue() does not work as expected for the parameters (0, 1)

[Melvin-Abraham]

TensorFlow.js version: 0.10.0

Google Chrome version: 66.0.3359.181

The tf.ClipByValue() function fails to clip the values in a tensor when clipValueMin is set to 0 and clipValueMax is set to 1.

let tensor = tf.range(0, 10)
// [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]

let t_clip = tensor.clipByValue(0, 1)

t_clip.print()
// [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
// Spits out the same tensor without clipping the values

[caisq]

cc @nsthorat @dsmilkov

[nsthorat]

@reighlan this is a good starter bug!

[dsmilkov]

Hi @Melvin-Abraham , I couldn't reproduce the error. See this codepen that uses tfjs 0.10.0 and gives correct output.

Can you share more info? Are you running this on a desktop/laptop or a phone? Also, are you running this on the webgl (default) or cpu backend? Thanks!

[Melvin-Abraham]

Well, the issue is there. The values greater than 1 are not clipped to 1. By the way, check this code on Codepen for comparison...

I am using laptop with default settings for tfjs.

[manrajgrover]

The code works fine for me for both webgl and cpu. This seems like an environment issue. Could you also add a screenshot of this page which might help?

[nsthorat]

Are you using a mac & Safari by chance?

[Melvin-Abraham]

@nsthorat I'm using Chrome in Windows 10 Pro.

@ManrajGrover I've included the screenshot of the debug page...

This is the codepen screenshot for tf.clipByValue()

[valishah]

Hi @nsthorat : I'm looking into good first issue labelled issues and seems most of them are done but not concluded to close. We could just check and close the issues which are no longer active/fixed. Ex: This

[dsmilkov]

Thanks for the ping. I think this issue should still be open since the user clearly gets the wrong result on Windows. If someone else can reproduce this on a Windows machine that would be great! Thanks!

[DimaRomaniuk]

Hi @dsmilkov @Melvin-Abraham I'm using Chrome in Windows 10 Pro. And can't reproduce this issue.

[dsmilkov]

Thanks for the pointer. So the only difference I could see is that @DimaRomaniuk has WebGL 2.0, while @Melvin-Abraham has WebGL 1.0. I tried forcing WebGL 1.0 via the url flags tfjsflags=WEBGL_VERSION:1, but still couldn't reproduce the error.

Will keep this issue open, until we can get a machine that can reproduce this bug.

[DimaRomaniuk]

Hi @dsmilkov @Melvin-Abraham the same result with flag ?tfjsflags=WEBGL_VERSION:1

[nsthorat]

It looks like you are on a very old version of tfjs, can you update to 0.13.0?

[Melvin-Abraham]

Thanks, @nsthorat for pointing the use of old version of tensorflow.js. The issue seems to be solved in tfjs v0.13.0 for me...

Actually, as per my testing, the issue was already solved after the release of tfjs v0.11.7