Different Values on Pixel and Sony Xperia Z1

[ehennis]

To get help from the community, we encourage using Stack Overflow and the tensorflow.js tag.

TensorFlow.js version 1.0. Pulled from <script src="https://cdn.jsdelivr.net/npm/@tensorflow/tfjs@1.0.0/dist/tf.min.js"></script>

Browser version

Desktop: Chrome 74.0.3729.131
Pixel 2XL: 74.0.3729.136

Describe the problem or feature request

When I run a predict on my desktop (both chrome and IE) I get 2.743 which matches my model's prediction in Google Colab. When I run from my phone I get a value, then on the second click I get a different value that stays through the rest of my clicks.

Code to reproduce the bug / link to feature request

Navigate to http://bet.eckronsoftware.com/. Once you see "Model Loaded" you can hit the 'Get Prediction' button. On the desktop you get 2.73. From your phone you will get a random wrong number.

The model was created using tf.keras.experimental.export_saved_model from within Google Colab.

[msektrier]

I can confirm it on my Sony Xperia Z1.

[aptlin]

Failed to reproduce on iOS 12.2.

[ehennis]

I didn't even think to check an iPhone but when I used my work iPhone running 12.2 I got the CORRECT results.
If you need any other testing let me know. If there is some sort of verbose logging that I can turn on I can easily do that. Or, give you any files you need. Or, if you have a test version of the tf.min.js I can run locally.

[dansitu]

I can reproduce on my Samsung Galaxy S10+ in Chrome. Result is 1.1265866[...] on the phone, 2.743607759475708 on my MacBook Pro in Chrome.

[dsmilkov]

@dansitu, @msektrier and @ehennis

Can you add the following at the beginning of your js program (before loading any tf.js model, or doing any execution): tf.ENV.set('WEBGL_PACK', false) and see if the precision problems go away. Thank you!

[ehennis]

Yes, this resolved the issue for me on my phone. I have updated my code at http://bet.eckronsoftware.com/ as well if anyone else needs to try it out.

[dansitu]

Thanks Daniel and Evan! This now works as expected on my S10+.

…

On Mon, Jul 1, 2019 at 1:40 PM Evan Hennis ***@***.***> wrote: Yes, this resolved the issue for me on my phone. I have updated my code at http://bet.eckronsoftware.com/ as well if anyone else needs to try it out. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#1586?email_source=notifications&email_token=AAC4SY3BNVADG67LRO65F53P5JTTNA5CNFSM4HNGAAD2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGODY7JGDY#issuecomment-507417359>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAC4SY244NZQPLYD7XDXUPDP5JTTNANCNFSM4HNGAADQ> .

-- Daniel Situnayake | Developer Advocate, TensorFlow Lite <https://www.tensorflow.org/lite/> | dansitu@google.com | tensorflow.org

[dsmilkov]

Thanks all for such quick replies! cc @annxingyuan

I want understand if this problem is fixed with the latest version of tfjs (with tensorflow/tfjs-core#1796), or if we still have it at master.

Can you depend on latest TF.js 1.2.2, while NOT using the workaround I sent you, and see if you still get the numerical issues?

Thanks!

[ehennis]

Removing the workaround and changing tfjs from 1.0.0 to 1.2.2 WORKED on mobile.

Here is my script code:
<script src="https://cdn.jsdelivr.net/npm/@tensorflow/tfjs@1.2.2/dist/tf.min.js"></script>