iOS 17 beta issue: abnormal search results.

[flipped0119]

In iOS 17 beta 3, regardless of the keyword used, the search results are always the same.

[mazzzystar]

It appears to be a bug introduced in the iOS 17 beta, where input text is being embedded into the same vector. Currently, I am unable to determine the exact cause of this issue.

[RupGautam]

I can confirm the results are the same, not always, but most of the time.

[mazzzystar]

@RupGautam
I've debugged the issue. I found that all text inputs were embedded into the same value in iOS 17 beta. However, it works correctly on iOS 16. I suspect it was the issue related with model loading, didn't figured out the problem yet.

[RupGautam]

@mazzzystar
"Deep green/blue water with mountains around"

"Deep green/blue water without mountains around"

Gives me the same results. ⁉️

[mazzzystar]

@RupGautam Right, you're using iOS 17 beta. As I've mentioned before, every input query in iOS 17 got the same text embedding vector output, which may be cause by model loading failure.

[mazzzystar]

@RupGautam @flipped0119
did you use Xcode 15 beta to compile? I've just tried iOS 17.0 simulator on my Xcode 14.3.1 and it works correctly. I do got abnormal result on my Xcode 15, it seems to be a Xcode issue.

[Dealizey]

@mazzzystar So has the bug been fixed with Xcode beta?

[mazzzystar]

@mazzzystar So has the bug been fixed with Xcode beta?

No, I'm saying that when using Xcode 14 + iOS 17 beta simulator to run the app, everything seems to be normal. While in Xcode 15 + iOS 17 the issue occurs.

But I only have one phone, and when your phone is upgraded to iOS 17 beta, you can't submit new app to App Store. So I couldn't test this issue on my my real phone.

[mazzzystar]

Confirmed, when upgraded to iOS 17 public beta, everything back to normal.

[mazzzystar]

@flipped0119 @RupGautam @Dealizey
Finally I found the issue. For iOS 17, ou need to explicitly set dtype=np.int32 for the input tensor when export model.

max_seq_length = 77

text_encoder_model = ct.convert(
            traced_model,
            convert_to="mlprogram",
            minimum_deployment_target=ct.target.iOS16,
            inputs=[ct.TensorType(name="prompt",
                                 shape=example_input.shape),
                                 dtype=np.int32)],
            outputs=[ct.TensorType(name="embOutput", dtype=np.float32)],
        )

[Dealizey]

@mazzzystar Great! When will this bug fix be published?

[mazzzystar]

@Dealizey
It's just published. https://apps.apple.com/us/app/queryable-find-photo-by-text/id1661598353

[Dealizey]

It's quite interesting that the Chinese App Store didn't show that update, but the US one did. Maybe it's because of the CDN configuration.

[mazzzystar]

@Dealizey Queryable & 寻隐 are using two models, so it's two app. In Chinese version, this update is still under beta testing.

[Dealizey]

Thanks! Hope the Chinese version will be published soon.❤️❤️❤️

[mazzzystar]

@Dealizey It is available now.

[Dealizey]

Got that. Thanks a lot!