Cannot use MediaPipeTasksVision pod alongside TensorFlowLiteSwift pod with Metal backend

[babldev]

Have I written custom code (as opposed to using a stock example script provided in MediaPipe)

No

OS Platform and Distribution

iOS 18.1.1

Mobile device if the issue happens on mobile device

iPhone 15

Browser and version if the issue happens on browser

No response

Programming Language and version

Swift

MediaPipe version

0.10.20

Bazel version

No response

Solution

Object Detection (any)

Android Studio, NDK, SDK versions (if issue is related to building in Android environment)

No response

Xcode & Tulsi version (if issue is related to building for iOS)

16.1

Describe the actual behavior

Runtime error: "objc[1064]: Class TFLBufferConvert is implemented in both /private/var/containers/Bundle/Application/D3732F45-70C8-458D-8BD7-3A8E99C8BB47/ObjectDetector.app/ObjectDetector.debug.dylib (0x106250610) and /private/var/containers/Bundle/Application/D3732F45-70C8-458D-8BD7-3A8E99C8BB47/ObjectDetector.app/ObjectDetector.debug.dylib (0x1062526e0). One of the two will be used. Which one is undefined."

Describe the expected behaviour

I should be able to use "pod 'TensorFlowLiteSwift', '~> 0.0.1-nightly', :subspecs => ['Metal']"

Standalone code/steps you may have used to try to get what you need

Use this Podfile

target 'ObjectDetector' do

  use_frameworks!
  pod 'MediaPipeTasksVision', '0.10.20'
  pod 'TensorFlowLiteSwift', '~> 0.0.1-nightly', :subspecs => ['Metal']
  
  # Pods for ObjectDetector
end

pod install

Build & run iOS app

Other info / Complete Logs

Most developers considering using Mediapipe in iOS apps will have separate TFLite models executed using TensorFlowLiteSwift directly. However, once the Mediapipe pod (eg MediaPipeTasksVision) is added to your podfile, existing usage of TensorFlowLiteSwift + Metal delegate will cause runtime errors. They both link to the same Obj-C code.

The Mediapipe iOS Pod should support a configuration that does not collide with TensorFlowLiteSwift Pod + Metal to allow greater iOS adoption with apps already using TFLite.

objc[1064]: Class TFLBufferConvert is implemented in both /private/var/containers/Bundle/Application/D3732F45-70C8-458D-8BD7-3A8E99C8BB47/ObjectDetector.app/ObjectDetector.debug.dylib (0x106250610) and /private/var/containers/Bundle/Application/D3732F45-70C8-458D-8BD7-3A8E99C8BB47/ObjectDetector.app/ObjectDetector.debug.dylib (0x1062526e0). One of the two will be used. Which one is undefined.
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1734382404.659640  236024 gl_context.cc:369] GL version: 3.0 (OpenGL ES 3.0 Metal - 101), renderer: Apple A16 GPU

[babldev]

Related #3936 #2378

[kuaashish]

Hi @babldev,

Apologies for the inconvenience. However, we have not tested compatibility with TensorFlow Lite Swift using the Metal backend. Unfortunately, we are unable to provide support for this issue based on our internal decision.

Thank you!!