Emgu TF Version History
Emgu.TF-2.10.1
Online Documentation
Change Log
- 2.10.1 is based on the Tensorflow 2.10.1 release.
- Release compiled with VS 2022 (instead of VS 2019)
- For xamarin demo app, use Environment.SpecialFolder.LocalApplicationData instead of Personal folder for downloading models.
- Moved native binary to "libs/runtimes/{system_arch}/native" folder
- For Emgu TF Lite
- If System DllDirectory is set, use it to look for tfliteextern binary.
- iOS binary is compiled with 11.0 SDK and now target only arm64. iOS 11.0 SDK no longer support arm.
- Rename Emgu.TF.Lite.runtime.macos package to Emgu.TF.Lite.runtime.xamarin.macos package.
- Rename Emgu.TF.Lite.runtime.ios package to Emgu.TF.Lite.runtime.xamarin.ios package.
- Rename Emgu.TF.Lite.runtime.android to Emgu.TF.Lite.runtime.xamarin.android pacakge.
- For Emgu TF
- If System DllDirectory is set, use it to look for tfextern binary.
- The CUDA release is built with CUDA 11.8 and support sm89 & sm90.
- Use Android 13 (instead of Android 11) to compile Emgu.TF.Platform.Xamarin.Android project.
Emgu.TF-2.8.0
Online Documentation
Change Log
- 2.8.0 is based on the Tensorflow 2.8.0 release.
- For Emgu TF Lite
- Added experimental support on 32bit Windows.
- Added XNNPackDelegate for tflite
- TF Lite is now build with VS2022 (instead of VS2019) on windows
- Replaced Emgu.TF.Lite.runtime.linux-arm nuget package by Emgu.TF.Lite.runtime.debian-arm. The new package is compatibile with both debian bullseye arm architecture and raspberry pi os arm architecture.
- Added Emgu.TF.Lite.runtime.debian-arm64 nuget pacakge. This package is compatibile with both debian bullseye arm64 architecture and raspberry pi os arm64 architecture.
- For Emgu TF Lite Unity
- More efficient implementation to convert WebCamTexture to drawable Texture2D.
- Fixed issue #66 Memory leak in TF Lite example for unity
- Added support for android x86_64 architecture (besides of the already supported x86, armeabi-v7a & arm64-v8a architectures).
- Added support for x86 architecture (besides of the already supported x64 architectures).
- For Emgu TF
- At library initialization, register a log listener to forward all native log message to TfInvoke.LogMegReceived event handler.
- Added functions to TfInvoke: IsBuiltWithROCm, IsBuiltWithNvcc, GpuSupportsHalfMatMulAndConv & IsMklEnabled
- Upgrade .Net 5 demo projects to .Net 6
- Support Debian arm 64 (Raspberry Pi OS - ARM64). No longer support building on Debian ARM (Raspberry Pi OS - ARM) architecture, the 2.8.0 release of tensorflow requires more than 3GB of memory to be compiled. It is no longer possible to compile it with a 32bit OS. For raspberry pi OS, only ARM64 is supported now.
- The GPU version is built with CUDA 11.6.112 (instead of 11.3.44) and CUDNN 8.3.2 (8.2.0) release
- For Emgu TF Unity
- Added support for android x86_64 architecture (besides of the already supported x86, armeabi-v7a & arm64-v8a architectures).
Emgu.TF-2.6.0
Online Documentation
Change Log
- 2.6.0 is based on the Tensorflow 2.6.0 release.
- For Emgu TF Lite
- Added Emgu.TF.Lite.Interceptor.ResizeInputTensor function.
- Emgu TF Lite nuget packages are now signed.
- For Emgu TF
- Emgu TF nuget packages are now signed.
Emgu.TF-2.5.0
Online Documentation
Change Log
- 2.5.0 is based on the Tensorflow 2.5.0 release.
- For Emgu TF Lite
- Fixed unity macos bundle deployment issue.
- Added native C++ implementation to convert image to Tensor
- Improve native binary loading logic
- For Emgu TF
- Fixed unity macos bundle deployment issue.
- Added native C++ implementation to convert image to Tensor
- Improve native binary loading logic
- Enable batch processing for Inception, Resnet and MaskRcnnInceptionV2Coco Model
- For CUDA release, updated CUDA to 11.3, CUDNN to 8.3.
Emgu.TF-2.4.1
Online Documentation
Change Log
- 2.4.1 is based on the Tensorflow 2.4.1 release.
- For Emgu TF Lite
- Renamed TfLiteInvoke.CheckLibraryLoaded() to TfLiteInvoke.Init(), TfLiteInvokeAndroid.CheckLibraryLoaded() to TfLiteInvokeAndroid.Init(), TfLiteInvokeIOS.CheckLibraryLoaded() to TfLiteInvokeIOS.Init()
- For Emgu TF
- Renamed TfInvoke.CheckLibraryLoaded() to TfInvoke.Init(), TfInvokeAndroid.CheckLibraryLoaded() to TfInvokeAndroid.Init(), TfInvokeIOS.CheckLibraryLoaded() to TfInvokeIOS.Init()
- For CUDA release, updated CUDA to 11.1, CUDNN to 8. Lower minimum compute requirement to 5.2
- Fixed issue #61
Emgu.TF-2.4.0
Online Documentation
Change Log
- 2.4.0 is based on the Tensorflow 2.4.0 release.
- The standard license terms has been change to dual license from GPL. See it here: https://github.com/emgucv/emgutf/blob/master/LICENSE The change is to allow commercial license owners to use our regular open source build and the nuget packages from nuget.org.
- For Emgu TF Lite
- The commercial release of Emgu TF Lite for windows is compiled with XNNPack for better performance. The nuget runtime package is Emgu.TF.Lite.runtime.windows.xnn
- Updated the Emgu.TF.Lite.Model nuget package version dependency for System.Drawing.Common
- Converted WPF demo from .NetFramework to .NetCore
- For Emgu TF
- Added a new Session constructor to load session from SavedModel.
- Added Resnet to Emgu.TF.Model
- Added TString API
- Converted CvInterop project from .Net Framework to .Net Core.
- Gpu release is build with CUDA 11 and CUDNN 8, targeting compute 6.0.
- Converted WPF demo from .NetFramework to .NetCore
Emgu.TF-2.2.0
Online Documentation
Change Log
- 2.2.0 is based on the Tensorflow 2.2.0 release.
- For Emgu TF Lite
- Added Emgu.TF.Lite.Model.CocoSsdMobileNetV3
- Android release is build with NDK-r20b (instead of r-18b)
- For Emgu TF
- Added TFInvoke.ListAllPhysicalDevices function.
- Android release is build with NDK-r20b (instead of r-18b)
Emgu.TF-2.1.0
Online Documentation
Change Log
- 2.1.0 is based on the Tensorflow 2.1.0 release.
- For Emgu TF Lite
- Added GpuDelegateV2 for Android
- Enable CocoSsdMobileNet demo to use camera view in iOS and MacOS
- Allow picking files on MacOS demos
- Added a Model Checker page in the Xamarin Forms demo, to load a tflite model and show the input/output tensor detail.
- Native binary targets NDK level 28 and requires at least Android 9 to meet the current Google Play Store submission requirement.
- For Emgu TF
- Android demo is built with Android SDK 10 instead of 9. Native binary targets NDK level 28 and requires at least Android 9 to meet the current Google Play Store submission requirement.