Weekly GitHub Report for Tensorflow - 2024-12-23 12:00:01
Weekly GitHub Report for Tensorflow
Thank you for subscribing to our weekly newsletter! Each week, we deliver a comprehensive summary of your GitHub project's latest activity right to your inbox, including an overview of your project's issues, pull requests, contributors, and commit activity.
Table of Contents
I. News
1.1 Recent Version Releases:
The current version of this repository is v2.18.0
1.2 Other Noteworthy Updates:
II. Issues
2.1 Top 5 Active Issues:
We consider active issues to be issues that that have been commented on most frequently within the last week.
- Linking an Android static library with TFLite GPU using CMake causes undefined symbol errors and can not get the correct install: This issue involves a bug where linking an Android static library with TensorFlow Lite's GPU delegate using CMake results in undefined symbol errors, preventing the correct installation of the necessary environment. The problem occurs on Ubuntu 20.04.6 LTS with TensorFlow version 2.18, and despite making suggested changes to the CMakeLists.txt files, the errors persist, indicating missing targets and undefined symbols related to OpenGL and EGL functions.
- The comments discuss troubleshooting steps, including ensuring the correct OpenGL and EGL libraries are linked in the CMakeLists.txt file. A user confirms that linking these libraries resolves the initial undefined symbol errors, but they encounter a new issue with the PReLU operation not being supported in TensorFlow Lite version 2.18.0. Another user references a related discussion on the PReLU problem, seeking further guidance on resolving it in the current version.
- Number of comments this week: None
-
transfer learning with TF hub: This issue is about a bug encountered when using TensorFlow Hub's KerasLayer in a Sequential model, where the layer is not recognized as an acceptable layer, resulting in a ValueError. The problem persists even when the correct parameters are used, and the error message indicates that only instances of
keras.Layercan be added to a Sequential model.- The comments discuss a solution involving the use of an older version of Keras to avoid compatibility issues with TensorFlow versions 2.17 and 2.18, which include Keras 3.0. A user suggests installing tf-keras 2.18.0 and provides a modified code snippet that resolves the error. Another user reports a similar issue with a different notebook, and further guidance is given to ensure the correct import statements and layer usage in the code.
- Number of comments this week: None
-
tf.autodiff.ForwardAccumulator._watch(primal, tangent) erroneously refers to dtype.is_floating which does not exist for a Keras layer.: This issue involves a bug in TensorFlow version 2.17.0 where the method
tf.autodiff.ForwardAccumulator._watch(primal, tangent)incorrectly referencesprimal.dtype.is_floating, causing a crash becauseprimal.dtypeis a string type that lacks theis_floatingattribute. The problem arises when using Keras 3.0, which is included in TensorFlow 2.17, and the error persists even after upgrading to TensorFlow 2.18, suggesting a need for a fix in the TensorFlow codebase or documentation.
- The comments discuss the issue's reproduction using example code from TensorFlow's documentation, with suggestions to switch from
tf.kerastotf-kerasto avoid the error. A workaround involving installingtf_kerasand setting an environment variable is proposed, but concerns are raised about the need for a permanent solution. A pull request is mentioned to update the example code, and further testing confirms the issue persists with Keras 3, prompting a discussion on making the API compatible with Keras 3.- Number of comments this week: None
-
tensorflow error in pycharm in ubuntu 22.04: This issue is about a user experiencing an error when running TensorFlow code in PyCharm on Ubuntu 22.04, despite the code being correct and running successfully on other platforms. The user is seeking help to understand why the error occurs and why certain TensorFlow functions are not recognized in their development environment.
- The comments reveal that the user is experiencing an error in PyCharm on Ubuntu, while the code runs fine on other platforms. Another user tested the code on Colab and Ubuntu without issues, suggesting the problem might be specific to the user's setup. The original poster confirms the code is correct but encounters errors related to function recognition in PyCharm, and they seek further assistance to resolve this.
- Number of comments this week: None
-
Division by zero error at random places if GPU is used: This issue involves a division by zero error occurring randomly when a program using TensorFlow's C API is executed on a GPU, specifically a Quadro RTX 6000, while it runs without issues on a CPU. The problem is difficult to diagnose as it appears to be deep within TensorFlow or CUDA, and the user is unsure whether the bug is in the TensorFlow binary or one of the CUDA libraries.
- The comments discuss potential collaboration to resolve the issue, with suggestions for a screen-sharing session to demonstrate the problem. A contributor suggests updating incompatible versions as per TensorFlow's documentation, and the original poster attempts to update CUDA but encounters further issues. Despite multiple tests and attempts to isolate the error, the problem persists, and additional debugging information is shared to aid in troubleshooting.
- Number of comments this week: None
2.2 Top 5 Stale Issues:
We consider stale issues to be issues that has had no activity within the last 30 days. The team should work together to get these issues resolved and closed as soon as possible.
As of our latest update, there are no stale issues for the project this week.
2.3 Open Issues
This section lists, groups, and then summarizes issues that were created within the last week in the repository.
Issues Opened This Week: 12
Summarized Issues:
- TensorFlow Functionality Issues with Multidimensional Arrays: The
tensorflow.experimental.numpy.kronfunction in TensorFlow version 2.18.0 fails to handle multidimensional arrays, leading to an "UnimplementedError" due to unsupported broadcasting. This issue arises specifically when input tensors have more than two dimensions, despite the function working correctly in NumPy. The problem highlights a gap in TensorFlow's compatibility with NumPy's functionality.
- Discrepancies in TensorFlow Layer Behavior on GPU vs CPU: The
tf.keras.layers.LSTMandkeras.layers.GRUfunctions exhibit inconsistent behavior when executed on GPU compared to CPU. The LSTM layer fails when using XLA on GPU, while the GRU function produces different output lengths due to the cuDNN-optimized kernel. These discrepancies necessitate disabling certain optimizations to achieve consistent results across platforms.
- TensorFlow Crashes and Errors Due to Invalid Arguments and Configurations: Several TensorFlow operations crash or produce errors due to invalid arguments or misconfigurations. The
LearnedUnigramCandidateSampleroperation crashes with an "Aborted (core dumped)" error whennum_trueis set too high, causing an integer overflow. Additionally, a custom loss function in TensorFlow 2.18.0 fails to receive all expected outputs, andtensorflow_hub.KerasLayeris not recognized in TensorFlow 2.8, leading to aValueError.
- Challenges with TensorFlow Lite on Android and Linux Systems: Users face difficulties with TensorFlow Lite on Android and Linux systems, including profiling operator performance and running benchmarks. The latest TFLite version does not display all operators in the Android Studio CPU Profiler, and attempts to run benchmarks using the QNN delegate on Android devices fail. Additionally, there is a need for support in utilizing XNN Pack to enhance model performance on x86 CPU/Intel GPU within a Python environment on Linux.
- TensorFlow GPU Crashes and Compatibility Issues: TensorFlow experiences crashes and compatibility issues on various systems when utilizing GPU resources. On an Apple M3 Max device, TensorFlow 2.16.2 crashes due to potential memory allocation problems with the Metal backend. Similarly, TensorFlow 2.18.0 on a Manjaro Linux system with CUDA 12.4 and an RTX 4090 GPU fails to run programs that work on Windows, likely due to CUDA environment misconfigurations.
- Import Errors in TensorFlow on Windows: Users encounter import errors in TensorFlow on Windows 10 with version 2.18 and Python 3.12. The issue involves a DLL load failure when trying to import TensorFlow, specifically due to a failed initialization routine of a dynamic link library. This results in an ImportError related to the '_pywrap_tensorflow_internal' module, hindering the use of TensorFlow on the affected system.
2.4 Closed Issues
This section lists, groups, and then summarizes issues that were closed within the last week in the repository. This section also links the associated pull requests if applicable.
Issues Closed This Week: 26
Summarized Issues:
- Build Errors on Linux and Windows: Users have reported various build errors when compiling TensorFlow from source on both Linux and Windows systems. On Linux, issues include missing TensorRT headers despite excluding TensorRT and a "No such file or directory: 'patchelf'" error on CentOS 8. On Windows, errors involve the
CreateProcessWfunction and misconfigurations with MSYS2. These issues highlight the challenges of configuring and building TensorFlow in different environments.
- Import Errors and DLL Load Failures on Windows: Numerous users have encountered ImportErrors and DLL load failures when using TensorFlow on Windows systems. These errors are often linked to missing MSVC redistributables, unsupported AVX2 instructions, or incompatible CPU architectures. The issues persist across different Windows versions and Python environments, indicating a widespread compatibility problem.
- issues/tensorflow/tensorflow/issues/74405
- issues/tensorflow/tensorflow/issues/74725
- issues/tensorflow/tensorflow/issues/76961
- issues/tensorflow/tensorflow/issues/77387
- issues/tensorflow/tensorflow/issues/79609
- issues/tensorflow/tensorflow/issues/80293
- issues/tensorflow/tensorflow/issues/82289
- issues/tensorflow/tensorflow/issues/82992
- issues/tensorflow/tensorflow/issues/83170
- Feature Requests and Enhancements: Users have proposed several feature requests to enhance TensorFlow's functionality. These include a reverse operation for
tf.image.extract_patchesinspired by the Swin Transformer and a TensorFlow 1.x wheel file for aarch64 architecture. Such requests aim to improve TensorFlow's usability and extend its capabilities for specific use cases.
- Compilation and Runtime Errors: Users have faced compilation and runtime errors across different platforms and versions. These include a type mismatch in the
ResolvePaddingfunction, a TensorFlow runtime error on Windows due to DLL load failure, and a JIT compilation failure on a Radeon GPU. These issues highlight the complexity of maintaining compatibility across diverse hardware and software configurations.
- TensorFlow Lite and Model Size Discrepancies: A user reported an issue with TensorFlow Lite where an exported int8 model was unexpectedly larger than the original model. This discrepancy raises concerns about the efficiency of model conversion processes and the need for better documentation or tools to address such issues.
- Platform-Specific Bugs and Compatibility Issues: Users have encountered platform-specific bugs, such as a missing symbol in the Metal plugin on macOS and a TypeError with TensorFlow Lite support libraries. These issues underscore the challenges of ensuring cross-platform compatibility and the need for thorough testing across different environments.
- Installation and Version Compatibility Problems: Users have faced installation issues due to version compatibility problems, such as attempting to install TensorFlow 2.10.0 on Python 3.13. These problems highlight the importance of clear version compatibility guidelines and the need for users to adhere to supported configurations.
- Spam and Irrelevant Content: The repository has also been targeted by spam, such as posts promoting unrelated websites. These issues are typically closed quickly to maintain the focus on relevant discussions and problem-solving.