Tensorflow new document publishing: add CLP, dtensor State of the art models are ready

Arrangement | Phoebe Coordinating editor | Du min
Produce | CSDN（ID：CSDNnews）

In recent days, ,TensorFlow Many exciting new documents have been posted on the official website , Include model repair 、 Model parallelism and TensorFlow Lite and TensorFlow Model Garden Update . Let's see what new things you can harvest ！

Model repair ： Counterfactual Logit pairing （CLP）

TensorFlow The AI team of has added a new model repair technology , As part of its model repair Library .TensorFlow The model repair library provides training techniques to intervene in the model , Such as changing the model itself by introducing or changing the model objectives . The first model repair technology initially introduced ：MinDiff, It minimizes performance differences between two slices . And this new technology is counterfactual Logit pairing （CLP,Counterfactual Logit Pairing）. This technique ensures that when sensitive attributes referenced in an example are deleted or replaced , The prediction of the model will not change . for example , In a toxicity classifier , Such as " I am a man " and " I am a lesbian " Such examples should be equal , Will not be classified as toxic .

Basic tutorial ：https://www.tensorflow.org/responsible_ai/model_remediation/counterfactual/guide/counterfactual_usage_steps

Keras course ：https://www.tensorflow.org/responsible_ai/model_remediation/counterfactual/guide/counterfactual_keras

API Reference resources ：https://www.tensorflow.org/responsible_ai/model_remediation/api_docs/python/model_remediation/counterfactual

Model parallel ：DTensor

DTensor Provides a global programming model , Allows developers to manipulate tensors on a global scale , Manage the distribution across devices at the same time .DTensor Through a named “ Single program multiple data ”（SPMD） Extended programs , Distribute the program and tensioner according to the slice instruction .

By decoupling the entire application from the sharding instructions ,DTensor Can be on a single device 、 Multiple devices 、 Even running the same application on multiple clients , While preserving its global semantics . If you remember TensorFlow1.0 Medium Mesh TensorFlow,DTensor Can solve Mesh Can solve the same problem ： The training model may be larger than the single core .

TensorFlow 2.9 Will always be built at night DTensor stay tensorflow.org Go public . Even though DTensor It's experimental , But you are welcome to try .

DTensor guide ：https://www.tensorflow.org/guide/dtensor_overview

DTensor Keras course ：https://www.tensorflow.org/tutorials/distribute/dtensor_keras_tutorial

API Reference resources ：https://www.tensorflow.org/api_docs/python/tf/keras/dtensor

TensorFlow Lite New functions

TensorFlow Lite The website has made some major changes , Including getting started documentation .

• Build the way for developers to grow

First ,TensorFlow Now press platform （Android,iOS And other equipment ） To provide a learning guide for developers , It makes it easier for developers to run on the corresponding devices TensorFlow Model .TensorFlow Not only for Android Users have added new learning roadmap and quick start guide , Earlier ,Google Play In service TensorFlow Lite The new beta version of has also added a guide . These quick start guides include Kotlin and Java Examples in , And the example code is upgraded to CameraX.

• Model maker

TensorFlow Lite Model Maker The library simplifies training with custom datasets TensorFlow Lite The process of modeling . It uses transfer learning to reduce the amount of training data required and reduce training time , Seven common tasks are preset , Including image classification , Object detection and text search .

TensorFlow Added a for text search New tutorial . This type of model allows you to accept a text query , And search the most relevant items in the text data set , Such as web database . On mobile devices , You can use it for automatic reply or semantic file search .

Besides ,TensorFlow Also released a complete Python Library reference .

• TensorFlow Lite Model page

Finding the right model for use cases can sometimes be maddening .TensorFlow Wrote more about How to choose the right model for your task , And the factors that need to be considered in making this decision . You can also find model links to common use cases here .

TensorFlow Model Garden： State of the art models are ready

TensorFlow Model Garden For visual and naturallanguageprocessing （NLP） Provides many of the most advanced machine learning （ML） Model implementation , And workflow tools , Allows you to quickly configure and run these models on standard datasets .Model Garden Covers both visual and textual tasks , And one called Orbit Flexible training cycle library . The model is equipped with pre built configuration to train to the most advanced level , And many useful specialized operations .

TensorFlow Will start recording your use of Model Garden All kinds of great undertakings , The first stop from this should be to understand Model Garden Overview of 、 The list of available models and the image classification tutorial begin .

Model Garden Overview of ：https://www.tensorflow.org/guide/model_garden

List of available models ：https://github.com/tensorflow/models/tree/master/official

Image classification tutorial ：https://www.tensorflow.org/tutorials/images/classification_with_model_garden

Other exciting new changes

Besides , Google and CSIRO cooperation , utilize TensorFlow The code base will strengthen the monitoring of harmful species on the Great Barrier Reef . adopt Kaggle competition , Machine learning developer collaboration training ML Model , To identify outbreaks of spiny crown starfish that degrade coral reef ecosystems . To protect the offspring of the Great Barrier Reef .

Another new change is about TensorFlow Compressed new tutorial , It uses neural networks for lossy compression . This example uses something like an automatic encoder to compress and decompress MNIST.

Picture source TensorFlow Official website

In the above update , What surprises you ？ Welcome to communicate with you ~

Reference material ：https://blog.tensorflow.org/2022/06/new-documentation-on-tensorfloworg.html