This article is reproduced from the AI Institute,Original address
Lei Feng Network AI Technology Review Press: TensorFlow is currently the most popular deep learning library, it is an open source artificial intelligence learning system. Tensor means tensor, which represents an N-dimensional array; Flow means a stream, which represents a calculation based on a data flow graph. The process of flowing an N-dimensional number from one end of the flow graph to the other is the process of analysis and processing by an artificial intelligence neural network.
Recently, Pagge Bailey (@DynamicWebPaige), a Goolge AI engineer and active advertiser of Google AI, summarized the main functions of TensorFlow's 30. The AI Technology Review of Lei Feng Network is organized as follows.
In the past few years, the knowledge system about TensorFlow has taken over my brain. TensorFlow does not have many features compared to other updated frameworks.
I have devoted my thoughts to this product and wrote down my thoughts. Here is a list that expands one by one. Are you ready for this?
1.TensorFlow Extension (TFX)
Everyone knows that I particularly like to use TFX and its full set of tools to deploy machine learning models into production environments. If you care about keeping models up to date and monitoring them, then you can take a look at the product and see its papers.
If you want to train your model on a small dataset, or improve generalization, you will need to use migration learning. The TFHub module makes it easy and can be used in the open source software store at https://tfhub.dev/.
The address of the TF Hub is:Tensorflow.org/hub/
3.TFX data validation
How to automatically ensure that the data used to retrain the model has the same format, source, naming conventions, etc. as the data originally used to train the model.
For online training, this is a lot of work!
4.TFX -TensorFlow transform
Similarly, you may want data for retraining to be automatically preprocessed: normalize specific features, convert strings to values, and more. Transform can not only do these operations on a single sample, but also batch data.
the website is:https://www.tensorflow.org/tfx/transform/?hl=zh-cn
5.TFX model analysis
I like to use the TensorFlow model analysis function to check the input data of the model or the problem that may occur on a small part of the data in the model reasoning process. I can use it to check the data carefully to make sure that all categories of data are not negatively affected.
the website is:https://www.tensorflow.org/tfx/model_analysis/?hl=zh-cn
The service makes it easy to deploy new algorithms + experiments, but still maintain the same server architecture + API. It can not only directly support models on TensorFlow, but also support other models.
the website is:https://www.tensorflow.org/serving/?hl=zh-cn
TensorBoard is a very cool visualization tool on the TensorFlow framework, and it comes directly with TensorFlow. It can visualize the log of the model running process, and has its own display panel for scalar, histogram, distribution, graph, image, audio and so on.
the website is:https://t.co/CEVbcJTHLP?amp=1
8.TensorFlow Lite (#TFLite)
Use #TFLite to deploy models on mobile phones and embedded devices. If you see an app on your Android phone that detects plant leaves with disease, or a small, AI-skilled robot, they are likely to use #TFLite.
the website is:https://t.co/suCsBIeQz4?amp=1
Swift on 10.TensorFlow
Swift can catch errors of type and shape mismatch before running the code, and has built-in automatic differentiation. It brings eager execution capabilities and increases the usability of TFs. I still need to use this more.
Keras is now integrated directly into TF, which is tf.keras. This means that if you don't want to use a low-level model, you can still use the user-friendliness of high-level APIs to build the graph + model. The 2.0 version will have more features!
Tensor2Tensor is an open source software library of deep learning models and datasets that make deep learning easier to use and facilitate machine learning research. It also provides a high-level guide on when and why to deploy these models.
13.XLA (linear algebra calculation acceleration)
XLA is a linear algebra-specific compiler that optimizes how TensorFlow is calculated. The result is improvements in computing speed, memory usage, and portability of mobile platforms.
But you must have a hardware accelerator first!
14. Edge TPU
A small ASIC that provides high-performance machine learning reasoning for low-power IO devices. For example: edge TPU You can perform state-of-the-art mobile vision models, such as executing the MobileNet V100 model at 2+fps, while also saving power.
15.Magenta (Leave the original English text here, Chinese translation does not know why...)
As a musician, Magenta made me very happy.
You can map 8 key inputs to an 88 key piano, automatically create melody accompaniment, use machine learning to display music visuals, transcribe tunes, produce new sounds, and more.
16. Seed Library
This feature is rarely seen by anyone.
The seed library is a growing collection of interactive machine learning examples that you can use, modify, experiment with and complement to satisfy your needs + use case studies. In the machine learning project of the seed bank, there are even examples of fairness and prejudice!
The address of the seed library is:https://research.google.com/seedbank/
17.GoogleColab Analysis Tool
This is not a unique tool for TensorFlow, but it is a very good tool and I have to mention it! It is an interactive Python notebook that can be used free of charge and can be used in CPU/GPUSwitch between /TPU or local/remote backend!
Deep learning is good, but as a data scientist, you may want to tell your model some domain-specific knowledge: Monte Carlo, variational inference, Bayesian techniques, vector quantization autoencoders, and more.
the website is:https://www.tensorflow.org/probability/
19. Model Park
This is a large collection of open source models of the GoogleAI and TensorFlow communities, including samples and code snippets. From the tree to the synthesis of neuron programs.
the website is:https://github.com/tensorflow/models
This is an easy-to-access framework for prototyping reinforcement learning algorithms. Focus on: simplicity, flexibility, reliability and reproducibility of the experiment.
Remarks: Google official has not officially released this product!
The Dopamine team even developed a series of GoogleColab notebooks to show how to use this product! This user-friendly approach also makes the framework more versatile.
Nype is a Python and C++ code base designed to make it easier to read, write, and analyze common genomics data such as SAM or VCF. It works seamlessly with TensorFlow.
For the 2018 TensorFlow Developer Summit, check out the You Tube video:
22.TensorFlow Research Cloud (TFRC)
This is a cluster of 1000 Google Cloud TPUs that provide a total of 180 megapascals of computing power to the machine learning community – and it's absolutely free, making its own contribution to the next breakthrough in machine learning.
This is not a specific product, but it is critical to the TensorFlow ecosystem.
Google AI's new focus on the community is: Edd-led feature mailing lists, social media, special interest groups, and TensorFlow directly enter new/changing features.
Do you know that all our files are placed on @GITHUB? Welcome the contributions and suggestions from all walks of life! Go and ask @billylamberta to learn how to get started!
Fly_upside_down, rstudio, and fchollet create an R interface for developers that uses the advanced #Keras+Estimator API and provides more control when you need to tune your network at a lower level.
There is even a book about this:https://tensorflow.rstudio.com/
An algorithm for structurally learning/optimizing weights for adaptive learning of deep neural networks. If you want to learn more about the adaptive machine learning kernel, AdaNet's tutorial is a good starting point!
Papers on AdaNet:https://arxiv.org/abs/1607.01097
Interpretability, that is, how deep neural networks make decisions, is critical for ethical machine learning and the use of deep learning in scenarios that have a significant impact.
https://github.com/tensorflow/lucid Here are GoogleColab tutorials, code, and distillpub articles.
28. Concept Activation Vector Test
A similar truth: Most interpretable methods display important weights in each input feature (for example, a pixel). Instead, TCAV shows the importance of high-level concepts (eg, color, gender, ethnicity), ie how humans communicate.
If the performance of your model is highly dependent on the input data, you can destroy these models by manipulating or contaminating the data. On @goodfellow_ian's blog you can see how to use cleverhans to provide a benchmark library for vulnerability assessment against attacks!
the website is:https://github.com/tensorflow/cleverhans
30.Rust + a combination of Haskell and C API
I mentioned rstats support earlier and want to make sure that other community projects are also mentioned (TensorFlowSharp was created by MiguelDigias).
Can view the URL: