The wonders of the new version of Tensorflow (1.2RC0)
Tensorflow recently released a new version with high-level APIs.
In a recent post Basic neural networks patterns it is worth for a researcher to know, by platform, I explained my interest for tensorflow and how it now integrates tf.contrib.learn and tf.contrib.keras (soon tf.keras) as contribution.
At the moment, with its new version 1.2RC0 presented at the Google I/O 2017, the annual developer conference in May 17–19, 2017, Tensorflow overpasses a lot of challenges to bring AI research to another level via its high-level APIs.
From a set of hours of videos from this conference, I retrieve the most relevant elements to share with you.
(Note: These functionalities are available only with the 1.2RC0+ version of Tensorflow. Original videos are linked in the titles.)
I- EFFECTIVE TENSORFLOW FOR NON-EXPERTS
We have now a set of advanced methods to easily implement, train, evaluate, predict and save models.
- New training next batch function in action : at 08mn,15s of the video.
- New Experiment function (17mn,00s) : allows to automatically try a lot of values for the training process as with CrossValidation and pick the best parameters.
- Setting up a cluster: github.com/tensorflow/ecosystem
- Full implementation code example available here : goo.gl/0OgXiL
- New Video Processing With Tensorflow : (27mn,06s). A new video API easily makes video analysis, such as Question-Answer from video, i.e ask a question about a video and automatically receive the correct answer.
It turns video frames into a vector, with pre-trained representations, allowing an access at any sequence (time) of the video for analysis and answers.
And how to code it ? Just see an example from the 34th min of the video as below.
II- FROM RESEARCH TO PRODUCTION WITH TENSORFLOW
With its new version Tensorflow tools allow researchers to easily turn their implementations into production :
+ Answer online requests at low latency
+ Serve multiple models and versions in a single process: now do not care about your models (libraries, binaries, inference APIs, CPU-GPU-TPU, Mobile, etc.), they will be served the same way. Serving is how you apply a Machine Learning model after you have trained it.
+ Achieve efficiency of mini-batching from training with requests arriving asynchronously.
Below some resources and tutorials to perform it.
- Project homepage : tensorflow.github.io/serving/
- K8s: tensorflow.github.io/serving/serving_inception.html
- Cloud ML: cloud.google.com/ml/
- Mailing list for feedback and questions : discuss@tensorflow.org
- Stack Overflow: #tensorflow
III- OPEN SOURCE TENSORFLOW MODELS
Here you will find a lot of examples of implementations, including the new Keras branch of Tensorflow, but also some pre-trained models available for free for researchers and anyone too.
Tensorflow Keras code samples
- Keras-tensorflow-workshop
Codelabs
- Image classification with tensorflow
- Deploy a custom Tensorflow image recognition model on an Android app
- Machine Learning Recipes with Josh Gordon
Shared research in tensorflow
1- Neural Audio Synthesis, Music Generation : magenta.tensorflow.org
2- Sequence to Sequence : The new Open Source Sequence-to-Sequence Framework in TensorFlow (tf-seq2seq) with clean and modular codebase, full test coverage maintaining and documenting of all its functionality.
3- Improving Inception : Improving Inception and Image Classification in TensorFlow.
4- Parsey Saurus (a set of new upgrade to SyntaxNet pre-trained models in Tensorflow) : A neural-network framework for analyzing and understanding the grammatical structure of sentences. A collection of pre-trained models for more than 40 languages which overcomes the challenge of building machine learning systems that work well for languages other than English.
As we have now these very useful improvements on the Tensorflow framework, anyone no matter its specific field could easily use it for applied research and deploy it. Congratulations to the team !
Thanks for reading this. If you liked it, please click the 💚 below so other interested people will be informed of it here on Medium.
