DeepLearning Images Revision M19. TensorFlow 2.0 experimental.

You probably have heard TensorFlow 2.0 is coming. Here is a nice overview of what exactly is coming. Some of you have also spotted that there are nightly builds of TensorFlow 2.0 are available. Even more, I know that some of you folks already have started using them! And not only using, there is at least one existing Deep Learning course build entirely around TensorFlow 2.0 (author of the course: Aurélien Geron)!

BTW if you do not know what is Deep Learning VM, read this first.

A lot of people right now, when they hear about TF2.0 are in a state like this:

Now, before we proceed, full disclosure: nightly builds of TensorFlow 2.0 are not supported! This release of Deep Learning images using nightly builds, therefore you should not expect anything to work, however, if you brave enough to proceed, I have a red pill ready for you.

Now the very best part. Our new Deep Learning Images not only have TensorFlow pre-installed, but they also come with the full course from Aurélien Geron pre-baked in! So you can start playing with TensorFlow 2.0 right away.

So, let’s begin.

Create TensorFlow 2.0 VM

In order to create TensorFlow 2.0 VM you just need to run the following command (after installing and configuring Google Cloud SDK):

You can find more details about the command and other ways to create a VM in the main article.

Connect To The VM

As soon as VM is up, wait a minute (just to make sure that VM is up and Nvidia drivers are installed) and connect to it:

There are some alternatives ways of connecting to the VM can be found here.

Run The Course

You can find Aurélien’s course in:

tutorials/tf2_course

Let’s play with very first TF2.0 notebook:

And do not forget time to time pull latest changes of the course:

Now, one can ask, this is nice, but can I use my TF 1.* code with TF 2.0? Well, it is somewhat complicated…

Convert Existing TF 1.* Code To TF 2.0

In order to covert code from TF 1.* to 2.0 there is a converting tool tf_upgrade_v2. It is already pre-installed on our VMs and can be used like this:

tf_upgrade_v2 --infile foo.py --outfile foo-upgraded.py

Execution of the line above will convert file foo.py into TF 2.0 compatible foo-upgraded.py.

How To Get Help

TF2.0 bugs: https://github.com/tensorflow/tensorflow/issues

Converting tool bugs: https://github.com/tensorflow/tensorflow/issues

Course bugs: https://github.com/ageron/tf2_course/issues

Deep Learning images bugs/feedback: https://groups.google.com/forum/#!forum/google-dl-platform

Experimental Image Families

With this release our images families map now looks like this: