
So here’s the thing; over the past week or so, we have been busy writing a series of articles exploring Google’s extensive AI Platform. Our goal? To show data scientists out there what’s possible, share our personal experience working with the platform, and provide easy to follow, hands-on demonstrations.
Here are the first two articles in the series:
Google Cloud AI Platform: Hyper-Accessible AI & Machine Learning
Google Cloud AI Platform: Human Data labeling-as-a-Service Part 1
So it came as quite a surprise when we noted this rather brief(!) release note from Google Cloud:

"Unified" we thought. "What on earth does that mean?!". "Please don’t tell us our articles are obsolete before the ink has even dried…"
What is Google AI Platform Unified?
Well, the official Google documentation describes it is as:
"AI Platform (Unified) brings AutoML and AI Platform (Classic) together into a unified API, client library, and user interface."
Initially, we were a bit confused with this – AutoML has always been part of AI Platform, at least in the official documentation. In fact, we had been using this image from Google Cloud to help summarise AI Platform services for a while now. And note it includes AutoML!

So clearly a bit more investigation was needed.
AutoML Tables
I thought I’d start by looking at a recent AutoML Tables Linear Regression model we had trained for a client project. I navigated to the AutoML Tables service via the usual link https://console.cloud.google.com/automl-tables.
You might be wondering why I use a url instead of searching for it in the Cloud Console service search bar? Well, Google have made this quite hard to find, and searching for "AutoML", "Tables", "AutoML tables" doesn’t find it!
Ironically, if you are looking for AutoML Natural Language, again nothing is found if you search for AutoML. Google, please fix this!
Now the first change; I’m presented with a big banner informing me these services are now available in the AI Platform (Unified).

Interestingly, this is optional, and all of my models are still displayed as before. I can also still train new models and upload datasets. Although Google advises any new models should be created on the unified version.
AI Platform (Unified) UI
I know, quite a mouthful that one! We assume Google is going to drop the unified thing at some point. Who knows!
So clicking on the banner above, and here are all of my AutoML Tables models, proudly sitting in their shiny new home.
Oh, erm, actually they’re not. Where are my models!

That would be far too easy, wouldn’t it?! It turns out, that all models and datasets need to be migrated from classic to unified. OK, let’s give that a go.
The first thing to note is Google lists a number of AutoML services that CANNOT be migrated. These are documented here. We were a bit disappointed to see this includes AutoML Natural Language models, as explained here:

The good news for our AutoML tables models, however, is these are supported. Huray! Although note below, there are some restrictions.

Migrating AutoML Tables into AI Platform Unified
OK so here‘s what I needed to do. These steps and more are available here.
-
Navigate to AI Platform Unified in the Google Cloud Console: https://console.cloud.google.com/ai/platform
-
Google have provided a handy migration tool. To access the tool, change the region dropdown to match the region where your models are located (ours are in us-central1). You should see Migrate to AI Platform (Unified ) displayed at the bottom of the screen:

Tip: Don’t see this? Either you have no resources in that region, or, the resources you do have are not compatible with unified (e.g. an AutoML Text Classification model).
- Click the SET UP MIGRATION button.
- I now see our AutoML Table models listed. That’s nice and easy so far. Click NEXT.

- This takes you to a review page. Here it details what is about to be migrated, in our case, 6 models and 7 datasets.
Tip: Nice to see that there is no charge for this migration. Another thing we liked, is a copy of your datasets and models is made during the migration process, so at least you can revert back to your original models if required.
- Click MIGRATE ASSETS. You should see this banner displayed, informing you the migration is in progress and that it can take up to an hour. An email is sent when complete.

Migrating our 7 models and 6 datasets took just 1 minute. Nice!
- Once complete, simply refresh the page and now you should see your models and datasets in their new home.
Our migrated datasets:

Our migrated models:

Viewing models in AI Platform Unified
The good news is we really like the new UI for viewing models. We will use one of our linear classification models as an example.
This was the (classic) AutoML tables view of the model:

Now this looks like this:

Note the following changes in the new UI:
- The IMPORT tab has gone. Importing data has now moved to Datasets.
- The TRAIN tab has also gone. This perplexed us for a while – as a reminder, here’s what the training tab looked like for this model in AutoML classic:

At first we thought this must have moved to the new Training item in the main menu in unified. But, clicking on the training pipeline for our model just took us back to the model page. Strange? We’ve asked Google and will see what they say.
- The MODELS tab has been replaced with the new MODEL PROPERTIES tab, shown here:

- We really like the new EVALUTE tab. Presents the key performance stats on our model really well:



Tip: The Feature importance (image above) is a great feature of AutoML Tables as it clearly shows how much importance the model is placing on your features. We often see some surprising results (proof that human bias can be flawed!). Great to see this is still here in unified.
- The TEST & USE tab in AutoML Tables Classis (below):

Has been reallocated in Unified as follows:
TEST & USE / BATCH PREDICTION has moved to its own BATCH PREDICTION tab in Unified.
TEST & USE / ONLINE PREDICTION has moved to its own DEPLOY AND TEST tab:

TEST & USE / EXPORT YOUR MODEL has moved to DEPLOY AND TEST (above) and is also available from the new Export link (below):

This lets you export your model as a Tensorflow model, exported to Cloud Strorage (GCS). This allows you to run predictions via BigQuery ML, which is a great feature that we use a lot as a team.
Conclusions
AI Platform (Unified) certainly came as a surprise, and I’m sure we won’t be the only people to share this view. At first it was a little confusing from the documentation (and one liner release note!), exactly what unified offered over classic. Hopefully, we’ve saved you some time digging through the documentation and made this a bit clearer for you.
The migration tool was easy to use, as is hopefully evident from our hand-on example. Once we started to migrate our AutoML models, the rationale to integrate AutoML tighter within the AI Platform started to make sense; and we generally like the changes, especially the cleaner UI and colocation of other AI resources such as the AI Platform Notebooks.
The training section we are still a little confused with – so we’ll let you know what we hear back from Google on this one.
In a follow-up article, we will show you what training a model now looks like in AI Platform Unified, and we’ll compare this with AutoML Tables. Should be interesting!
Next steps
- Keep an eye out for Part 2, where we will explore training a model.
- Read the Google Cloud AI Platfrom Unified documentation
- Learn more about Ancoris Data, Analytics & AI