PRIYANKA: Hi, I'm
Priyanka [INAUDIBLE], and this is AI
Simplified, where we learn to make our data useful. In this video, we're going
to imagine a company that has a large amount of
data which they want

to use to make
meaningful predictions, and take their business
to the next level. And we also know that their
team has varying levels of machine learning expertise. To accelerate this
AI innovation, they are looking
at Vertex AI, which

provides tools for every step
of the machine learning workflow from managing data sets to
different ways of training the model,
evaluating, deploying, and making predictions. In this episode, let's
see the first step of the machine
learning lifecycle,

which is the data sets. You don't need to
worry about data sets. If your use case
is generic enough, you can just take advantage
of the pre-trained machine learning APIs. And to learn more about them,
I've included a link below.

In order to create custom
machine learning models, we must first have
a collection of data to train our models with. That is what data sets are for. Data sets make data discoverable
from a central place. They provide the ability to
annotate and label the data

right from within the UI. You can track the lineage
for data governance and can compare model
metrics between the models. Now, to create a data set,
we first select our data and upload it, import it,
into the managed data sets.

And then we can make
modifications and use it to start training the model. Currently, there are four
supported data types -- image, tabular,
text, and videos. Now, image data sets currently
support image classification, which is where models
predict one or many labels

from an image. Now, for example,
identifying the types of dog treats from images of treats. Object detection is where the
model draws bounding boxes around the items in the image. For example, you're
identifying the location

of different vegetables
within a picture of salad. Image segmentation
is where the model can assign a label to pixel
level regions in an image. To ensure your model
performs well in production and there's no
training serving skew, make sure the
training images are

close to the input
your users will send to your deployed model. For example, if your images
are going to be low resolution, be sure to have blurry
images, low resolution images, in there. Include multiple
angles, backgrounds,

and all sorts of resolutions. It's recommended that you
include at least 1,000 images per label. But, you can always get
started with 10 per label. The more examples you provide,
the better your model will be.

Now, tabular data sets
currently support regression, where you predict a numerical
value, and classification, where you predict a
category associated with a particular example. Forecasting is where the
model predicts the likelihood of sudden events or demands. For example, you want to predict
sales over a period of time.

Now, tabular data sets support hundreds of columns and millions of rows. Text data sets currently support classification, which is where you assign one or more labels to an entire document. In a customer support example, an email could include classifications as great service or a suggestion.

Now, entity extraction is where the model identifies custom text entities within a document. For example, maybe you want to identify all the phrases in a customer support email that reference pricing. Too expensive or great value. Sentiment analysis is to understand

the overall sentiment expressed in a block of text. For example, if a customer was happy or upset or frustrated. Now, will your data sets currently support classification, which provides labeled predictions for an entire video, shots, and frames?

For example, in a football game, you would be able to label the section with commercials and the section with the game. Now, action recognition is where the model identifies clips in the video where specific actions occur. For example, in a puppy football game, the action of a dog

making a touchdown. Now, object tracking is where you get labels, bounding boxes, and time stamps for objects you want to track in a video. For example, in the same game, you want to track the football's motion throughout the video.

Now, we just covered the basic functionality here. I have included the link below for more details and best practices on data preparation. At this point, I think we want to see how to create and manage the data sets in the console.

So let's dive in. In the console, we get to Vertex AI and click on the data set where we see image, tabular, text, and video data sets for the different objectives we talked about earlier. Let's try to create an image data set.

Say we want to do image classification. Before we begin, it's good to check the Data Guide that shows us the requirements and recommendations. You can import files directly from your computer, and they would be stored in Cloud Storage.

We can then add the corresponding labels for our images in the next step. If you already have labels, use the Import File option to import a CSV with your image URLs and their labels. The entries in the CSV would look something like this.

If your data is not labeled, and you would like human help to label it, then you could use the data labeling service. Let's upload a few files here. Once the files are uploaded, we can create labels and assign them to the images.

We can also analyze the images we have in the data set, number of images per label and a few other properties. Depending on the type of data, your options might vary slightly. For example, if you look at tabular data,

you could upload a CSV file from your computer, from Cloud Storage, or select a table from BigQuery directly. Once you select the table, the data is available for analysis. We can see the number of rows, columns, etc.

You would do the same
for a text data set. Once uploaded, we
can see the labels or add new labels to the data. And we can also
analyze a data set. The video data set
interface looks similar. I have some labeled sports
files here for golf, cartwheel,

horse riding, and others. That was a high
level overview of how you can manage different
data sets in Vertex AI. In the next video, we will look
at the next steps of the machine learning workflow to
build and train ML models.

In the meantime, let's
continue our discussion in the comments below. I am excited to hear all
about your ML use case and your data sets.

How to manage ML datasets with Vertex AI

Welcome to the official YouTube channel for Google Cloud’s technical community. Explore our how-to videos, see exciting demos, stay up to date with product news, and find new developer tools helping you build what’s next.


How to manage ML datasets with Vertex AI

Scripts