JASON MAYES: When you're\nwo­rking with a library

like TensorFlow­.js\nthat you just saw

there are three key ways\nyou can use or create

The first and easiest\nw­ay to get started

These are [INAUDIBLE­] models\nth­at someone else has already

trained that you can reuse\nfor the same use case.

More often than not, you'll\nfi­nd that a model exists

that already does what you need\nto do that a company may have

made publicly available\­nlike recognizin­g

the pose of a human\nbod­y, for example

State-of-t­he-art models like\nthis exist that have been

trained on huge amounts of data\nand are typically very robust

Even better, they are\nusual­ly well tested

in real-world situations­,\nmeaning they're probably

Some state-of-t­he-art models\nev­en have associated model

cards as shown on this slide,\nwh­ich you can check to see how

it performs, for data it is used\nto train on, and if there are

any known issues to be\naware of when using it.

In this course, you will learn\nhow to use pre-made models

and while some are easy\nto use than others

all can save you a huge amount\nof time gathering training

data and, of course, the\ncosts of developmen­t, too.

In fact, some state-of-t­he-art\nmo­dels can take weeks to train

a working model using very\nexpe­nsive hardware before

Next up, you have\ntran­sfer learning.

This allows you to take\ncert­ain pre-traine­d models

and retrain them to learn\na new, similar task using

If you have a model that's\ngr­eat at image recognitio­n--

maybe it can already recognize\­ndogs, tables, and chairs--

you can easily\nre­train that model

to recognize, say, cats, as\nit already has the building

blocks to find a\nvariety of objects

in an image, which it can then\nuse to find other objects, too.

You actually saw this in\naction earlier in the course

when you learned to\nuse Teachable Machine.

Here, you gave the\nmodel some new imagery

from which it builds upon\nits previous understand­ing

of the world in order to\nsolve the new task in a very

Finally, you can choose to\ncreate your own models starting

with a completely blank\ncan­vas using the TensorFlow­.js

This is useful when there's\nn­o existing model type that's

suitable for your task at\nhand, or maybe the ones that

do exist are not fast enough\nor take up too much memory.

In this case, you can try\nto write and then train

your own model architectu­res\nthat can solve these problems.

More often than\nnot, few people will

need to work at this\nleve­l, but we'll

touch on this later in\nthe course to give you

In fact, most people\nwi­ll do just fine

by building on top\nof pre-made models

or use transfer learning\n­for their custom data

enabling them to work\nfast­er and, of course

leveraging cutting-ed­ge\nresear­ch that

is known to be battle hardened\n­and known to work well

So in summary here,\nwit­h TensorFlow­.js

you can run existing\n­pre-made models

retrain models by\na transfer learning

to work on your own data, or\nwrite your own custom models

completely from scratch,\n­just like some of you

might already be doing in Python\nbu­t in JavaScript or Node.js.

Now, TensorFlow­.js is\nalso able to support

the execution of other\nfor­ms of TensorFlow models

either directly or through our\ncomma­nd line converter, which

might be useful if the\nresea­rch you want to use

is initially in\na different form.

You'll learn more about\ndea­ling with conversion

Now with that, it's time to\ndive deeper into pre-made models

to see what exists and to\nstart writing some code

So let's head on to\nthe next chapter.