WEBVTT - This file was automatically generated by VIMEO

0
00:00:07.680 --> 00:00:12.500
Hi. In this video we will learn more about tune search Cross validation.

1
00:00:15.950 --> 00:00:20.000
Tune search is a randomized search from library called Array,

2
00:00:20.580 --> 00:00:25.480
and it's main benefit is that in addition to purely randomized

3
00:00:25.480 --> 00:00:30.280
search in a sticky learn, we can specify various optimizers.

4
00:00:31.910 --> 00:00:36.570
So basically the search for the best hyper parameters may not be

5
00:00:36.570 --> 00:00:37.410
completely random,

6
00:00:38.370 --> 00:00:43.270
but there is some logic behind it where some parameter combinations

7
00:00:43.490 --> 00:00:46.270
are thrown away because behind the scenes,

8
00:00:46.490 --> 00:00:51.190
the optimizer evaluates that continuing exploring parameters in a

9
00:00:51.390 --> 00:00:56.120
specific direction doesn't make sense and that some combination of

10
00:00:56.120 --> 00:00:59.160
parameters are more promising to explore more.

11
00:01:00.770 --> 00:01:01.603
Using Ray,

12
00:01:01.820 --> 00:01:06.440
we can also scale our hyper parameter optimization to multiple

13
00:01:06.440 --> 00:01:11.340
machines as well. Here,

14
00:01:11.720 --> 00:01:14.820
uh, we put more parameters, uh,

15
00:01:14.840 --> 00:01:19.500
to test and replace the sampling distributions from SciPi Library

16
00:01:19.720 --> 00:01:22.980
to tune methods. Instead of, um,

17
00:01:24.540 --> 00:01:25.880
an inter parameter,

18
00:01:26.260 --> 00:01:30.350
we specify number of trials and we may also specify,

19
00:01:31.050 --> 00:01:33.690
uh, the optimizer. In this case,

20
00:01:33.970 --> 00:01:38.250
I choose to use the bio and optimization with Hyper Bend,

21
00:01:38.920 --> 00:01:41.380
but tune search supports more of them. For example,

22
00:01:41.700 --> 00:01:45.220
a pure bio and optimizer from a circuit optimized package,

23
00:01:46.370 --> 00:01:48.220
Opta never ground, and others

24
00:01:51.200 --> 00:01:56.140
we call the feed method and we can check what was the best parameters and we can

25
00:01:56.460 --> 00:01:57.820
evaluate the model on this data.

26
00:02:04.710 --> 00:02:08.730
In our next video, we will learn more about the light G B M algorithm.
