Name: 分類生成器テストの試み - 型破りなニューラルネットワーク p.5 (Classification Generator Testing Attempt - Unconventional Neural Networks p.5)
Uploaded: 2021-01-14T12:27:52.000Z
Duration: 12 min 6 s
Description: VoiceTubeの動画で発音を聞きながら英語表現を覚えよう！学べる英語：

命令形

And welcome to part five of our shenanigans with the neural networks tutorial series in this part, what we're gonna talk about is the results from training the general this call it a classifications generator.

So, uh, let's talk about it so you could see what I'm running right here.

I guess you can't see this on the screen.

So, basically, what it's doing is every time it gets the classification, right, Like that was the one predicted it was a one.

And then eventually, it might get a correct one.

Anyways, um, I've been running this for a while.

As you can see, it has leveled out quite a bit.

But if I do smooth this out, you should see that it's still slowly tapering down still, so I'd like to actually let it keep going.

In fact, it's still it's like less than 10% done with at least the 50 e pox that we wanted to go through.

So I kind of wanted to let it continue going.

But I didn't want to do that if it was just a waste of time.

So I actually did not know if this was going to be successful.

My first attempt was unsuccessful completely.

And then I wondered if we gave it a little more of an opportunity to get things right.

If that would help in, sure enough, it has.

Um, at least the code that I used to get this I don't think it's wise.

There's really no point for us to write it.

So I just kind of want to show you the logic that I'm doing to detect whether or not we got something right.

And also we should probably talk about, um what the chances of getting it right are.

So, for example, with a regular classifier, you would say the classifier is better than random.

If it makes it makes the prediction right.

In the case of M nus, there's 10 predictions.

So anything over 10% is you're better than random now with a generative model, though I mean, it could.

It's like an infinite number of things that could generate, So we'd still like to see something over 10% accuracy.

But, um, really any accuracy is is pretty good, because it could have generated anything.

Um, as we can see, it seems to get zeros pretty well, probably because it just over predicts for a zero.

Um, but as you saw there, got the three right and so on in the nine.

It's on a roll that did look like a four.

Anyway, uh, here is the code that I just kind of changed up.

That's what this is was running right at this moment.

Let me just fit it to the screen so I don't talk about something that you can't actually see and, um, not too much here.

As you can see, It's the same kind of logic that we've seen before.

And again, if you want this code, you can go to the text based version of the tutorial.

It's actually not there right now because I made some changes.

And basically, all it's doing is just gonna It's just gonna generate through the exes, which, in this case yet right there, the validation.

So these are numbers it's never seen before.

Um, so it's gonna generate through those, and then it's going to make a classification and output it.

Now, our logic for doing this isn't necessarily the greatest logic.

I'm still very much in the development stage.

But for example, what constitutes our logic is whether whether we got it right or not is, um, wait, I'm sorry.

When he pulled the updated one type, like, commented out this logic, but not shown the other logic One moment.

If arcamax of the label is just in sample dot decode my you know, for the last 90 elements we just say, Yeah, we did a good job.

Um, but obviously, there's a few reasons why, um, Why did it do that?

Anyway, And that was on a model with 18,000 steps.

And now we're already at 21,000 actually, so I'll probably transfer over again in a little bit.

Um, Anyway, what I just wanted to do is, you see, like, is it worth me continuing to train this model for a couple of days or not?

So I'm gonna actually let this one continue.

I think we'll continue with the next topic on the tutorial.

Will probably come back to this one for the results.

But right now, getting about 23% accuracy is pretty good.

But anyways, as I was saying, um, so in the case of like a zero, for example, there's frequently zeros.

Um, so a lot of times to just starts predicting another number like this is a perfect example where I mean, it got the Coghlan's and it started to sort of make that prediction.

But then it just started drawn another zero, basically, or just who knows what it's doing?

But that's not totally right, right, And we're classifying that as being correct.

So take that with a grain of salt, are accuracy.

But just from running it visually, inspecting it as well.

I'll worry about making something a little more official.

We'll probably use something like regular expression or something to find something that fits the correct format and then decide Is this correct?

字幕リスト動画再生

分類生成器テストの試み - 型破りなニューラルネットワーク p.5 (Classification Generator Testing Attempt - Unconventional Neural Networks p.5)

bunch

expression

permanent

basically