Numenta Live Streams, 27 Apr 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Freeman42's anomaly model params question part 2

Description

https://discourse.numenta.org/t/issue-with-getting-anomaly-parameters-instead-of-swarming/5754/30 -- Watch live at https://www.twitch.tv/rhyolight_

A

B

A

And return my desktop audio back on it is on I'm gonna. Take a break. I'm gonna play some nighttime low. Five, that's laughs.

A

And we're gonna get back to what I was doing before the research meeting you log off of the double its charging it up. Okay, thanks for joining I will be back and we're going to jump back into trying to build hot Jim and Python. Three I was.

B

Actually, in the middle of a.

A

Python debugging session, dealing with a a question on the forum from Freeman about this. These anomaly params so I'm sort of partially through trying to figure out what's up with this and but I haven't figured it out yet so I might just kind of put this on pause, I hate, putting things on pause, or maybe I'll just finish it out, and try and figure out how to how to do this. The I'm, hoping within one minute there, the where it says annum.

A

It's a research meeting up here on the top of the screen, will change I, hope and because it should change to something different.

B

A

Should say, let's build something with HTM any moment now, I think it just.

A

Let me make sure, because I've been using this tool to switch the topic of the twitch feed, and sometimes it doesn't do the right thing Matt from dementia. It is live streaming. This work also sleeve. It's simple. This is fine, but don't be behind the scenes. That's fine! Okay, software, robotics design, it's good enough! It's good enough!

A

Okay, so I was trying to figure out why the model parameter. Sup I feel like we're close to figuring out. What's wrong with this. Why are the model parameters.

B

A

That we're getting right here this is where I was at and I was about to I was I, was jumping into the let's go to call stack the OPF and I'm and I think that I'm getting too far into it, I, don't think I want to jump into the oh yeah, that's how I was getting into model run and that's not really what I want to do that.

A

The problem is, has got to be some type of bad model configuration for new pick for the OPF model that we're using to create the model, because when we run it and we get a result back we're not getting predictions back, let's, let's run several times over and over and let's see what happens if we just run without a break point: oh oh yeah we're still trying to pull them all this stuff, hello, ducky, Daniel, thanks for the follow I appreciate that doing some I'm actually stuck in Python 2 right now, and a Python 3, but I'm eventually gonna be Python 3 I'm, trying to debug an issue in Python 2.

A

Let's get out of clothes all the stuff that I'm not using and right here. This isn't working so I'm, just gonna copy copy it out and we're just gonna watch the counter accumulate, hopefully we'll at least get that we're not going to get predictions out, but we're gonna say at least it's processing. So it's running, but it's not getting predictions out, I wonder if I can add a debug statement it does it stops. Ok, so so I can look into the result.

A

Now, after it's had a chance and just make sure it's still got no anomaly score.

A

Okay, it's even it's got a prediction number, but it's not returning anything. It's not returning any inferences. Ok, so there's gonna be something wrong with the model. Parameters are reason to create it. So let's look into new pic again, because I know that there are example, model parameters.

A

In a Python file find bottle programs, hi I, know a ton of these yeah anomaly, so I want a prediction model. So this is the one that's returned by swarm. Let's do this one. Let's look at this.

B

A

Params, okay, get out of this.

B

A

Here's the example that we're using the coders consumption- this doesn't really matter. It's I, don't think the problem is with the encoders I Jose I am so I'm a community manager for an open-source project that deals with machine intelligence and one of our forum. Members is having a problem.

A

Getting anomaly parameters or creating a model that predicts temporal anomalies makes predictions of temporal well, it gives you anomaly indications of how anomalous the temporal stream of data is and he's having a problem with it and I'm trying to debug the problem so that I can answer him on the forum, and so it boils down to model parameters which is like the configuration that you give to that you use to create a model and the model being a collection, essentially of like simulations of neurons, like populations of neurons and we've got a population of neurons that runs a process called spatial pooling and another called temporal memory sort of on top of it- and this is the like the basic core of what HTM is HTM- is hierarchical, temporal memory, so I'm looking specifically into these parameters, so.

B

A

Input with the 0, so that's what we what I said last time. We should expect that and fit with. So let's go back to the right before we create the model and looking at the the parameter. Where do we create the model? Creat? The model here, discrete model, okay, right here, I'm gonna pause. This is where we want to look and see what the model crams look like.

A

So here we have the prams model. Config close this and model parameters.

A

Jose thanks, thanks for following out I'll do JavaScript and Python on this channel, but it's all going to be somewhat related to hierarchical some portal memory for more about that here is a link and also I've got a lot of good videos on our YouTube channel HTM school you want to actually. Maybe this is a better, a better link if you want the whole series of videos on HTM and how it works.

A

Okay. So let's compare now the model params that I found darn it? Okay, that's not what I want where I have too many things open. Here's the model params that I found.

A

Alright I was going through all these before anomaly params. That's not what I want I want to look at I've been through several of these inference types, a pro multi step.

A

Let's make sure that we have the right inference type here.

A

Hey Falco thanks for joining I've been going a long time already today kind of that's too Pro multi stuff. They make sure I, don't have a typo in here and prints type temporal multi-step model parameter configuration furan, s-- type, inference type to promo, stuff, okay, just making sure sensor, params I, think are not the problem. Something else is the problem.

A

A

Do not think sensor parameters are the problem. Tm enabled true SP forensic, whose strength the only thing I was worried about was input, strength or if it with and that's not, it should be a problem number back to columns. 40. That's right, that's boost strength.

A

This is three. This shouldn't affected, it I mean predictions. I should still be getting predictions out of it. So why am I not getting predictions? Okay? Let's look at the TM parameters. Activation threshold cells per column. Thirty-Two input width is the number of columns mini columns over many columns. Newton EPS account twenty usin em to count 20 the none of this stuff train s peanut. Only if requested false, that's probably an artifact from solder day, Pam.

B

Lee, we don't even use anymore.

A

Classifier parameters may be okay, if it could be, it could be in the classifier parameters that we're having a problem, see how cramps. Let's look at that? Okay.

B

We've got a region.

A

Named SDR classifier region steps one five, but we're only doing one. We didn't specify five, so that should be okay.

A

The Alpha is different, but it's still controls how fast the classifier that should be problem: training, SP net only for question.

A

I'm still having an issue of figuring out, why we're having a problem here? It's he's absolutely right.

B

A

There's no there's no predictions coming out. Okay, here's! What we're gonna do we're gonna go into new pic we're.

B

To go into examples.

A

Opf, climate, hot Jim, simple, okay, so there should be something here: I can just run Python hot Jim, and it's just going to run. It I want to make sure this.

B

A

Okay, so it's working! So let's look at this.

B

A

Our hot gym- let's, let's inspect the this- is like the simplest HTM you can make right now. Is this little the simplest.

B

A

This should be doing predictions I'm pretty sure. If our model Krabs is right, it's a prediction model tempura multi-step. So this this is where you should start.

A

B

Let's inspect the result right here, because we should be able to get into this result where.

A

Do I put the breakpoint there is. We should be able to go into this result. Let's run run the file and oh, no module name new pic: oh oh yeah, yeah I'm in Python, 3, I'm, gonna, switch to Python 2 and then we're going to rerun so I like the interpreter.

B

A

I told you to please: um okay, we're still is still something wrong. I think I can't rerun. It I think I have to hurry over with the right interpreter here.

B

A

Let's look at our result now and see if we've got in some predictions because we should, and if we do inferences, we do okay, so I'm gonna Freeman. If you're watching I'm gonna point you right here, this script works I, don't know what's wrong with the other script, but this script works the wood and it's there's something and there's some difference between the model. Is it them? It must be the model parameters.

A

So I, don't know, I, don't know what the issue is, but I'm I'm totally gonna. Point you to a different example, like the simplest hot gem prediction example is in like I said examples: I, don't remember where it was now.

A

Sorry too many windows I had one I'll find it it's under opf clients, hot gym, simple right here. It's hot gym! Try this script! Okay,.

A

Okay, Freeman sorry about this Ross Rose, Vaughn I just spent some time on this and I cannot.

A

Figure out what is I, took your code and ended up having the same problems: hey there, so maybe I don't have to well I'll point you to this I'm going to point you to that's like dictating this to you, while I'm typing to this really simple, simple hot.

B

A

That does seem to work form for me. Okay, so here's a deal since you're. Just since you just got here I'm trying to figure out what was wrong. I took your code, I I I took your whole script, your whole run script and I and I simplified it down and I even having the same issues at your head. When I get here and I try to pull a prediction out of the multi-step best predictions, there's nothing there. So I had the same issue I've compared the model parameters right.

B

Here that we get, because this is what I supposed.

A

Once you get these params you've got to change the model. Configs model params in French type-2 temporal multi-step, to make it a little two-step model, but that doesn't seem to help that doesn't seem to help and and so the data also there is a I'm using the rec center data and I just put C 0 C 1 up here, just so it'll, because that's what the get parameter is saying does that's what the canned columns that have names that it creates?

A

Hey, you're, welcome, mark I thought you had already been gifted a subscription before so sorry it took so long.

A

So that's where I'm at and but but so here's what I did I was like having insanity problems and so I went and found like the canonical simplest, hot gym example, which is just this really small script. You can make it smaller by taking all the metrics out and it works. So, but it's it's loading the parameters from a canned file. It's not using that function, but so there's some difference.

B

A

This and and like you should be able we should be able to.

A

Let's see if we can do this, while we're here while we're here, let's go into your run and and let's grab this this function and in the other one, let's try and use it instead of.

A

Using these right- let's this may be like maybe I can maybe I can replicate this. So so here's.

B

A

Yeah, so in create model where we're getting these model. Trams, basically, do you know of any usage of that function anywhere, maybe yeah there's. Definitely we use it in the mint to apps.

A

Minta yeah and I mean we should be able to put that function in here. I know, I've got it in my in this repository there. It is so here's here's where it's we're use it, and this is in, like.

A

Don't ask me what these are. I would mess with this stuff in a very long time in HTM engine, though we're using it, which, which runs.

B

A

For stocks and H tips, studio also uses. This is that a unicorn is an issue studio, so there's another place and the pram kinder this would be. This is the delay the place. I would probably look come on.

A

Do this again, but just told me, it was in here and then nevermind this. This is where I would look anyway, but.

B

A

If, let's see if, when we create this model,.

A

Right so I'm gonna do this and I'm not gonna, send it the data I'm just gonna. Send it a minute, a max I'm, not gonna. Do the data thing I should be able to just send it a minute uh max I hope so here in this other create model, I'll say: params equals get whatever. And what does this want? It'll take min Val, Max Val, so I'll just send it a minute. A max I.

B

Think this should work.

A

So then, zero Max, Val I, don't know 100 I'm, just gonna put something in there and that's I think that's all I need to give it and then.

B

A

To give it a resolution, so that will give me back some params and then I can say grams like we did in the other one grams model, config model, friends, I, think here, let's see what happens with this now. So let's run this hot gym.

A

And so we get type error. It takes at least one argument. Oh you have to give it something so I'm just gonna give it, but it ignores it.

A

It'll ignore it if it hasn't been valued, Beck's value as far as I can tell okay what happens at work.

A

Where am I simple, yeah.

A

Wait a minute something's weird: this is my hot job: I want to run. It run the Python file where's! Oh, it went out here, I missed it. There we go okay model model, Factory Kiara model, so that config I'm giving it as a right what am I giving it. Then, let's stop right here all right again: yes, that the okay I have to create a new configuration right, add a configuration.

A

A

Model config model params, so.

B

Damn I didn't have hi I'm Sam.

A

Auto config model params and that's not what it once. What is this terminal area? Yeah CP yeah with the zero thing?

A

Oh this. This is something that.

B

A

Something silly.

B

A

So I've got my prams and why doesn't this work again? Let's go that worked now. It's still so here, hot gym, 64.

A

Here so it's looking for something called model in here, so I'm, obviously I'm not sending it the right thing. So, whatever is sending into this, it should have a model on it. So, let's see what is the thing with the model? The model config yeah, so I think I'm just sending it one mint deep or something I've got to do it again.

A

Okay, hey so we got somewhere. Let's see what the result looks like. Oh there's a so we got to change the C right, a C in the. So this is going to be C. C, 0 I, don't have a C.

B

A

This is gonna, be C 1. This is going to be C, 0, C, 0.

A

Okay, go another key air, so another C problem. Let's see model input.

A

A

Oh right, the zip headers the headers are coming from the file, so I'm gonna actually make header is equal, C, 1 to C or season seats, I'm, just gonna fake. This cuz I just want it to work. I'm just I just want to prove that this one, okay there's another consumption. This is going to be c1 go again.

A

There is no encoder for this set of predicted field, consumptions, there's more consumption, so you'll some show the field consumption, that's C,.

A

Let's try again Romo Sarah I hope, there's no encoder for the field. Consumption.

B

A

Consumption, consumption of them everywhere ever cut, let's just place it with C one five, one, two, three, four: five, yes.

A

One more time with feeling it's working: okay, so let's inspect the result, object and make sure that it is that it has what we want run again. Let's look at the result. Please have inferences. Please have multi-step that doesn't it doesn't have.

A

Okay, so there's so so.

B

A

Here's what I can tell you what I thought that you could easily take the model parameters returned to you by that function and make a modification to them so that it would add the predictions to it, but it obviously doesn't it doesn't have the predictions to it. I, don't know why not what I would suggest you do is.

A

Start with some model parameters that are not coming from that function, I guess I I, don't know what the answer is yet I'm still very confused, but there's definitely something wrong with that function, and it's not returning it's not returning you something that you can get predictions out of now. That being said, I wouldn't call it a bug because that function was not was not designed to give you a model that returns predictions. It was designed to give you a model that returns anomaly. Indications.

A

So my suggestion would be to follow this example. This simple hot gem example and use model parameters starting from there and then see how that works and then tweak them because it's honest I think it's better to have a set of model parameters that you set to the side, and you know what the values are and then you change them over time. So I'm gonna leave it at that I'm gonna mark.