GeoSci.xyz SimPEG, 11 Apr 2017

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: SimPEG meeting April 11

Description

Combo objective functions and joint inversions

A

No Kate me was fun to drive you're, not I, didn't.

B

Hear from in them.

A

A

um I, don't know if you guys might start with things on your list or the thing I want to talk about it. Drank versions and I have a combo.

C

A

Functions: yeah.

C

A

A

Lot my screen so Franklin you can see the key see my screen.

D

A

Ok, so what I wanted to go over is the objective function, stuff in particular, with three versions: I was looking at your directives and also your script just to get some ideas on like we're constraints are and what kind of things we want to be able to do so. The first thing I well there's two things that came up is both which we're talking about the validation of directives and then like working with properties of the objective function. I'll start.

B

With the validation, because.

A

I think that's more straight forward, um so in this case there are things like the ordering matters as well as potentially like the type of objective function. You feat something to like the irls. You should only give it the sparse, if you give it anything else like it's, not actually doing a veteran primary. um So in this case, what I've done is added a.

A

Directive, I've added a validate to the base directory, so all it does so validating like by default. It will just return true, so the default directive is balanced and then, if you want to go and do something like this was on the irls.

A

So if all I did was check, if there's a linear preconditioner and then if there is it needs to be before so, this will just fail. Yeah.

E

A

And then print out, like please put your directive before that's good. Otherwise, if there isn't a linear preconditioner, this is a question for you. All. I did is basically say that convergence might be slow, so I throw a warning but I'll still let it run. Okay,.

C

That's fine yeah every directives gonna be different, but you pretty much just put a switch there for us yeah. That's good yeah,.

A

So like this, this validate is specifically on the irls.

C

Directive yeah.

A

C

Art and put it right after your when you create your about your your directive. Oh it.

A

Doesn't actually go, excuse me, um so it doesn't go like in the initialize or anything like that. It's just a separate method. So here there is just a new method called death validate okay.

C

That takes itself.

A

And it takes a directive list I.

C

See and so we can have any directives any check on your directives inside that function. Yeah exactly so. It.

A

Can be anything about the creation of the directive and then also why I gave it. The list is because it does like it cares about the other things in there. So then, all that happens that when this is actually called, is within the so this is now when you create a directive list. All it does is now the directive list has a method called validate on itself, and it's just going to go through all of its directives and say: are you valid yes or no? Oh man.

B

They're all valid, then.

A

We're getting out right right right on.

C

That already existed right. There activists the skull by optimization or something yeah.

A

Yes, a directive list already existed, but now its lowest.

C

A

A validating my sweet sweet and that um Franklin down the road can play with properties nicely. Potentially um we don't have properties injected into the directives yet, but that's on the list. So where is this yeah.

B

Where it, where is it, where are your medications, see an indirect ofstuff? No, but your branch? What branch.

A

Oh I, I'm doing a pull request on to your mag amplitude, real, quick, just bringing them yeah.

B

C

Thank you. That's great yeah.

A

So that's like Chris, you jerk yeah, it's nice to have because, especially when you start playing with multiple directives and things like that, don't like yeah, so dude cool okay. So that's that one. This is one that I would appreciate. Thoughts on I was chatting with Rowan a bit about it last night, um mostly because I just started playing with some like base functionality and Python and as soon as you start messing with like set attribute and get attribute. That scares me, but you can do some really cool things. I.

E

A

So because what I noticed um John and your example is um so do this, so you had done something like this and then defined you sparks, but that's fine still break. Is there one times break 1, plus 2 times rate you so we're not going to do a joint inversion for two different parameters, um and then you want to set like the MRF, for example, um and so in this case, if you just do, raiga and r f equals.

A

At ed, doesn't actually in its current state that doesn't talk to write 1, I'm right to all that that has done is basically just added a property to reg called em breath. Ok, so.

B

Does actually do anything.

A

But I totally agree that it should so like we should be able to if you've created something that is composed of two regularization, both of which have an mrf. You should be able to just set them and.

C

Then it would just trickles down yeah.

A

So this is what we were getting at when I sooner talk about exposed, should try and expose properties, um and so I think, when I was talking to Roland I think link is actually better and we can get into the details of why um so I'll show you an example of how, like I, think this should work. So in this case we have our big one and write to you and right now, if I set two different em breves, oh I, do you break one dot com, ref plunge or two da ra? It's used.

A

Ok! So now I wanted to find my regularization, which is now composed of those two things and now I want to be able to access em wrath on both. So what the idea here is to use something like a link to basically then expose em ref from the children objective functions to the top level. So now, if I do something like this, if I said reg got em breath on my combo and I want to then see what reg one got em ref is.

A

It is now that taint propagated. Ok. What.

C

Is going to happen if, if rag1 and rag2 are not and steel on, the broad sense wrecked, you could be a three NC right, not every resident is always on is always on yeah.

A

Regularization should always beyond the model, so the sizing, the sizing issue.

C

Ok, sorry of wire, secure enough. That's the thing: yeah I CSU.

C

Well, okay, because in my case my that's, my reg too yeah is is a combo okay! Oh.

A

You made to break.

C

C

So will that propagate all the way through the combos as well, if.

A

If we've set look, if we've chosen to expose it, yes are chosen to expose that specific property. Yes,.

C

Okay, well, that's good! Oh.

B

Hi, yes, those it's.

C

A deeper evil: what's link, never seen link before what is it you that.

A

Is right now that is what is exposing the MRF property? Okay,.

C

A

The lower level guys to the top level I.

C

Guess you've created that you've created that thing. It's.

A

See I created a sketchy version of it, okay um and so right now, oh, it doesn't like what we pull up this comment, so I was trying to throw in last night and there's some weird things potentially in terms of behavior that we want to just clear up so right now we exposed em breath or the link well. This is where I've switched to lately, but if we started with thinking about exposing amurath and all that that was going to do is basically give you a downward propagating yes same. Yes, the.

C

Upward that the exactly yeah.

A

That's a good, and so um if we set em ref is a and we print that that would have propagated all the way down. That's fine yeah, um but then in the in its current status you actually could go in and change right, one and right to dependent.

C

A

And that's confusing so.

C

You mean you changing from a top level and that's that re changing it individually. So your know.

A

Is if, if you decided to do that, so if I set reg 1m ref, it's a yeah.

E

A

I set right to dem records, be, and now I'm looking at the top level function. What should that return, yeah and so the way I think we work around, that is we actually, instead of exposing and just giving you a downward propagating thing, we tied them all together and basically say now. If I change em rack on the top level that changes in the plot level, I change in the bottom level, it's also going to change it in the top level.

A

So we tie these things together through a link and then give you then an option to unlink.

C

Okay, so we're giving me an example: what's going to happen, ok.

A

So, in this case, if we set red got em ref is a and print that you just get back eight and then you get back pay for all of the children objective functions.

A

Now, if we sit reg 1m ref is B. That's going to propagate it to all of the things that it's tied to.

C

What do you mean whether so.

A

If I tied together right one and raped you oh great.

C

A

Gonna come p. Thank you. Can it become be as well.

C

But that's not good, so that's already helping else. Does it.

A

Well, I think I guess the like my comment here is like I would be inclined to throw a warning at you. If you do this, if you're, if you to write one goddamn ref like throw a warning that says this is linked to these other objective functions, we are changing those as well, um because I like you, should be working at it from the top level. If you explicitly hide these things together.

A

But what this achieves is, then, that there's no ambiguity of what right got, mrs, should be returning like the combo objective function. It is well defined which return.

C

Yeah I thought the our goal was to be able to change things from the top of the combo objective function. Right you just without creating lists and moving through lists. Yeah.

A

C

That is that helping us yeah.

A

This helps us because in this case, reg dot M ref actually exists and is then tied. We title these things together,.

C

Mm SS, okay, so if you want to operate on the top level, I su s. If you still want to trigger each one, if you want to change individual and individual components, you still need will sting to loop through it right, yeah, yeah now.

A

That's something that we can think about is because maybe this is too complicated like maybe it would just be better to at this well. The other thing that we could do is provide if you call right up amref. It then just gives you back that list of all of the objective functions and rest on the subtle thing. There is that if you have added objective functions together that don't have an mrf or just returned aunty list, the sizes should be the same, so it could return num.

C

No I know yeah.

E

C

Exactly like, you would be able to answer it in that list, those that have a maraca ganado back there and those that don't have a member of our gonna happen. Mt mt entry or not. Okay, that's.

A

Fair, it's.

C

Just be again same things happening right now in the joint inversion right. So it's a you call it what's deep brown I would want to have it lists. Always you know, because each each misfit functions should return something. When you asked what my deep red is right, can we always return the lists? Why are we not returning less, even if it's not if it's not a combo or obese, always your.

A

C

Because it makes things simpler then it's always then.

A

We're only thing is always going to view through.

C

A list and then extract things ever yeah.

A

C

This, the ones with sub 1, aroma.

C

A

C

We can you that's right, you don't do.

A

That is this, like this.

C

This is still good right, because this still will still be operating on the letter yeah on the comp board. I love it absolutely. Okay,.

A

So we want to be returning lists and you do want to. Then we still want to be able to know.

C

We go down top level, but then, when you call them from up y and then you always see everything right, I think so I would vote for that. I was kinda stuff. You guys you're, no.

A

I mean it's like up for. Discussion is yes.

E

A

Is we're defining how we want to interact with this?

A

A

A

Franklin, have you played around with like tying and linking good yeah yeah.

D

So they actually a branch- oh right, buddy, Framm or version. That's that's linking ok. So I guess I should man.

D

Yeah I mean so it's like if property is not linked and then you give it a instance and a property name, and you can just give it like as many as many instances in property names as you want and it'll link them all together. So when you change one, it changes them all and it can be directional or it can be like just changing one changes the other, but changing the other doesn't change. One cool.

D

See ya, it's it's on a on a pull request. If you want to see it, but um yes, I mean it seems like that would be helpful. It looks kind of like that is what you're doing with link I haven't. Looked at all if thats MPEG codes, yeah.

A

Well, it's sort of like it's very much a first pass implementation at this point.

A

Yeah because there's there's link and then there's also like the ability to expose it on the parent objective function and the way I'm doing that now um it is through setattribute and get attribute, which is your. What I was her to asking you vote.

D

So many words, oh so.

E

D

When you add rag1 and rag2, the thing that it creates doesn't have an M ref by default, correct, okay, cool yeah.

F

D

The exposed thing kind of turns it into something that you can see, and then you have to link whatever is exposed that make sense. Yeah.

A

Exactly yeah, so it's like, which I think is the functionality we want so.

A

And if there's a question to of like, if we actually want to expose.

A

All of the properties like by default, instead of just instead of having just eat it, yeah.

D

E

I guess I don't know.

D

What um I don't know enough of what is going into adding these regularization together so like? If you add, can you add to regularization that look totally different into a third one yeah, you can't just be like. Take the overlapping stuff or something or well good,.

A

Yeah- and you don't necessarily want to is like those in some ways, the point of creating two different ones so that you can give them different attributes and different properties. But then, basically, a top-level just say like evaluate this thing that I've created. Oh.

A

But what we could do is basically like exposed like hesitant and I. Don't have handle my um we could. We could expose everything at top level um and then basically just give you lists back always.

B

Yeah, we need an examples: yeah I'm, not sure right now, man, yeah.

C

Okay layers: that's.

F

A

Yeah Franklin, if you could even just ping me with which pull request you're looking at for linking that's something I, would appreciate playing around with yeah sure.

D

I'll, just put it in slack.

E

A

Let me just share my screen here.

A

Okay, cool I will take a look through this.

A

D

I think that's cool. Sorry go ahead, go.

A

D

Just gonna say: yeah, it's basically just directional link and I'm link. That's.

B

Awesome: okay,.

A

And so then, this is a.

A

Prop this is like a property or an attribute of alba property. Is that we were going to say that.

D

So the link the link is just sort of like a function in properties. Little X connects two different properties.

F

A

Cool none like perfect.

A

Did you want to go through the joint inversion example? You said you got something up and running well,.

C

Yeah everything is, you still need to fix a scaling directives? Okay, but it rocks. Your version runs. That's great. Okay,.

B

C

Look at the the way I, that's that's where I'm stuck right now is the way the predicted our handle rest. You have to miss your function. Yes,.

E

C

E

They're added their.

C

Added together, how does that I was that where, where are they stored? Basically, why are you predicted latest or are they sent back to their to their individual misspelled functions or no.

A

There's some weird and I think you're right and that this needs to be probably completely rethought. So I believe that deep red should.

C

Be in your inversion problem and problem.

A

So right now, I believe if you call the pred on thee.

A

Yeah I- really don't like this actually so get deep red right now will return a list if it is a combo, I spit or.

B

A

Just returned ah effective mm-hmm, so that's on the info. Prob I think the datum misfit. So this is again like on the top, the top level, if you were to call like what is deep red of spectrum and miss petunia of yeah you're like comp, isn't that wouldn't give anything as of right now? Can.

B

You repeat that if you're calling be proud of, oh, you can see, must get 1, plus D misfit you of the coma yeah.

C

A

Won't do anything I think.

C

No return something, but it's on your vector. You know one of the two, so if I would be like, if you, if you ask for a demon, spell deep red like your misfit deep breath, it returns something. But it's not twice the data length if you want like it, doesn't have both in them.

C

Okay, um but I, realized that if I call get deep red, then it will return you this for both, but that's in and.

A

You're calling that from where are you calling are from.

C

So you are, you join together to mess with function vm as a component spit right.

A

C

You create the misfit. One is a problem, one ministry to it's a problem too yeah.

F

C

Add them together, it's it's a complementary function: yeah, oh yeah! You should be able to call that thing. So if they call, if I do this mr. function not be proud, because, usually it's a misfit, the deep breath returns. You predict the beta yeah yeah. If you do this on the combo, it returns me a list with only one element in it, which is one of the two but yeah I don't get both basically but now, I realize if I call inverse problem instead getting prepped. You issue return it to me. Okay,.

A

You should shoot it shouldn't hurt, that's really scary, because it shouldn't in the in the current status. So let's try this.

C

Maybe maybe I'm wrong.

A

Hey so if we create and then let's import no problem.

B

So you'll need to survey thing uh yeah. You probably need to so yeah, let's create a potential field. So how do I do that? A.

C

Probably close yep the magnetics just.

A

C

And then the happy girl, that's true. That's.

A

It and then where's that guy j.

C

I've heard about that I'm gonna do.

A

No she's, my mish.

A

C

That they'll need a survey to rain right.

A

So I'm do survey.

E

C

Right I, don't remember how to you.

C

Always Louis study from can you look in the in the driver, yeah.

A

C

Cab driver and then go down all the way down to love the date. Then how does it the.

B

B

Face mag darling personally, okay, should we think up? I will copy this guy early.

C

And just put like, like XYZ candy yeah,.

E

C

And being Newell, the divine me it's also a.

A

Three-Digit sort of need to be that's.

C

Like amplitude than that azimuth angle, just three numbers and any three numbers scheduler, oh.

A

It needs to be an array.

C

What's the difference be a server in every and that's why it.

A

Oh I guess a rape probably doesn't as a ray no.

C

I, like em p and p like yes, oh so.

A

Empty dot, our is going to just make a vector yeah.

C

A

In this case, we're expecting for for the receiver locations we're expecting a 2d array, okay, and so, if something comes in, that is not TV, it's unsure, which is something we should think about now.

C

What I was asking, what's the difference between this fiscal and peed array and versus NP as the ring? Oh.

A

C

I'm always using as the rating with them I.

A

Think, actually, azeri is probably what you want in this case is.

B

Use your alluring back there two day there ya, go.

A

Array, I guess, is a little more flexible if I were to give it a list of two things um or like if you, if you give it that.

C

Beam beam idea, if you.

A

Give it this sorry, forgive NP daughter ray this. Oh.

C

I think so, it is then transfer into an hour yeah.

A

I believe so that's soon, I.

B

Don't know if that actually works, let's try. I have be my last via list BS to be a list are eventually let's get one.

A

Oh Artie, its me I was wrong.

A

Ok, so we've heard that guy so could you do miss.

A

Kate, a misfit of spray, how that works? Oh damn into it. We want L to.

C

Where is there.

E

B

No, you need augs in there. What oh.

A

C

Economic function, that's cool.

A

Aight we need a mapping where's the problem, so pretty much.

C

I'm on sky map here and I, just whipped like yeah.

A

See and sure we probably want.

B

A

Fine, ok, so then we've got one data misfit now.

C

If you had them, yeah create the master one yeah exactly.

A

Chris um probably should fail in the future. We shouldn't necessarily pie the same survey to to gate and misfits like that's something we should think about when we're combining objective functions. It should they'll, probably work right now and.

C

You repeat that right now, two problems at the same survey: yeah.

A

Sure, probably well, no we've only created one problem, but we've given it to two data. Misfits and.

E

A

We're gonna add those together. That seems sketchy if.

C

There would be things you have a thing, the same, the same as for function: yeah, I,.

A

Like potentially in an inversion, if you could be directive at two different data, misfit that have the same problem like that, could cause some weird things to happen. So that's something we should think about. Okay, so two.

F

Guys, okay, yeah, okay,.

A

So now we had a few minutes um so.

B

Everything you asked for deep brother.

B

Who SD brother only the improv should have okay? Well, can you can you do the we have a new program, hood the.

A

Survey has deep red.

A

So the survey has deep red and the in Prague has deep red.

C

Okay, I'll double check, though of them not acting. What I was doing. Okay, sorry about that yeah.

A

No worries on what that is.

C

A

Again, hopefully, something we can achieve with the the linking is be able to expose deep red ah and give you back a list of things. Okay,.

C

I'll double check some reason: I was only able to access one of one of the two but I'll find the other problem, and I will help ya.

A

If you did surveyed now.

C

You might have been named problem that the president returned only 11 to 20 cannibal cherries, I'll double-check, though I'm, to spend too much on okay.

A

Yeah, let me know.

B

A

I mean survey makes sense to have deep red because there is there's one problem. One model like it's: it's a single forward simulation at that point, so it should be able to it has to be paired. So if you haven't paired it, it would fail to get d Pryde and the pairing is like a. What are your thoughts on that? Is that a weird, because in some ways, like I, think we should probably just an out talking to Rowan about the church bay? It's just you basically like my simulation is, are now something like.

C

The problem in sort of a sort of area and.

A

Then you would ask like simple fields: hmmm same guy.

C

And then no more problem with physics, how.

B

Would you need a name, a problem I.

A

Would said, though, I don't know free we're settled we'll deal.

B

A

In a different yeah, I think the objective.

B

Function, stuff.

A

Is changing enough things because.

C

That would be dope pretty readable right. We do a simulation of yet.

A

C

This survey, yeah Henry 30 fields, I, actually really.

A

C

That that's really readable.

A

Yeah, that's fair, okay! Well when words, when we get to that went into this, that's that's what that should look like.

B

Where is that problem with physics and I drama from the problem is problem.

A

Cool okay, um if you have time later today, I would be game to just like sit down yeah.

C

I'll take an hour to work work that example, we're really close, really really close. Yeah.

A

C

It's inverting, but we just need all the knobs to be to be working yeah, that's very fair! It's a lot of us man we're going to yeah, that's getting pretty complicated bad! It's that it's I.

E

Mean the code is.

C

This surface is nice and slick with the starting to touch a lot of different components. There yeah.

A

And like trying to make sure that we can do that in a reasonably same way,.

A

C

A

C

Safety checks that we can possibly do yeah.

A

C

Will will face problems as the as the surface, it's hard to foresee everything right, yeah, yeah, we're pretty close I would.

A

Say so? Yes, so the validation, the validation floor, request I think can go in the other side.

E

A

Kind of I wouldn't float in, yet it's liquid, the exposing stuff. We need to think through some things before we actually started hitting on that yeah.

B

A

They can that can win sure cool um anything else. You guys want to go over or a seat. Give me also well give me popped in briefly, is gone and he's gone.

C

I'm sorry working at the shaggy on the dairy waiting for the DC problem.

A

C

Actually, we benchmarked the same peg against DC 82d. Oh.

F

C

I mean it's slower, but the results are very similar. So that's that's.

B

Encouraging that's good.

C

Yeah, so his 22 and add Diaz is doing a pretty good job. That's.

B

Good, oh, we need to look how much is lower and.

C

I don't know I can I can. I can time it, but this program factor like three right now: okay, yeah yeah.

A

Fair enough, there's some things things to do there, cool.

F

A

It's going to work thanks guys thanks for joining Franklin.

D

Sure no problem I have a better idea now to to give some more thought to that issue too. Okay,.

A

Yeah I would appreciate hearing your thoughts on that because they're I feel like there are there's some interesting things happening, there's also some stuff, that's potentially quite dangerous in terms of exposing stuff, so yeah, possibly.

C

Don't expose yourself friendly thanks.

F

D

Cool see you later thanks.