GeoSci.xyz SimPEG, 6 Jun 2017

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: SimPEG meeting June 6

Description

Improving SimPEG

A

How do you guys want to do this? Do we want like to make much it really wanted to sort of like go out and Google our first yeah.

B

Let's talk about that, okay, maybe.

A

We'll doc started so we keep.

C

Track of the things.

D

C

Done Tibo good.

E

C

F

Shiny second there's um yeah super CEQA, but maybe oh.

C

D

C

F

Wasn't it done yes, exactly.

D

But there might be.

A

Alright so I miss start a Google Doc and let's do in development proposals.

G

G

A

Still like almost right, 22 you're, just kind of sad about oh yeah, that's.

F

One of the year.

G

G

B

G

B

G

Progress: it's taken over to acknowledge 2010.

G

A

Oh yes, I, remember trying yesterday and the key is mentioned. If you think they're bugging in the e/m code thought we should talk about that, and also just like other things right now that are sort of sticking points or things that we can rally on them like that it improved sort of the day to day stuff, I. Think because I don't like a lot of the issues we have right now, I think are sort of very forward-looking, which is awesome, and that, like helpful in that stuff, is just like annoying right now, I'm, not perfect.

A

Yeah, so sorry do you want to start us off because you cut some stuff yesterday, uh yeah.

B

Once though, I was rewriting being I want reviewing the don't know what like not be writing I. The completely is a separate color from goodness, okay, but mostly a yes for a tutorial for empty one, and that I found I might get quite a few places where you are very inefficient. Giver.

H

B

First, that's like a can. You and I could potentially.

A

Get here, yeah.

B

Can you get well yes or okay? Yeah? Oh yeah, I'll join okay.

B

A

In your general in slap I, didn't ask you by email, nobody.

B

B

B

Can you yes so, for instance,.

B

The first thing, I think, is probably an interesting point that people may not know Oh like it's a it's a specific for e/m code, and so, if you look at the base class and that there is this lines, so you like this is what you store in in the class.

B

Then, whenever you update your model and you want to remove it so I think that you're expecting that's only happening when you're after your first employee for an inversion, but the one iteration includes the the field, computation and J vac and Jacob and multiple of them because, like you're doing see jr., it depends upon how many cgi creation are you doing so there there could be potentially ten five and five five of them each JPEG and trick that we should keep you making Jen for instance, and sometimes like that makes me computing that makes matrix is cheap.

B

But if you think about like how many matrix vector product, we do especially a lot yes. So throughout.

F

The chain yeah.

B

Yeah and then uh yeah when I was rewrite. I was writing the MT code based on mostly the structure of the e/m code, and that I was confused. Why cuz I got it was updating a lot.

H

B

Computed a back JT back, it was updating and the reason. Why is that? Because that that time like it's, so whenever you do, the substa set updates, whatever.

A

We said yeah whenever.

B

We set off now whenever we set the self that model, if it updates so but like at this stuff. That model is always in like in back and JT bag. I, don't think we don't we not sure, but we don't need II. I, don't know. I tie ended up removing self a model in my J back and Jacob Eggman for the empty code yeah, which I think better yeah, because it like only fields need that or so you can.

B

You can only can take it, but you don't you don't need to set the self that model.

H

F

B

What we need to do, we just need to comment this.

F

Line and vests on them yeah, it's.

B

Like I am only required for your fields and then that's only when you actually update your model. Yeah.

A

B

A

Mission, doing just commenting it out is that if you actually do just want to compute the sensitivity still.

B

Fine cuz like it, you need to learn. You need to pass on them to a field right in the field that self that model is reply kits so you're doing twice so they're like there's no point doing that, and then that actually so you do unnecessary, yeah, mace meters, computation, yeah and and the like, probably mass matrix was fine, but I I was taught I was trying to do. Was something along. This line see.

F

Okay, so one second, can you just one question so you're fields here? Are we said that they're fixed right now right as long as as long as you don't do itself that model F is computed only once for Jay, Jay, Beck, NGP van. So what's.

B

Actually happening.

F

In the inversion in CG, princess.

B

Your and cg f, this F.

F

Is constant work? Yes, yes, that it doesn't change exactly, and then you just reset so at the end of the right, stop secular, yeah,.

B

So like it's kind of good to remember it's only when we call the fields, that's actually, when your model is change it and then we need to be complete the results, but what's happening in that inversion code. We we we first compute the fields and pass it through. It is good I, so actually what Thank You Jana, but.

F

B

Sometimes people want to just evaluate the JPAC in the.

H

Other side, without knowing the field saying that, then so, if you don't pass F.

B

That, like a smart.

F

Enough to compete that.

B

In an account, so that's like a big deal. No.

F

Yeah, did you recompute every time.

A

Oh we're not repeating this.

F

Not knowing that all.

A

The operators really.

C

A

F

The only thing.

A

That we might want to do just as a safety. It's basically call so that while I check that it is actually in fact it may pass acid yeah is, if you were to pass right now. If you were to pass F and we get different end. Oh, we wouldn't catch that. So, like you pass death and the end good Cratchit, we.

B

A

Catch that yeah.

B

A

That's sort of the only end upper case that we should watch for yeah.

B

A

If you're passing half then yeah.

B

H

May not even need.

B

To change that, but the like, we, we need to change the base code, yeah, which actually like an ignite that that deletes stuff yeah. So.

D

G

D

B

Checks, whether and it's the same, then you don't need to delete yeah.

B

F

Thing and that so I should speed up those up as I give you. Oh I.

B

Don't know I think I've invested.

F

For actual submit code, yeah.

B

um Okay, so that's one thing: the other thing I have fun.

B

And the reason so I think probably mass matrix might be fine.

B

But like, for instance, if I want to do.

B

If I want to do some time, like this, for instance, say if you.

G

Want to store your factorizations.

B

Say a matrixes like this.

B

D

A potential way that you could.

B

Do so so you set it as a property and then store your a in.

H

B

A transpose iam so.

H

If you have a large.

B

Enough memory that's much better way because the like.

H

Sometimes factorization takes.

B

A while, then well, for instance, like you, just need the one factorization for each iteration yeah.

E

So that's because because.

B

Small doesn't change so you can actually store it. Otherwise,.

F

You losing using the gain of doing it directly, so yeah yeah.

B

F

You re computing in factors every time, yeah.

B

F

Like a yes or no yeah, with a bat, isn't like good reason, why are we doing into X up? Is the back in the back? Step is much faster than doing all the iterations of.

A

Permit absolutely is Han yeah accuracy. Oh it's.

B

Like we're still getting that because we have multiple like Assad, but we're not recomputing for multiple right inside that the for, like I, say, feel when they're can being the field in JPEG and JT bag, or they could be that's their same.

D

B

You could reuse we're kind of recomputing. So if you consider that like say, if you have a ten of those up like a computation, the ten times of every factorization which could slow down your speed a lot of goods, so.

A

Basically, what we said the problem: we just had a keyword that is store factors, and then we do this wrong. Yeah.

B

Something like that, that's a I think we probably considered as an option with both the frequency and time domain code, yeah and.

E

That'll speed up quite a bit.

B

I guess in the case, when you about the enough memory, they'll definitely speed. A blog yeah did.

F

You solve your products when you're doing backs all right. You were getting something like creates slowdown because you were in forward for something is then you saw saw that one you remember. You were I.

B

Don't like it that was I think that problem was kind of mingled but I. Think it wasn't it wasn't. It wasn't just that matrix vector product I think it was something inside that the maze matrices because, like whenever we are looping over it, wasn't actually stored full in snippets, but, like part.

D

Of it and then.

B

Like it contact me, then then complete I, guess that we were doing in that for loop. Oh.

A

We shouldn't fish now the incarnation yeah.

B

Yeah yeah yeah so that that has been fixed. It.

E

Because it's defined another way to that, you, like your program, was reflexive Mac. The metrics metrics product was faster than the matrix vector.

H

E

One was implemented, like is a very efficient library, why's everyone yeah.

B

Well, I wasn't 100% sure cuz whoops re are post, it was he's taking led Finland me yeah Canon's yeah. There was two guys like the one and Jake Jake yeah.

E

His G actually.

B

Replied there could be a fundamental issues inside PI, but the our re showed that the it's just that it was a problem of like a. How are you slicing yeah? So if you kind of do a trick where you copy a man like, do it better that actually helped.

A

Decide yeah he has around.

A

B

A

So what the matrix make sure this products that actually accesses the Blas library, especially like your very base level Chico, whereas.

H

A

Major spectra, product who's accessing so I think I've, we're digging. My snow machines probably would make a huge difference, but others like it absolutely nice occasions, especially, is we're calling makes.

C

Expect some yeah I.

B

Haven't yeah heaven like actually fix it so that I get it I think we may not want to change it, but we may want to change actual side pie code yeah that.

F

Then we don't need to change that in.

B

Our code so without big of a fix with the unite and, although I think cuz what the why I I figured I think that's actually make a big slowdown, the code, because it's a picture, that's a big difference, and then what we misunderstood was our major expected. Pollock is cheap, but there wasn't cheap. There was my feeling and for especially for the DC code or IP and DC code, because what I was doing, cuz I was exactly doing the same DC code, but I have a multiple time channel.

B

Then your matrix vector product is ten times more or twenty times more and yes, I did the simple calculation. How long does it take? This inversions was like it.

H

Was not a big size.

B

But the twenty days or 10 days to actually compute like duty full inversion well so I what I ended up doing I store the sensitivity matrix rather than like doing this major component, so I just computed this pulse and steering it once then.

F

It's a real lecture to.

B

It yeah, then, then I think that's like that's actually like. So then I could do really fast, like for us, relatively small size problem so but I get that actually happening in every all the code, all the e/m code, so DC code, although.

H

Nonlinear version.

B

Code, which is yeah yeah I, think that we need to fix that I haven't, checked, I haven't visited you.

A

On issues I cannot.

F

Yeah yeah I can kind of remember that's when you get the new your sources you're right at size, room I, just.

E

Put this back over floating.

F

I, remember seeing it us again, it was pretty even compelling share my.

A

They assume, oh, is it in fact.

F

A

F

Now it's actually.

A

On people's radar, so this guy, if you have a look basically and compared so this guy, sees hardly matrix-vector like it's looping and opulent blast. Whereas if we look at these.

F

Yeah there is a certain function: CSR Macbeth men.

F

Well, that's but that's the C language. Now, let's see there.

A

F

Junk I'll be seeing the rest is by them. Oh.

G

A

Are handbags right? Oh.

C

F

C

B

F

B

F

That's the violin for I.

B

Think that's a glass that other parts are mostly see it's when they're looking over. There was our C code and then you know in the loop they use that blas package I think that's potentially major effective product.

A

Yes, late weekend and.

F

That was a slow part of it. This.

A

Is this one's fast.

F

This one's birth, yeah.

A

G

C

B

Actually, a former per simple.

C

B

A

So this isn't it's like what, if all we get reject your case actually build is intensely songs that were full-on to initiate inside line and then like I. We can try it like this. It shouldn't be too hard at Lee I'm, not familiar velocity. I imagine probably figures uh or or something I.

B

Think, probably not I should have done that yeah making any cuz I was a little bit.

H

B

Because the is feelin guy he he actually should he or she. He showed a certain way that, like he could do it yeah, so I wasn't sure. Actually, what's what's the real issue, yeah so I'm still like not exactly sure whether, like that that's a problem or not so.

H

That was a reason why I kind.

B

Of stopped, but the it's like a heat. The only problem he showed like like it just is slicing sliced or a considered a certain thing, and he just need to copy that I've been making your a and then do a loop that actually helped so I wasn't I, wasn't sure and then like he actually found that concatenation in the mass matrices yeah so yeah, something I. Wasn't we.

A

B

Maybe could be oh hello. Oh oh.

G

A

Yeah, so, if trying to test again, because it should now should be preparing with a shocking reaction on occasion, yeah.

B

I'll I'll test it again.

A

B

That's another their thing: a fun drink so.

B

Sam I'm not really happy with the like. All of this set basic a survey class how it works like you're, the source takes their sieve, your class or I get a benefit. What what takes? What it's fine, but the like yeah, for instance, say here. If you look at the he bell and what's actually happening, it's kind of hard to see, but I'll show you what I've done, for instance,.

B

So here's what I was playing so say in in the e/m code, there's a real and imaginary part and like it. So this is the projection. So f is my field I evaluate to a location where you I got this and like you, don't need to do it.

B

Do this matrix vector product twice, but how the current code works, because you have a receiver for real part and then another is referred imaginary part, and by doing that, we need to do that plus it's twice and also, let's not make that fields, but when you're doing that you've got like a JPEG and g2 that also unique. For instance, you need that to the two: that's all and I'm.

F

All bunch of major tech, your father was in theory both imaginary all,.

B

Right so, but like, if you actually put the both, then that kind of solves an issue, and the other thing is because, um like we don't need that much evaluation, because each well how it works. It receiver has all the location and then battled.

H

Evaluate every.

B

Time like, but um so for instance, how UBC could works, for instance, say they you, they find all.

H

The unification.

B

One evaluation matrix then a valley once and whatever in time figures. So that's kind of I can make it more in a way. It's efficient, I, don't think that's very user-friendly and finding the data. It's not really like easy compared to how sick backwards, but ah yeah in a certain level that what you really slow down. Your speak for.

H

Instance, like good new scope,.

B

Say you have a whole.

H

B

Evaluation matrix because you have a field H field and you have a to polarization and if you do that, every time that actually.

H

Like that, let's go.

C

Yeah, that's a lot of the matrix.

B

A

I've been thinking about this as well today, I, don't I also have not necessarily like how the source case may be reasons for water service, not all of them. Let me see what I'm team, like here's, your receivers, would statement all frequencies so that why is diversity? A science? That's like over cereal, ah yes in something, what I would perhaps see is basically I'm Hank. We basically like this receiver is less acute all of these cruises and this source, like should yet always your fingers and then.

A

It's also a question for, like probably, must set up a survey there because like to be measuring both a ex and ey right now, liturgy separate procedures which I think it makes sense in the investing point of use. You want to Gartner next one, um but there's this name location, exactly saying burglars, matrix right now,.

C

A

F

C

F

But that we're having expensive efficiency, yeah.

A

So into you like, would you want to then here with Jerry, certainly by saying like.

A

Stever, like almost hop locations then here to consumers, is that I have all of these.

H

A

And now I'm just can say that a yes on their UI, three or imaginary exp. One said you know. How would you want to air right? How do you want to set up your way yeah.

B

Yeah, it's uh it's kind of hard a depends on.

F

B

Survey right so, for instance,.

F

Like say, isn't, location always the core debates, this food- that is true so you're.

B

A stake that the lines see definitely.

A

B

With the unique locations I think and but there's a subtlety, for instance, it's exe why? It's not oh right.

H

Because they're.

B

Living in a different place, so we need a different like a relation matrix yeah to do that so yeah there's a bit of salty on like it's, not just the location, yeah ah yeah.

B

D

B

Yeah yeah I'm not sure how I was like uh sort of using it. For instance, DC code, sometimes I generate.

H

To one receiver cloth one receiver.

B

Class has a unique alternate location and where I can change I just computed but like when I'm.

F

Putting a wd matrix.

B

Like but they like, this data.

C

Weighting matrix, I.

F

B

Put the impact like infinity for the location where I don't want to see so I think that's potentially, that could be in behind the scene like so that's a potential way that we could. We can probably use it exactly same structure that we have now, but under the scene we we actually complete unique locations, yeah and, and then.

F

That's a good point, though, like you sang precip emptied your e^x is with two poles or it H is a sicko point, but in theory all like the poles of the eaves and location of the H could all be unique, are all unique locations and you're just doing different operations on those we're going to asking the fields on all of those points? Are you going to keep them or not? That's that depends on the area on your circuit type of it.

B

Some something.

F

B

F

A single projection: are you purchasing all your fields on all the calls, and you should get it taking out for your data transformation. Vl do.

A

We want to also have like supposed to be able to make anything, some, perhaps being able to say the line number and yes.

B

A

That we want I thought.

B

A

B

Think I would I would do in a like in a different place. Okay, so I would, like a word say, kind of a like a CSV type by format where.

H

You have a line like a connects.

B

What's that type of format like never line line, ID and all those stuff, but yeah I kind of want to separate in with this, if Aksarben class, maybe there could be a easy I/o that can transfer this to a here or this to bear, but when we're taking the lines and data I think I kind of want to do it in a different place, yeah yeah. Otherwise it's very convoluted and yeah.

B

A

Attitude a Greek city just get confusing them and what we actually require cuz. We don't need to necessarily file woman, just ones factor numbers that make sense. Yeah.

D

B

Yet so that it goes.

F

It would be a big major overall, the right to change this then and in TEM, and how that surveys creative well.

G

B

I, don't I, don't think it'll be a what once you've got the right structure. I think it's mostly just parsing things and moving whateleys.

A

If you think we can always the wheels.

C

A

Gate at a big step, we're not instead right take a big step forward. That sort of pervert passive should be too.

A

Yeah you were this weekend with the objective function: stuff, there's some great things with how always property and.

F

But then how they understand I kind of like the way, that's because eventually and each of those guys could live on different measures right, yeah,.

H

F

If we try to change all of them from the from the top, it might not, it might not work it's just because it was document. Doesn't it yeah.

A

So it's what if you guys, are aware that dollar in two years he was wanting to throw lobbies something on the regulation, actually change out looking degree I'm great.

F

A

That's it um right now, if you try and access it from like a couple gently function like once, you sit the singing, isn't quite right to.

G

The commo objective.

A

Function has a regular station mesh. Then each of you little pieces that make it up. Also memory.

G

A

It should be the same modest, they really are only like doing it and then they're not actually tied together. So if you go to change something on, one of them haven't propagate that ah should.

F

They be the same I guess because they are I, think that on it saying that model parameters, room yeah.

A

F

A

What I'm inclined to think is that, if you want to create regularization from different meshes that you probably should be finding like.

C

A regularization and.

F

Yes, yes, yes,.

A

B

A

Salute in a lot of ways, I think that this example that you sent she can operate it angers us back that you should just the right match on your collar regularization, simple and how hot you want. Yeah.

F

Yeah yeah, because of them all primary ler are giving the finishes that they're not linked to two different objective function.

A

Yes, okay, I think.

G

A

Would be calm over the organization, and this is told everything the way I sort first saw it. It's like we should for me how to be able to treat it. The same has like a single regularization, but it should have services.

A

So if you call team al4 smallness, they ask the same, even though smallness.

C

Is only one.

A

Wild or university off from happy.

F

Is there any other pieces to their share? I said the right. Actually right mesh is the only thing that they're all.

A

Mappings Empress skills: well, the algorithm accelerates, so it's the opposite. Different I can pull it up. Oh.

G

F

Then she just dropped out there.

C

This is not being.

A

Alright, so the way it is currently.

A

H

A

In that, if there's a lot of things that they share, so if we look at Li combo regularization.

A

So we set the Outlawz.

A

Then they share an M breath if you're in my thing soon they shared reactive and so as well as needed, right mesh. But the weight of these were the way that they share. These things is, like I, haven't, actually properly cleanse them, and this is something that with ranking for episode. One is that like when we omit this I, make sure that all the things are the same, but they won't necessarily like their.

F

Model pointing.

A

To the same options, everything.

F

If you change your drink, of course, the guys still going to keep its original, then is that so that's.

A

Or lose integrate yeah.

F

Because now we need to loop, everything I want to change something we need to route through all the I'll, be the children yeah.

C

A

um That's something that we can run with right on and Sikhs.

H

I think he's actually recently implemented.

A

That properties, but I just don't know how. So if you might be able to give us sort of a one-line at this platen, it's not how to probably look for destitute what we should be.

B

F

So when the MTA yeah, but in what way, exactly so like what we have right now in this immigrant work, how is that? How is the code you're working on different Oh, so I good Nia's, published an example there? What DMT, when oh.

B

I haven't I, haven't like I haven't, played with exact same problem, I could yeah I could.

D

B

Yeah just work: that's not.

F

B

F

To have a full understanding, it was yeah.

B

Well, more, like a good, we want to make a tutorial, so how good nice doing it? It's more like that I kind of tighten along with his 3d coat so I think it's I, don't know, I, don't exactly know how it works, but he complete the primary field using the analytic code and do a secondary formulation to compute. If you got some area.

H

B

Something that's.

D

H

D

I haven't really go.

B

Through so yeah, it doesn't like a resonate with the tutorial.

B

H

B

And go and the point uh I was thinking configure just sure.

H

From A to Z, like.

B

Say how you scanner the script of simulation code and sensitivity code instead now.

H

B

Into a simplex structure, since the problem class received class early class, how you parse it.

H

B

Then generate the problem, class and actually monthly version today,.

F

B

Was what I was that kind of my scope? What's that, haha so we ended up writing it, but I yeah we're not sure what we may not fit that into animal tutorial, but we could. We could put those as a people with want someone want to see and someone's actually from a girl code to us in my problem.

H

B

Yet that might be yes, odorless.

F

B

Well, works yeah, it can invar for applicant face a little bit. Difference proved a little bit.

F

No I still have to finish it. Look at the key are working on rotation. So more thing. Focusing research purse.

H

Is to notice with the grounded source, the Ian problem: that's Augie a permanent month or so ago. I think I've been I've prepared. Then, comparing that code with smoothly UBC pillows and then also comparing it with the DC code and I think that as we're doing, the integration of the source between the Phalanx I think the current is actually going in the opposite direction. I think it's in because, if I set, my like a elector didn't want to be electro and I.

H

Look at it for the DC case, where I look at it in case for low frequency, like the polarization sometimes, which is a sine clip, I'm happy between those two guys ever so I. Think if I switch, which is my a which might be in one of the two, then they match. Okay, but I can't use the same family for the DC code and after the air source you. So if I need to flip it in one village, so.

A

Which one is currently in sensibly, where who has a being depositing.

H

Lantern think of decency, okay, because the DC matches very well with that you see DC I'm gonna do.

H

All right so I think there's just something small yeah in a way that we're setting up the source and the ground it's routine, one where we're getting grating forward between the two old rules. So.

B

Dc matches with the analytics right like because it's just a matter of like what is false,.

H

And what is actually minus.

B

And DC Malchus web.

H

Analytics DC matches with UBC and analytic and I mean I'm able to get when I was looking at the en stuff. I was able to get it to match to, but it wasn't quite a straightforward right because we have the different coordinate systems between the UBC code and the simple code, and then we also have like they're using the different convention for the via the minus I Omega T versus the positive I Omega T. What are you transforming?

H

So there is like a negative, complex conjugate that you need to be taken to play certain things, but.

A

It over you can see if we look real hard. It looks like yeah.

B

I think that's something worth archived yeah like it, probably just at least one datum and whatever I says signs and stuff that you need to be supposed together. What are the conventions that we're using yeah because, like what, if you see the problems in UBC code like we.

H

Don't know actually.

B

And they have different conventions for different codes, yeah, so he's truthfully. That has changes and all that comes and change it. So Roman.

E

B

Working on air yeah was working on stairs that oh yeah.

C

B

Anyway, you need to rediscover all the signs it. It's actually not a trivial yeah issue say it really needs serious like a fast right.

F

Now and actually.

B

He'd run the code yeah, it takes a while, if you don't have a ground truth like it's very hard to.

A

Yeah, well, is it because a couple things we could do is in an augmentation.

D

And change we definitely.

A

Helpful these things there may not be a big weekly imminent. ah The other thing that would be really helpful is just like showing comparisons to analytics, basically saying because like when we compute dipole response, we should be matching work on reaching all the exact interventions is going on, and so it should be r XO.

B

Exact there's a thumb.

A

In some examples it is what I was thinking of, just like the type of emphasis yeah.

B

A

Is a right handed coordinate system.

B

Yes, so we should like.

C

A key that's trip ancillary head: according.

A

Fits your silly basically showing that we should give a benchmarking and something else to literature.

A

Yeah I know that you've heard much of those examples. So if you have small on that, we could use you to have the gosset basically say: here's some expectations.

F

All right, a big row, you're good. You know why why XYZ other things she done yeah, so, let's flip it up, do.

A

You want hate, unison, aghast, it's just like a finger things. Yeah yeah.

F

A

C

Because I've been as I question.

F

If you've done the rhythm yeah, what's the order, we're counting this with this or this way and yeah yeah UTM.

B

D

B

Like III consistency, all the Kosh insistence, yeah.

A

B

The edge of this here.

C

G

C

To do us- and it's true and.

B

People mentioned the like a mapping yeah for.

H

B

Mapping derivative.

D

B

It actually it provides a matrix, so it doesn't think actually a vector right. So that's a potentially kind of when you're mapping is a functional. There's.

A

Something here it should know this, just something I implemented a while ago, yeah.

B

A

I would be surprised because.

B

The how you could you could go ahead: people a good, quick attention: cuz yeah, like a the mace matrix. How, if you look at the mace matrix derivatives, that's where it.

E

Fails so I think the constructor.

B

Actually there but the actual code assume it provides a matrix yeah. So we cannot quite do a certain functional into your mapping. So yeah, that's potentially we need to change it and that's.

A

G

Simple change, I think.

A

We just need to take.

B

A vector and then I think it.

A

Is actually no grace clause that says.

C

B

G

Yeah Acacio.

E

When I was doing, this wicked project can apply, I have to just use. I define my kind of my own mapping at some Congress to be able to use like full year where a transform, which is you, can define it as a as a graduate students. That is the section on one yeah.

D

E

A

Cool the antes in oh, that's! So before you wrap up. How do we want to use like happiness and get actual.

B

So that's in Google Doc now or we need to put it in Google them so.

A

Thing zeros God's here and say that.

G

So we have actually less year, oh and that's suitable well.

B

I think I can just transfer into issues and.

A

So what will be helpful? There is, if it maybes for the name on image person brought up. If we can't, if you don't mind, transitive next to issue and then also, if you know wearing the code, things need to change like if you can articulate that as well. Oh yeah, they mess that helps point these out, because there are people who are interested in jumping on board area. He was this student in 350, you great, including County, awarded and popping out, and so in these types of things help articulate where it is issue time.

A

This is a good it's! No! The code base I also make a valuable contribution to all of us. Really.

C

A

B

Right on yeah, and potentially, we probably need some actual, like a kind of efficiency test for the cousin I've done a couple times. Yeah there's a profiler yeah that one vote before it's all like really handy like in terms of visualizing, but like it's actually pretty nice, just rather than doing typing things where, like it, actually profile things and check how what how much time that you're spending for certain yeah, yeah I think especially M codes I think we really need to do it for.

H

A relatively big size problem and also DC.

B

Code that we're like it is normally.

H

B

I think there could be some part where you got a pretty big memory, which you're not supposed to or like a certain part, is not very inefficient yet because I am I feeling like we should supposed to get similar efficiency with the UPC code for three equal x at least a tensor code, if you're using very similar solver, yeah, yeah, so I think we can do it, but like that there are some parts: we're not just going yeah.

A

Well, but um there's a color, you can add if you want to suspect from the docs, which is good an example that you want to see. Think that the other thing is is if we actually have like folder full of sort of efficiency tests where.

G

We want to have stuff.

A

That is a little bit bigger. We can change Chavez to basically like only test on all requests or things like that to make sure that before we virginity thing, we haven't at least like destroyed specifications, Oh efficiency, oh yeah,.

B

Yeah yeah I think that's probably invited now, probably all right time to do it, we're pretty evolved and yet I think it's pretty stable. Now, yeah.

A

Yeah, so in this change we make we should be worse. Yeah.

B

Exactly exactly is that yeah.

C

B

We've done like a to go yeah.

H

I thought I was.

B

Happy yeah with that yeah, but I change it, but let you expect all it should be pretty simple art yeah, but no it's different. Surprise.

D

Surprise, but if it is it projects.

A

Oh, oh, okay, yeah, so.

G

A

Type up some issues and start hitting on some of these things.

F

Sound anything.

A

That you personally want to tackle this type assault recipe champion, and then you know it's movie star on your radar. Otherwise you float around.

C