Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Nested Vectorization #1465

Open
pgkirsch opened this issue Feb 29, 2020 · 9 comments
Open

Nested Vectorization #1465

pgkirsch opened this issue Feb 29, 2020 · 9 comments

Comments

@pgkirsch
Copy link
Contributor

Starting a design discussion thread for nested vectorization.

Generic motivating example

A user wants to optimize the design of an aircraft. This aircraft has several different missions to complete, and the user wants to optimize over all of these missions simultaneously (the "outer vectorization"). For each mission, the aircraft has a max-payload configuration and an average-payload configuration (the "inner vectorization"). Of course the aircraft needs to be able to complete all missions with the max-payload configuration, but performance on each mission should be optimized around the average configuration. To further complicate things slightly, the climb and descent phases are discretized using VectorVariables. This means that some variables have as many as three dimensions.

Design challenges to address

Inputs

In the example given above, the payload is necessarily a 2-element substitution (average payload and maximum payload). But when the outer vectorization is introduced, GPkit currently requires that substitution to be 2xn, where n is the number of missions (the "degree(?)" of outer vectorization). Ideally this wouldn't be the case.

Slicing specific levels of vectorizations

As described in #1464, as_view (which works fine for one level of vectorization) seems to not be designed with nested vectorization in mind (please correct me if this is wrong). Maybe it's a simple fix, but would be good to know if there is perhaps something more scalable than the current as_view approach.

Printing and accessing solutions

The trusty workhorse SolutionArray.table method starts to struggle with communicating information beyond the inner vectorization. It's hard to tell which values refer to which slices of the vectorization. Maybe this is just the natural limit of plaintext solution tables, but I'm interested in discussing ways the table could be made more readable (more explicit), even with three-dimensional variables.
Even the output of sol(somethreedimensionalvariable) is hard to interpret.

Diff-ing

Closely related to the above: how do we adapt sol.diff to effectively capture changes in the solution when some variables have three dimensions.

Discuss!

I suspect one of the possible solutions could be something along the lines of Named Vectorizations, i.e. the idea of a key-oriented vectorization as opposed to an index-oriented vectorization, but I'm sure that would present lots of challenges with which I'm not yet familiar.

I'm sure we'll think of more challenges too, so this is just a start. Interested to hear people's thoughts!

@1ozturkbe
Copy link
Contributor

1ozturkbe commented Mar 1, 2020

@pgkirsch firstly, I have little experience with as_view (pretty much didn't know of its existence until I checked the documentation a few minutes ago).
For a practical implementation using only two levels of nesting, I personally would Vectorize the Mission class over the same aircraft, with twice as many missions due to max and average payload cases. Your objective would be the sum of the average payload fuel weights, plus epsilon times the average max payload cases for solution stability. Your epsilon might be very small or significant depending on how well behaved your problem is.
This is an opinion from experience in ML, but the human brain really doesn't deal well with tensors. Although there may be a practical software way to create and use tensor variables they are really not intuitive to use. You have already pointed out that you cannot print 3d data on a 2d medium such as a screen. As for a practical implementation I will comment below.

@1ozturkbe
Copy link
Contributor

Very simple MWE of working 3d vector subs to base discussion on:

from gpkit import Variable, Model, Vectorize
import numpy as np
with Vectorize(2):
    with Vectorize(3):
        with Vectorize(4):
            v = Variable()
            u = Variable()
m = Model(np.prod(v), [v[:] >= u[:]])
m.substitutions.update({u: np.random.rand(4,3,2)})
sol = m.solve()

@1ozturkbe
Copy link
Contributor

1ozturkbe commented Mar 1, 2020

@pgkirsch What about the substitutions most bothers you? Now that I have tried, they seem to work really well out of the box., and it's relatively clean from a practicality standpoint. As for good printing of solutions, that's a challenge that other models face too, since large vectorizations are flattened beyond 2d. But I'm of the opinion that such solutions should implement problem-specific visualizations instead of terminal outputs. But for terminal, you could try this:

        varkeys = [k for k in sol.program.gps[-1].varlocs if somekey in k.models]

to only print variables of a certain model (indicating level of vectorization).

@bqpd
Copy link
Contributor

bqpd commented Mar 2, 2020

obligatory citation, Tony Tao's thesis has a great exploration of multidimensional vectorization: https://dspace.mit.edu/handle/1721.1/120427?show=full

He had 3d vectorization down pretty well: flight states sharing airplanes sharing tooling

@bqpd
Copy link
Contributor

bqpd commented Mar 2, 2020

motivating example

an aircraft with several different missions to complete

for each mission a max-payload configuration and an average-payload configuration

climb/descent phases are further discretized

Sounds good! I tend to think of it was what variables are shared, i.e. not discretized at each level. So climb phases share a mission, missions share an aircraft. idk but it helps me keep track of it.

Design challenges

Inputs

the payload is necessarily a 2-element substitution
(average payload and maximum payload are both fixed variables).
But when the outer vectorization is introduced GPkit currently requires that substitution to be 2xn, where n is the number of missions. Ideally this wouldn't be the case.

I'm pleased to tell you that it is not supposed to be the case!

import numpy as np
from gpkit import *

with Vectorize(2):  # payloads
    payload = Payload()

desertgoosequail = Airplane(payload)

class Mission(Model):
    def setup(self, airplane):
        # writing constraints in here will be tricky, but
        # I think it can be done without having to manually read
        # input variable size by stretching those input variables
        # across new dimensions as necessary
        mo = MissionObjectives(airplane)
        with Vectorize(10):
            apm = airplane.performance_model(mo)
        return [mo, apm]

with Vectorize(3):
    missions = Mission(desertgoosequail)

m = Model(missions.fuel.sum(), [missions])

Slicing specific levels of vectorizations

As described in #1464, as_view (which works fine for one level of vectorization) seems to not be designed with nested vectorization in mind (please correct me if this is wrong).

Welll it was designed with nested in mind (see my response on that issue) but that doesn't mean it's good at it.

Printing and accessing solutions

The trusty workhorse SolutionArray.table method starts to struggle with communicating information beyond the inner vectorization. It's hard to tell which values refer to which slices of the vectorization. Maybe this is just the natural limit of plaintext solution tables, but I'm interested in discussing ways the table could be made more readable (more explicit), even with three-dimensional variables.

This is really hard! Large design spaces like this are sometimes best understand by focal points, and clustering; we could cluster results together and do a kind of PCA analysis?

Diff-ing

Closely related to the above: how do we adapt sol.diff to effectively capture changes in the solution when some variables have three dimensions.

Been musing about something like this for evaluating "paragons" i.e. making it a first-class thing to solve a model with multiple objective functions, then compare those solutions in a table.

Discuss!

I suspect one of the possible solutions could be something along the lines of Named Vectorizations, i.e. the idea of a key-oriented vectorization as opposed to an index-oriented vectorization, but I'm sure that would present lots of challenges with which I'm not yet familiar.

Unfortunately it doesn't make the tables and display any easier, just the code more readable!

@pgkirsch
Copy link
Contributor Author

pgkirsch commented Mar 31, 2020

Oof sorry for the month-long delay in response.

I'm not sure your example corresponds to what I'm experiencing. It might be easier to discuss with eyes on real code in the other repo, but here is my attempt at generalizing it, this time using aero drag as a perhaps slightly more accurate analog:

...
with Vectorize(numberofmissions):
   ...
    with Vectorize(numberofaircraftconfigurations):
        self.Drag()
...  

where

class Drag(Model):
    ...
   constraints = [
        ...
    ]
    substitutions = {
        'drag coefficient': np.tile(np.array([[1], [2]]), numberofmissions), # previously [1, 2] sufficed
        ...
    }
    return constraints, substitutions

I assume Model.setup can see all levels of vectorization above it, so surely this tiling isn't necessary. What am I doing wrong?

Solution Arrays

I have two proposals for solution arrays, one short term and one maybe longer term:

1. Array Legend

A legend of sorts that can optionally be printed at the top of sol.table(), showing what each element of all multi-dimensional solution values correspond to. More concretely, in my application, I have the following sizes of variables:

aircraftmass  :   x
averagefare :    [x  x]
cruisedrag :    [[x  x]
                  x  x]]
climbvelocity:   [x  x
                  ... 
                  ...
                  x  x] # confusingly this one seems to have shape `numberofmissions x numberofaircraftconfigurations*numberofclimbincrements`

The first two are readable but, without digging into GPkit source code or inferring from the values, it's impossible to tell which element is which (how values are tiled) for the latter two.
It would be helpful if the legend showed, for example:

Variables of dimension `numberofmissions x numberofaircraftconfigurations`
----------------------------------------------------
[[mission1_config1, mission1_config2]
 mission2_config1, mission2_config2]]

Note that this would require the concept of a named vectorization as described above.
Is this even possible to do in a generic way? What kinds of challenges do you foresee?

2. SolutionArray.savehtml()

As you know better than most, nothing replaces interactivity. Hovering over a value and seeing which vectorization it corresponds to would be sweet. A browser-viewed solution array would also open up options like using colour to either differentiate the different missions (for example) or to add a third axis.

And maybe an html solution array would have some nice interactivity to allow filtering so you can only look at one slice at a time if you want.

I could be in a position to help you with this, but I'd want to know your ideas for the best way of doing things. This is at least partly inspired by the wonderful work that the good people of plotly have done to change solution viz.

Accessing Solutions

Closely related to the above, it would be great if a user could access a slice of a solution using either an index (or perhaps the named vectorization key). Off the cuff (and absolutely brushing over the complexity I'm sure this would create), here is an example:

a = AirlineSystem([mission1, mission2, mission3], [config1, config2])
sol = a.solve() # returns a 3x2 SolutionArray
mission1sol = sol.slice(mission1) # returns a 1x2 SolutionArray

Sorry this is a lot. I can spin out anything you think is deserving into its own ticket. Wanted to put all here for now because I think it's all related.

@pgkirsch
Copy link
Contributor Author

Modified my previous comment to remove the "Diff" section. The KeyError I mentioned seems to be a fairly narrow issue with sensitivity diffs only, so doesn't belong in this discussion. It would still be nice to be able to explicitly diff one mission from a multi-mission solution against the single-mission-only solution, but I think that can be discussed as part of the "Accessing Solutions" section. A nice-to-have feature for diffs might be:

a_simple = AirlineSystem([mission1], [config1, config2])
sol_simple = a_simple.solve()
a_complex = AirlineSystem([mission1, mission2, mission3], [config1, config2])
sol_complex = a_complex.solve()
sol_simple.diff(sol_complex['mission1']) # use of `[]` is intended as pseudo-code for discussion purpose

@bqpd
Copy link
Contributor

bqpd commented Jul 10, 2020

@pgkirsch these are great ideas! We definitely have the structure to store index "keys" like this already, between the global VECTORIZATION counter and varkey metadata...the tricks for (1) and (2) above would be more of interface questions, and would be much harder than Accessing Solutions.

It seems to me like a good place to start would be a class (or metaclass) which serves as a "range" and one which serves as an "index". For example:

class Dimension:        
    def __get__(self, idx):
        name = None
        if hasattr(self, "idxnames") and self.idxnames:
            name = self.idxnames[idx]
        return Index(self, idx, name)

class Index:
    def __init__(self, dim, idx, name=None):
        self.dim = dim
        self.idx = idx
        self.name = name

class ConstraintIter(Dimension):
    def __init__(self, constraints):
        self.constraints = constraints  # each mission is a constraintset
        self.shape = (len(constraints),)
        if isinstance(constraints, dict):
            self.idxnames = constraints.keys()

## declaration
missions = {"Short Hop": mission1, "Sprint": mission2, "Long Haul": mission3}
missiondim = ConstraintIter(missions)

## vectorization
with Vectorize(missiondim):  # equivalent to Vectorize(3)
    ...

## variable slicing
x[missiondim[0]]  # x_(:, Short Hop, :)

@pgkirsch
Copy link
Contributor Author

@bqpd thanks for this and sorry for the slow response on my end. I have come back and looked at this probably a dozen times and am still struggling to follow a few things:

  1. Why Dimension and Index are separate classes
  2. How to flesh out the example you provided/use these new building blocks in context
  3. How to tie them into slicing a SolutionArray object.

To explore 2., I have tried the follow:

In [1]: from gpkit import Model, Variable, Vectorize                                                      

In [2]: %paste                                                                                            
class Dimension:        
    def __get__(self, idx):
        name = None
        if hasattr(self, "idxnames") and self.idxnames:
            name = self.idxnames[idx]
        return Index(self, idx, name)

class Index:
    def __init__(self, dim, idx, name=None):
        self.dim = dim
        self.idx = idx
        self.name = name

class ConstraintIter(Dimension):
    def __init__(self, constraints):
        self.constraints = constraints  # each mission is a constraintset
        self.shape = (len(constraints),)
        if isinstance(constraints, dict):
            self.idxnames = constraints.keys()

## -- End pasted text --

In [3]: x = Variable("x")                                                                                 

In [4]: y = Variable("y")                                                                                 

In [5]: from gpkit import ConstraintSet                                                                   

In [6]: mission1 = ConstraintSet([x >= 1])                                                                

In [7]: mission2 = ConstraintSet([x >= 2])                                                                

In [8]: mission3 = ConstraintSet([x >= 3])                                                                

In [9]: missions = {"Short Hop": mission1, "Sprint": mission2, "Long Haul": mission3}                     

In [10]: missiondim = ConstraintIter(missions)                                                            

In [11]: with Vectorize(missiondim): 
    ...:     cs = ConstraintSet([y >= x]) 
    ...:                                                                                                 
In [13]: m = Model(y, cs+mission1+mission2+mission3)                                                      

In [14]: sol = m.solve()                                                                                  
Using solver 'mosek_cli'
 for 2 free variables
  in 5 posynomial inequalities.
Solving took 0.167 seconds.

In [15]: print(sol.table())                                                                               

Optimal Cost
------------
 3

Free Variables
--------------
x : 3
y : 3
...
In [16]: sol(x[missiondim[0]]) 
    ...:                                                                                                  
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-16-fd8a35524b9e> in <module>
----> 1 sol(x[missiondim[0]])

TypeError: 'ConstraintIter' object is not subscriptable

In [17]: x["Short Hop"] # ??? How do mission names get used?                                                                              
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-17-d586d69427aa> in <module>
----> 1 x["Short Hop"] # ??? How do mission names get used? 

TypeError: 'Variable' object is not subscriptable

Apologies if I'm missing the point or making some basic mistake.

What am I not understanding?

As for 3., I think it might be useful to start from the MWE @1ozturkbe provided and show the functionality I'm hoping for and you can tell me how feasible it is.

from gpkit import Variable, Model, Vectorize 
import numpy as np                                                              
with Vectorize(2):                                                              
    with Vectorize(3):                                                          
        v = Variable("v")                                                       
        u = Variable("u")                                                       
m = Model(np.prod(v), [v[:] >= u[:]])                                           
m.substitutions.update({u: np.random.rand(3,2)})
sol = m.solve()
print(sol.table())
# Optimal Cost
# ------------
# 0.02721
#
# Free Variables
# --------------
# v : [ 0.339     0.911     0.378
#       0.866     0.459     0.587     ]
#
# Fixed Variables
# ---------------
# u : [ 0.339     0.911     0.378
#       0.866     0.459     0.587     ]
sol1 = sol.slice([:,1]) # a method that returns a SolutionArray
sol2 = sol.slice([:,2])
print(sol1.table())
# Optimal Cost
# ------------
# (Excluded)
#
# Free Variables
# --------------
# v : [ 0.911
#       0.459 ]
#
# Fixed Variables
# ---------------
# u : [ 0.911
#       0.459 ]
print(sol1.diff(sol2)) # really want to be able to diff slices
sol1.save(...)

Does this seem feasible?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Development

No branches or pull requests

3 participants