Multiple constraints minimization

[dudifrid]

First of all - thank you for the great repo!

Now, I have a question: is there any support in multiple constraints minimization? It is often the case, since I am minimizing a function that gets tensor as input, so I need multiple constraints for the relations between the input tensor's elements.

So far I've tried: list of dictionaries, dictionary of lists, dictionary that returns tensor with > 1 elements, and dictionary that returns list of tensors. They all failed.

Here's my toy snippet of code with one constraint:

from torchmin import minimize_constr
import torch
import torch.nn as nn

eps0 = torch.rand((2,3))

res = minimize_constr(
    lambda x : nn.L1Loss()(x.sum(),  x.sum()),
    eps0, 
    max_iter=100,
    constr=dict(
        fun=lambda x: x.square().sum(), 
        lb=1, ub=1
    ),
    disp=1
)

eps = res.x

Thanks in advance!

[rfeinman]

Alas, minimize_constr() does not currently support multiple constraints. It would be possible to add this functionality, namely by allowing fun to output a vector as opposed to a scalar (in which case lb and ub would need to be vectors of the same size). However, this would require a change to the way that jacobians and hessians are computed behind the scene; jacobian computations would now look like what were previously hessian computations, and I'm not sure what exactly hessian computations would look like (if they are possible at all).

This is a low priority right now, given that the majority of pytorch-minimize is devoted to minimize(). The minimize_constr() utility is merely a wrapper for scipy that was included as a temporary convenience. I have been meaning to write custom implementations of constrained minimization in pure PyTorch.

[dudifrid]

Thanks for the rapid reply! I need only linear constraints (but for very complicated neural network-dependent objective function), and I'd love to here from you what is best for my needs: some kind of workaround (I tried but didn't come up with any, so far) \ different suitable libraries \ is it best to implement this myself (from your answer it sounds difficult, but maybe it is easier for linear constraints?)

[ilan-gold]

@rfeinman I am a bit of a newbie at optimization, but hopefully my math background will help me hit the ground running. I might be interested in taking this on if you're looking for someone to implement this in torch. I have a question though - I have seen people use parameters in a "logX" space or similar so that the parameters that the optimizer sees are unbounded, but they are then transformed by X**parmas to just the positive reals when evaluating the objective. What is the theoretical grounding here, if any, as a workaround? And would something similar work by using a re-scaled sigmoid of some sort (in order to put an upper bound on things)?

If you feel given your expertise that these are both bad ideas/workarounds, I would be interested in implementing constrained optimization for multiple bounds provided it is not a completely insane amount of work and you'd be willing to provide some guidance :) I have a fair bit of open source experience and I just finished up adding some features to a GPU-based numerical integration package, so I feel comfortable offering this. I also knew nothing about numerical integration before I started that and here we are 😄 so I hope I can catch on here too if need be :)

[ilan-gold]

P.S I'm interested only in box contraints, so from I can see L-BFGS-B would be the route to go, I think.

[ilan-gold]

P.P.S I would also be interested in BFGS-B if that exists. No attachment to limited memory particularly

[xavinatalia]

P.S I'm interested only in box contraints, so from I can see L-BFGS-B would be the route to go, I think.

Excuse me, I want to apply L-BFGS-B in pytorch settings and use cuda, are there any implementations of algorithm L-BFGS-B? Specifically, I prefer to the algorithm equivalent to scipy.optimize.minimize(method='L-BFGS-B')

[gnorman7]

I've created a wrapper around this function (scipy.optimize.minimize(method=’trust-constr’)) for my own purposes, supporting multiple constraints. However, I don't compute Hessians, instead using the scipy bfgs approximation. If there's still interest in this, I'll make it open source sooner rather than later.