[DO NOT MERGE] [WIP] multiobjective inspiration #105
Conversation
facebook-github-bot
added
the
CLA Signed
| import os | ||
| import json | ||
| from nevergrad.optimization import optimizerlib | ||
| from nevergrad.functions.multiobjective.core import MultiobjectiveFunction |
There was a problem hiding this comment.
from nevergrad.functions import MultiobjectiveFunction
| if nevergrad not in [None, 'CMA', 'DE', 'PSO', | ||
| 'TwoPointsDE', 'PortfolioDiscreteOnePlusOne', | ||
| 'DiscreteOnePlusOne', 'OnePlusOne']: | ||
| 'DiscreteOnePlusOne', 'OnePlusOne', 'random', 'loss-covering', 'hypervolume', 'domain-covering']: |
There was a problem hiding this comment.
this mixes algorithms and selection mode, I dont get it and the docstring is not up to date.
Also, other modes dont seem to be used anyway since it loops on all selection options
There was a problem hiding this comment.
fixed, now there is an additional "moo", in which case we loop over various Pareto sampling modes.
| gradientDecay = 0.1 | ||
|
|
||
| nImages = input.size(0) | ||
| assert nImages == 1 |
There was a problem hiding this comment.
For not mixing the different multiobjective functions.
Anyway the original code, I believe, is not good for the case nImages > 1..
| if not randomSearch: | ||
| loss.sum(dim=0).backward() | ||
|
|
||
| if nevergrad: |
There was a problem hiding this comment.
not modified, but: to check if it's None, always do if nevergrad is not None
| assert nImages == 1 | ||
| # assert randomSearch | ||
| thelosses = [1., 3., 3.] # These numbers should be discussed... | ||
| target = MultiobjectiveFunction(lambda x: thelosses, tuple(thelosses)) |
There was a problem hiding this comment.
this is constant? maybe use lambda x:x? and then target(losses) instead of target.compute_aggregate_loss?
There was a problem hiding this comment.
When thelosses will be updated, this function will return something different. Tested with:
y=[1]
def biz(x):
print(y)
return y
biz(2)
biz(2)
y=[3]
biz(2)
There was a problem hiding this comment.
I would have though that overriding y would unlink the external y from the one used internally but indeed...
Still, this is an ugly hack you should reaaaaaaally avoid doing for both clarity and avoid extremely weird bugs
There was a problem hiding this comment.
The point is that Nevergrad is not equipped for Ask&Tell in multiobjective mode. I did not find a better solution for now.
|
@Brozi : |
|
|
||
| loss = -lambdaD * model.netD(noiseOut)[:, 0] | ||
| sumLoss += loss | ||
| combinedLoss[2] += loss |
|
|
||
| if nevergrad: | ||
| for i in range(nImages): | ||
| optimizers[i].tell(inps[i], float(sumLoss[i])) |
There was a problem hiding this comment.
you could have an if nevergrad == 'moo' to avoid breaking the code when optimizing hte mono objectif
|
|
||
| else: | ||
| optimalVector = torch.where(sumLoss.view(-1, 1) < optimalLoss.view(-1, 1), | ||
| optimalVector = torch.where(combinedLoss.view(-1, 1) < optimalLoss.view(-1, 1), |
There was a problem hiding this comment.
you probably want combinedLoss.sum(dim=0) and same thing for optimalLoss here.
Otherwise you're taking parts of varNoise and parts of optimalLoss but those parts are not related to the losses that are better.
| varNoise, optimalVector).detach() | ||
| optimalLoss = torch.where(sumLoss < optimalLoss, | ||
| sumLoss, optimalLoss).detach() | ||
| optimalLoss = torch.where(combinedLoss < optimalLoss, |
There was a problem hiding this comment.
same, I think you should also take .sum(dim=0)
|
|
||
| print(str(iter) + " : " + formatCommand.format( | ||
| *["{:10.6f}".format(sumLoss[i].item()) | ||
| *["{:10.6f}".format(combinedLoss[i].item()) |
Update inspirational_generation.py
Update inspirational_generation.py
No description provided.