Fixes for CHOP #14
Fixes for CHOP #14
Conversation
solvers/chop.py
Outdated
| self.X = torch.tensor(X).to(device) | ||
| self.y = torch.tensor(y > 0, dtype=torch.float64).to(device) | ||
| scaler = StandardScaler() | ||
| X = scaler.fit_transform(X) |
There was a problem hiding this comment.
this should not be done in the solver but in the dataset as it changes the problem
There was a problem hiding this comment.
Can it be an option in the dataset?
There was a problem hiding this comment.
the libsvmdata package does some preprocessing like this. I would just offer the good dataset (normalized or not).
There was a problem hiding this comment.
Right, it seems this was the cause of the infinity values... (Since the scaler wasn't applied at eval time) Modifying.
There was a problem hiding this comment.
@GeoffNN dont change the solver but the dataset. Otherwise solvers will not be comparable
There was a problem hiding this comment.
Done; I added standardization as a parameter in Covtype, so it runs on both versions of the dataset. We can also just keep the standardized version if you prefer.
| requirements = ['pip:https://github.com/openopt/chop/archive/master.zip', | ||
| 'pip:scikit-learn'] |
There was a problem hiding this comment.
| requirements = ['pip:https://github.com/openopt/chop/archive/master.zip', | |
| 'pip:scikit-learn'] | |
| requirements = ['pip:https://github.com/openopt/chop/archive/master.zip'] |
|
|
||
| if self.standardized: | ||
| scaler = StandardScaler() | ||
| X = scaler.fit_transform(X) |
There was a problem hiding this comment.
I would simplify and always standardize.
There was a problem hiding this comment.
Ok -- only for Covtype? It would make sense to do it for all non-sparse datasets, right?
|
I'm still working on the stochastic variant -- I'll share the figures for both ASAP |
|
I don’t know about Madelon but the others should be already scaled |
objective.pyThis doesn't fix the issue mentioned here.
This still needs work for the stochastic case.