MAINT refactor imputation of inactive hyperparameters into function #176
Conversation
smac/configspace/__init__.py
Outdated
| from ConfigSpace.util import get_random_neighbor, get_one_exchange_neighbourhood | ||
| from smac.configspace.util import impute_inactive_hyperparameters |
There was a problem hiding this comment.
I haven't found such a file. maybe still missing?
|
btw: have you checked that the random forest can handle -1 for categorical parameters? |
|
@sfalkner could you please comment on the question from Marius? It boils down to: does the random forest code from commit 21157707735bc8a11ebf3790a8f5d8de2fcab860 (basically version 0.2 from last April) support -1 as a value for categorical and numerical values? |
|
No it doesn't. What would be the purpose of that value anyway? |
|
An outlier value representing a missing value. I'll update this to impute the default, but still be fast. |
|
You could use 0 for undefined and use 1...num_categories for the actual values...Don't know if that helps, though. But I guess you wanted the same value for numerical and categorical variables. |
|
Now it's the default again (as it was before). When can decide when the deadlines are over. |
mfeurer
force-pushed
the
maint/speedupincativehpimputation
branch
from
c635b6f to
2a336af
Compare
1 similar comment
smac/runhistory/runhistory2epm.py
Outdated
| @@ -296,12 +296,12 @@ def _build_matrix(self, run_dict, runhistory, instances=None, par_factor=1): | |||
| for row, (key, run) in enumerate(run_dict.items()): | |||
| # Scaling is automatically done in configSpace | |||
| conf = runhistory.ids_config[key.config_id] | |||
| conf = impute_inactive_values(conf) | |||
| conf_array = impute_inactive_hyperparameters([conf])[0] | |||
There was a problem hiding this comment.
this is an vector and not an array, right?
There was a problem hiding this comment.
Which is the same class in numpy, but I get your point. Will change this.
smac/configspace/util.py
Outdated
| from smac.configspace import Configuration | ||
|
|
||
|
|
||
| def impute_inactive_hyperparameters(configs: List[Configuration]) -> np.ndarray: |
There was a problem hiding this comment.
the name of the method is a bit misleading, since it does more than imputation; it also converts it into our internal numpy representation.
There was a problem hiding this comment.
Docstring is there, how about convert_configurations_to_array?
1 similar comment
|
Thanks! |
Refactor imputation into it's own function @mlindauer