bmcfee

bmcfee

bmcfee

# SKLearn's one-vs-rest class requires that the (binary) estimator object implements decision_function() or predict_proba().
# If you want the internal estimator to contain a parameter sweeping layer (so that each ovr classifier gets optimized separately), 
# this fails due to the following chain of events:
#
#  1. OVR checks `hasattr(estimator, 'predict_proba')` at construction time
#  2. `hasattr()` tries to call `getattr(estimator, 'predict_proba')` and fails if an exception is thrown
#  3. Because the estimator has not yet been fit, it has no `best_estimator_` property, so it throws an exception and fails
#  4. `hasattr()` misinterprets this exception, and returns false.
#
# We can circumvent this probelm by putting a wrapper on the predict_proba and decision_function methods, but this is a dirty, dirty hack.

bmcfee

bmcfee

bmcfee

bmcfee

bmcfee

bmcfee

bmcfee