def write_binaries(rows):
# row is dict from split name (train/test) to path data can be found
for pair in rows:
i, row = pair
with as_local_paths(row.values(), with_query=True) as local_inputs, \
as_output_files([path + '.xgb' for path in row.values()]) as local_outputs:
for local_input, local_output in zip(local_inputs, local_outputs):
write_xgb(local_input, local_output.name)
def trainWithFiles(fold: Mapping[str, str], train_matrix: str,
params: Mapping[str, Any], **kwargs) -> 'XGBoostModel':
"""Wrapper around xgb.train
This intentionally forwards to trainWithRDD, rather than
trainWithDataFrame, as the underlying method currently prevents using
rank:pairwise and metrics with @, such as ndcg@5.
Parameters
----------
fold :
Map from split name to data path. All provided splits will be
evaluated on each boosting iteration.
train_matrix: str
name of split in fold to train against
params : dict
XGBoost training parameters
Returns
-------
mjolnir.training.xgboost.XGBoostModel
trained xgboost ranking model
"""
with as_local_paths(fold.values()) as local_paths:
matrices = {
name: xgb.DMatrix(path)
for name, path in zip(fold.keys(), local_paths)
}
dtrain = matrices[train_matrix]
evallist = [(dmat, name) for name, dmat in matrices.items()]
metrics = cast(Mapping, {})
booster = xgb.train(params,
dtrain,
evals=evallist,
evals_result=metrics,
**kwargs)
return XGBoostModel(booster, metrics, params)