def _create_device_quantile_dmatrix(feature_names, feature_types, meta_names,
missing, worker_map, max_bin):
worker = distributed_get_worker()
if worker.address not in set(worker_map.keys()):
msg = 'worker {address} has an empty DMatrix. ' \
'All workers associated with this DMatrix: {workers}'.format(
address=worker.address,
workers=set(worker_map.keys()))
LOGGER.warning(msg)
import cupy # pylint: disable=import-error
d = DeviceQuantileDMatrix(cupy.zeros((0, 0)),
feature_names=feature_names,
feature_types=feature_types,
max_bin=max_bin)
return d
(data, labels, weights, base_margin, label_lower_bound,
label_upper_bound) = _get_worker_parts(worker_map, meta_names, worker)
it = DaskPartitionIter(data=data,
label=labels,
weight=weights,
base_margin=base_margin,
label_lower_bound=label_lower_bound,
label_upper_bound=label_upper_bound)
dmatrix = DeviceQuantileDMatrix(it,
missing=missing,
feature_names=feature_names,
feature_types=feature_types,
nthread=worker.nthreads,
max_bin=max_bin)
return dmatrix
def dispatched_predict(worker_id):
'''Perform prediction on each worker.'''
LOGGER.info('Predicting on %d', worker_id)
worker = distributed_get_worker()
list_of_parts = _get_worker_parts_ordered(
has_margin, worker_map, partition_order, worker)
predictions = []
booster.set_param({'nthread': worker.nthreads})
for data, base_margin, order in list_of_parts:
local_part = DMatrix(
data,
base_margin=base_margin,
feature_names=feature_names,
feature_types=feature_types,
missing=missing,
nthread=worker.nthreads
)
predt = booster.predict(
data=local_part,
validate_features=local_part.num_row() != 0,
**kwargs)
columns = 1 if len(predt.shape) == 1 else predt.shape[1]
ret = ((delayed(predt), columns), order)
predictions.append(ret)
return predictions
def dispatched_get_shape(worker_id):
'''Get shape of data in each worker.'''
LOGGER.info('Get shape on %d', worker_id)
worker = distributed_get_worker()
list_of_parts = _get_worker_parts_ordered(False, worker_map,
partition_order, worker)
shapes = [(part.shape, order) for part, _, order in list_of_parts]
return shapes
def mapped_predict(partition, is_df):
worker = distributed_get_worker()
booster.set_param({'nthread': worker.nthreads})
m = DMatrix(partition, missing=missing, nthread=worker.nthreads)
predt = booster.predict(m, validate_features=False, **kwargs)
if is_df:
if lazy_isinstance(partition, 'cudf', 'core.dataframe.DataFrame'):
import cudf # pylint: disable=import-error
predt = cudf.DataFrame(predt, columns=['prediction'])
else:
predt = DataFrame(predt, columns=['prediction'])
return predt
def _create_dmatrix(feature_names, feature_types, meta_names, missing,
worker_map):
'''Get data that local to worker from DaskDMatrix.
Returns
-------
A DMatrix object.
'''
worker = distributed_get_worker()
if worker.address not in set(worker_map.keys()):
msg = 'worker {address} has an empty DMatrix. ' \
'All workers associated with this DMatrix: {workers}'.format(
address=worker.address,
workers=set(worker_map.keys()))
LOGGER.warning(msg)
d = DMatrix(numpy.empty((0, 0)),
feature_names=feature_names,
feature_types=feature_types)
return d
def concat_or_none(data):
if data is not None:
return concat(data)
return data
(data, labels, weights, base_margin, label_lower_bound,
label_upper_bound) = _get_worker_parts(worker_map, meta_names, worker)
labels = concat_or_none(labels)
weights = concat_or_none(weights)
base_margin = concat_or_none(base_margin)
label_lower_bound = concat_or_none(label_lower_bound)
label_upper_bound = concat_or_none(label_upper_bound)
data = concat(data)
dmatrix = DMatrix(data,
labels,
missing=missing,
feature_names=feature_names,
feature_types=feature_types,
nthread=worker.nthreads)
dmatrix.set_info(base_margin=base_margin,
weight=weights,
label_lower_bound=label_lower_bound,
label_upper_bound=label_upper_bound)
return dmatrix
def mapped_predict(data, is_df):
worker = distributed_get_worker()
booster.set_param({'nthread': worker.nthreads})
prediction = booster.inplace_predict(data,
iteration_range=iteration_range,
predict_type=predict_type,
missing=missing)
if is_df:
if lazy_isinstance(data, 'cudf.core.dataframe', 'DataFrame'):
import cudf # pylint: disable=import-error
prediction = cudf.DataFrame({'prediction': prediction},
dtype=numpy.float32)
else:
# If it's from pandas, the partition is a numpy array
prediction = DataFrame(prediction,
columns=['prediction'],
dtype=numpy.float32)
return prediction
def dispatched_train(worker_addr, rabit_args, dtrain_ref, evals_ref):
'''Perform training on a single worker. A local function prevents pickling.
'''
LOGGER.info('Training on %s', str(worker_addr))
worker = distributed_get_worker()
with RabitContext(rabit_args):
local_dtrain = _dmatrix_from_worker_map(**dtrain_ref)
local_evals = []
if evals_ref:
for ref, name in evals_ref:
if ref['worker_map'] == dtrain_ref['worker_map']:
local_evals.append((local_dtrain, name))
continue
local_evals.append((_dmatrix_from_worker_map(**ref), name))
local_history = {}
local_param = params.copy() # just to be consistent
msg = 'Overriding `nthreads` defined in dask worker.'
if 'nthread' in local_param.keys() and \
local_param['nthread'] is not None and \
local_param['nthread'] != worker.nthreads:
msg += '`nthread` is specified. ' + msg
LOGGER.warning(msg)
elif 'n_jobs' in local_param.keys() and \
local_param['n_jobs'] is not None and \
local_param['n_jobs'] != worker.nthreads:
msg = '`n_jobs` is specified. ' + msg
LOGGER.warning(msg)
else:
local_param['nthread'] = worker.nthreads
bst = worker_train(params=local_param,
dtrain=local_dtrain,
*args,
evals_result=local_history,
evals=local_evals,
early_stopping_rounds=early_stopping_rounds,
**kwargs)
ret = {'booster': bst, 'history': local_history}
if local_dtrain.num_row() == 0:
ret = None
return ret
def __init__(self, args):
self.args = args
worker = distributed_get_worker()
self.args.append(
('DMLC_TASK_ID=[xgboost.dask]:' + str(worker.address)).encode())