def using_quantile_device_dmatrix(client: Client, train_dir, model_file, fs, do_wait=False):
'''`DaskDeviceQuantileDMatrix` is a data type specialized for `gpu_hist`, tree
method that reduces memory overhead. When training on GPU pipeline, it's
preferred over `DaskDMatrix`.
.. versionadded:: 1.2.0
'''
colnames = ['label'] + ['feature-%02d' % i for i in range(1, 29)]
df = dask_cudf.read_csv(train_dir, header=None, names=colnames, chunksize=None)
X = df[df.columns.difference(['label'])]
y = df['label']
print("[INFO]: ------ CSV files are read from" + train_dir)
if do_wait is True:
df = df.persist()
X = X.persist()
wait(df)
wait(X)
print("[INFO]: ------ Long waited but the data is ready now")
# `DaskDeviceQuantileDMatrix` is used instead of `DaskDMatrix`, be careful
# that it can not be used for anything else than training.
start_time = time.time()
dtrain = dxgb.DaskDeviceQuantileDMatrix(client, X, y)
print("[INFO]: ------ QuantileDMatrix is formed in {} seconds ---".format((time.time() - start_time)))
del df
del X
del y
start_time = time.time()
output = xgb.dask.train(client,
{ 'verbosity': 2,
'learning_rate': 0.1,
'max_depth': 8,
'objective': 'reg:squarederror',
'subsample': 0.5,
'gamma': 0.9,
'verbose_eval': True,
'tree_method':'gpu_hist',
#'nthread':1
},
dtrain,
num_boost_round=100, evals=[(dtrain, 'train')])
print("[INFO]: ------ Training is completed in {} seconds ---".format((time.time() - start_time)))
history = output['history']
print('[INFO]: ------ Training evaluation history:', history)
output['booster'].save_model('/tmp/tmp.model')
fs.put('/tmp/tmp.model', model_file)
print("[INFO]: ------ Model saved here:{}".format( model_file))
def test_empty_partition(self,
local_cuda_cluster: LocalCUDACluster) -> None:
import dask_cudf
import cudf
import cupy
with Client(local_cuda_cluster) as client:
mult = 100
df = cudf.DataFrame({
"a": [1, 2, 3, 4, 5.1] * mult,
"b": [10, 15, 29.3, 30, 31] * mult,
"y": [10, 20, 30, 40., 50] * mult,
})
parameters = {"tree_method": "gpu_hist", "debug_synchronize": True}
empty = df.iloc[:0]
ddf = dask_cudf.concat(
[dask_cudf.from_cudf(empty, npartitions=1)] +
[dask_cudf.from_cudf(df, npartitions=3)] +
[dask_cudf.from_cudf(df, npartitions=3)])
X = ddf[ddf.columns.difference(["y"])]
y = ddf[["y"]]
dtrain = dxgb.DaskDeviceQuantileDMatrix(client, X, y)
bst_empty = xgb.dask.train(client,
parameters,
dtrain,
evals=[(dtrain, "train")])
predt_empty = dxgb.predict(client, bst_empty, X).compute().values
ddf = dask_cudf.concat([dask_cudf.from_cudf(df, npartitions=3)] +
[dask_cudf.from_cudf(df, npartitions=3)])
X = ddf[ddf.columns.difference(["y"])]
y = ddf[["y"]]
dtrain = dxgb.DaskDeviceQuantileDMatrix(client, X, y)
bst = xgb.dask.train(client,
parameters,
dtrain,
evals=[(dtrain, "train")])
predt = dxgb.predict(client, bst, X).compute().values
cupy.testing.assert_allclose(predt, predt_empty)
def using_quantile_device_dmatrix(client: Client, X, y):
'''`DaskDeviceQuantileDMatrix` is a data type specialized for `gpu_hist`, tree
method that reduces memory overhead. When training on GPU pipeline, it's
preferred over `DaskDMatrix`.
.. versionadded:: 1.2.0
'''
# Input must be on GPU for `DaskDeviceQuantileDMatrix`.
X = X.map_blocks(cp.array)
y = y.map_blocks(cp.array)
# `DaskDeviceQuantileDMatrix` is used instead of `DaskDMatrix`, be careful
# that it can not be used for anything else than training.
dtrain = dxgb.DaskDeviceQuantileDMatrix(client, X, y)
output = xgb.dask.train(client,
{'verbosity': 2,
'tree_method': 'gpu_hist'},
dtrain,
num_boost_round=4)
prediction = xgb.dask.predict(client, output, X)
return prediction