Hendrik/xgboost
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Support multiple batches in gpu_hist (#5014)

Browse Source 
* Initial external memory training support for GPU Hist tree method.
This commit is contained in:
Rong Ou
2019-11-15 22:50:20 -08:00
committed by Jiaming Yuan
parent 97abcc7ee2
commit 0afcc55d98
15 changed files with 559 additions and 134 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
tests/python-gpu/test_gpu_updaters.py
View File

@@ -19,17 +19,19 @@ def assert_gpu_results(cpu_results, gpu_results):
datasets = ["Boston", "Cancer", "Digits", "Sparse regression",
"Sparse regression with weights", "Small weights regression"]
test_param = parameter_combinations({
'gpu_id': [0],
'max_depth': [2, 8],
'max_leaves': [255, 4],
'max_bin': [2, 256],
'grow_policy': ['lossguide'],
'single_precision_histogram': [True],
'min_child_weight': [0],
'lambda': [0]})
class TestGPU(unittest.TestCase):
def test_gpu_hist(self):
test_param = parameter_combinations({'gpu_id': [0],
'max_depth': [2, 8],
'max_leaves': [255, 4],
'max_bin': [2, 256],
'grow_policy': ['lossguide']})
test_param.append({'single_precision_histogram': True})
test_param.append({'min_child_weight': 0,
'lambda': 0})
for param in test_param:
param['tree_method'] = 'gpu_hist'
gpu_results = run_suite(param, select_datasets=datasets)
@@ -38,6 +40,19 @@ class TestGPU(unittest.TestCase):
cpu_results = run_suite(param, select_datasets=datasets)
assert_gpu_results(cpu_results, gpu_results)
# NOTE(rongou): Because the `Boston` dataset is too small, this only tests external memory mode
# with a single page. To test multiple pages, set DMatrix::kPageSize to, say, 1024.
def test_external_memory(self):
for param in reversed(test_param):
param['tree_method'] = 'gpu_hist'
param['gpu_page_size'] = 1024
gpu_results = run_suite(param, select_datasets=["Boston"])
assert_results_non_increasing(gpu_results, 1e-2)
ext_mem_results = run_suite(param, select_datasets=["Boston External Memory"])
assert_results_non_increasing(ext_mem_results, 1e-2)
assert_gpu_results(gpu_results, ext_mem_results)
break
def test_with_empty_dmatrix(self):
# FIXME(trivialfis): This should be done with all updaters
kRows = 0