Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Enable HDBSCAN gpu training and cpu inference #6108

Merged
merged 4 commits into from
Oct 17, 2024
Merged

Enable HDBSCAN gpu training and cpu inference #6108

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
e51419d
passing tests
divyegala Oct 11, 2024
d4b9610
use new hdbscan release
divyegala Oct 16, 2024
01b098e
Merge remote-tracking branch 'upstream/branch-24.12' into hdbscan-gpu…
divyegala Oct 16, 2024
83a156e
remove extra pip install
divyegala Oct 16, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 4 additions & 2 deletions conda/environments/all_cuda-118_arch-x86_64.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,7 +24,7 @@ dependencies:
- fmt>=11.0.2,<12
- gcc_linux-64=11.*
- graphviz
- hdbscan>=0.8.38,<0.8.39
- hdbscan>=0.8.39,<0.8.40
- hypothesis>=6.0,<7
- ipykernel
- ipython
Expand All @@ -51,6 +51,7 @@ dependencies:
- numpydoc
- nvcc_linux-64=11.8
- packaging
- pip
- pydata-sphinx-theme!=0.14.2
- pylibraft==24.12.*,>=0.0.0a0
- pynndescent
Expand All @@ -69,7 +70,6 @@ dependencies:
- scikit-learn==1.5
- scipy>=1.8.0
- seaborn
- setuptools
- spdlog>=1.14.1,<1.15
- sphinx-copybutton
- sphinx-markdown-tables
Expand All @@ -78,4 +78,6 @@ dependencies:
- sysroot_linux-64==2.17
- treelite==4.3.0
- umap-learn==0.5.6
- pip:
- hdbscan @ git+https://github.com/scikit-learn-contrib/hdbscan.git@master
name: all_cuda-118_arch-x86_64
6 changes: 4 additions & 2 deletions conda/environments/all_cuda-125_arch-x86_64.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,7 +26,7 @@ dependencies:
- fmt>=11.0.2,<12
- gcc_linux-64=11.*
- graphviz
- hdbscan>=0.8.38,<0.8.39
- hdbscan>=0.8.39,<0.8.40
- hypothesis>=6.0,<7
- ipykernel
- ipython
Expand All @@ -47,6 +47,7 @@ dependencies:
- numpy>=1.23,<3.0a0
- numpydoc
- packaging
- pip
- pydata-sphinx-theme!=0.14.2
- pylibraft==24.12.*,>=0.0.0a0
- pynndescent
Expand All @@ -65,7 +66,6 @@ dependencies:
- scikit-learn==1.5
- scipy>=1.8.0
- seaborn
- setuptools
- spdlog>=1.14.1,<1.15
- sphinx-copybutton
- sphinx-markdown-tables
Expand All @@ -74,4 +74,6 @@ dependencies:
- sysroot_linux-64==2.17
- treelite==4.3.0
- umap-learn==0.5.6
- pip:
- hdbscan @ git+https://github.com/scikit-learn-contrib/hdbscan.git@master
name: all_cuda-125_arch-x86_64
12 changes: 10 additions & 2 deletions dependencies.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -512,7 +512,7 @@ dependencies:
packages:
- *cython
- dask-ml
- hdbscan>=0.8.38,<0.8.39
- hdbscan>=0.8.39,<0.8.40
- hypothesis>=6.0,<7
- nltk
- numpydoc
Expand All @@ -526,7 +526,15 @@ dependencies:
- statsmodels
- umap-learn==0.5.6
- pynndescent
- setuptools # Needed on Python 3.12 for dask-glm, which requires pkg_resources but Python 3.12 doesn't have setuptools by default
Copy link
Member Author

@divyegala divyegala Oct 16, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

dask-glm was removed by PR #6028 so this is now unnecessary

- output_types: conda
packages:
- pip
- pip:
- hdbscan @ git+https://github.com/scikit-learn-contrib/hdbscan.git@master
- output_types: pyproject
packages:
- hdbscan @ git+https://github.com/scikit-learn-contrib/hdbscan.git@master

test_notebooks:
common:
- output_types: [conda, requirements]
Expand Down
22 changes: 14 additions & 8 deletions python/cuml/cuml/cluster/hdbscan/prediction.pyx
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1,4 @@
# Copyright (c) 2021-2023, NVIDIA CORPORATION.
# Copyright (c) 2021-2024, NVIDIA CORPORATION.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
Expand Down Expand Up @@ -166,19 +166,25 @@ def all_points_membership_vectors(clusterer, batch_size=4096):

# trained on gpu
if not hasattr(clusterer, "_cpu_model"):
# the reference HDBSCAN implementations uses @property
# for attributes without setters available for them,
# so they can't be transferred from the GPU model
# to the CPU model
raise ValueError("Inferring on CPU is not supported yet when the "
"model has been trained on GPU")
clusterer.import_cpu_model()
clusterer.build_cpu_model()
clusterer.gpu_to_cpu()
# These attributes have to be reassigned to the CPU model
# as the raw arrays because the reference HDBSCAN implementation
# reconstructs the objects from the raw arrays
clusterer._cpu_model.condensed_tree_ = \
Comment on lines +172 to +174
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This happens in the setters in the hdbscan library, right?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Essentially it happens in the getters. Here's the issue, consider CondensedTree object:

  1. We use setter to assign CondensedTree object to self.condensed_tree_ from cuML to hdbscan
  2. The getter for self.condensed_tree_ checks if it has a value already. If it does, it assumes that it is raw numpy arrays and creates another CondensedTree object without any value sanitization

That's why I re-assigned the raw arrays, so when hdbscan library internally calls the getters it reconstructs the object correctly.

clusterer.condensed_tree_._raw_tree
clusterer._cpu_model.single_linkage_tree_ = \
clusterer.single_linkage_tree_._linkage
clusterer._cpu_model.minimum_spanning_tree_ = \
clusterer.minimum_spanning_tree_._mst

# this took a long debugging session to figure out, but
# this method on cpu does not work without this copy for some reason
clusterer._cpu_model.prediction_data_.raw_data = \
clusterer._cpu_model.prediction_data_.raw_data.copy()
return cpu_all_points_membership_vectors(clusterer._cpu_model)

# gpu infer, cpu/gpu train
elif device_type == DeviceType.device:
# trained on cpu
if hasattr(clusterer, "_cpu_model"):
Expand Down
5 changes: 2 additions & 3 deletions python/cuml/cuml/tests/test_device_selection.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -932,9 +932,6 @@ def test_nn_methods(train_device, infer_device):
@pytest.mark.parametrize("infer_device", ["cpu", "gpu"])
def test_hdbscan_methods(train_device, infer_device):

if train_device == "gpu" and infer_device == "cpu":
pytest.skip("Can't transfer attributes to cpu for now")

ref_model = refHDBSCAN(
prediction_data=True,
approx_min_span_tree=False,
Expand All @@ -951,11 +948,13 @@ def test_hdbscan_methods(train_device, infer_device):
ref_membership = cpu_all_points_membership_vectors(ref_model)
ref_labels, ref_probs = cpu_approximate_predict(ref_model, X_test_blob)

gen_min_span_tree = train_device == "gpu" and infer_device == "cpu"
model = HDBSCAN(
prediction_data=True,
approx_min_span_tree=False,
max_cluster_size=0,
min_cluster_size=30,
gen_min_span_tree=gen_min_span_tree,
)
with using_device_type(train_device):
trained_labels = model.fit_predict(X_train_blob)
Expand Down
4 changes: 2 additions & 2 deletions python/cuml/pyproject.toml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -113,7 +113,8 @@ classifiers = [
test = [
"cython>=3.0.0",
"dask-ml",
"hdbscan>=0.8.38,<0.8.39",
"hdbscan @ git+https://github.com/scikit-learn-contrib/hdbscan.git@master",
"hdbscan>=0.8.39,<0.8.40",
"hypothesis>=6.0,<7",
"nltk",
"numpydoc",
Expand All @@ -125,7 +126,6 @@ test = [
"pytest==7.*",
"scikit-learn==1.5",
"seaborn",
"setuptools",
"statsmodels",
"umap-learn==0.5.6",
] # This list was generated by `rapids-dependency-file-generator`. To make changes, edit ../../dependencies.yaml and run `rapids-dependency-file-generator`.
Expand Down
Loading