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

Remove 'sample' parameter from stats::mean API #2389

Open
wants to merge 15 commits into
base: branch-24.08
Choose a base branch
from
+112 −196
Open

Remove 'sample' parameter from stats::mean API #2389

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
15 commits
Select commit Hold shift + click to select a range
0099a3a
forwards mean, var, stddev & sum implementation to reduce, also added…
mfoerste4 Jul 8, 2024
11afafc
add Kahan sum for coalesced kernels
mfoerste4 Jul 9, 2024
c2f5d8f
improve Kahan sum for strided kernels, also compensate on block
mfoerste4 Jul 9, 2024
e2ff5ac
adjust heuristic for coalesced reduction kernel choice
mfoerste4 Jul 15, 2024
c2cc888
Merge branch 'branch-24.08' into kahan_neumeier_sum
mfoerste4 Jul 16, 2024
6ad9a78
add comments
mfoerste4 Jul 18, 2024
6f0b012
add comment for second API
mfoerste4 Jul 18, 2024
1fceb1a
Merge branch 'branch-24.08' into kahan_neumeier_sum
mfoerste4 Jul 18, 2024
5b9837b
enable sample parameter for column based
mfoerste4 Jul 18, 2024
15e937d
Merge remote-tracking branch 'refs/remotes/origin/kahan_neumeier_sum'…
mfoerste4 Jul 18, 2024
0461fd3
fix covariance should always use mean with sample=false
mfoerste4 Jul 18, 2024
f2a76be
fix stddev and var for sample=true
mfoerste4 Jul 24, 2024
21d71b6
Merge branch 'branch-24.08' into kahan_neumeier_sum
mfoerste4 Jul 24, 2024
876b9ea
remove sample paramemter from mean API
mfoerste4 Jul 24, 2024
75630d8
Merge branch 'branch-24.08' into remove_mean_sample
tfeher Jul 24, 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
5 changes: 2 additions & 3 deletions cpp/include/raft/stats/detail/mean.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,10 +27,9 @@ namespace stats {
namespace detail {

template <typename Type, typename IdxType = int>
void mean(
Type* mu, const Type* data, IdxType D, IdxType N, bool sample, bool rowMajor, cudaStream_t stream)
void mean(Type* mu, const Type* data, IdxType D, IdxType N, bool rowMajor, cudaStream_t stream)
Copy link
Member

Choose a reason for hiding this comment

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

What's the reason for removing the ability to compute a sample-based mean vs a population-based mean? This feature seems useful.

Copy link
Collaborator Author

Choose a reason for hiding this comment

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

This is correct for some statistical metrics like stddev or var, but IIUC that theory does not hold for the mean. The sample mean is the closest we have to the population mean.

Copy link
Contributor

Choose a reason for hiding this comment

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

Both sample and population mean shall have 1/N factor. That gives the unbiased estimate. (This is unlike stdev or variance where we need to use different factor for sample and population mean).

{
Type ratio = Type(1) / ((sample) ? Type(N - 1) : Type(N));
Type ratio = Type(1) / Type(N);
raft::linalg::reduce(mu,
data,
D,
Expand Down
2 changes: 1 addition & 1 deletion cpp/include/raft/stats/detail/scores.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -59,7 +59,7 @@ math_t r2_score(math_t* y, math_t* y_hat, int n, cudaStream_t stream)
{
rmm::device_scalar<math_t> y_bar(stream);

raft::stats::mean(y_bar.data(), y, 1, n, false, false, stream);
raft::stats::mean(y_bar.data(), y, 1, n, false, stream);
RAFT_CUDA_TRY(cudaPeekAtLastError());

rmm::device_uvector<math_t> sse_arr(n, stream);
Expand Down
16 changes: 4 additions & 12 deletions cpp/include/raft/stats/mean.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
/*
* Copyright (c) 2018-2023, NVIDIA CORPORATION.
* Copyright (c) 2018-2024, 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 @@ -38,17 +38,13 @@ namespace stats {
* @param data: the input matrix
* @param D: number of columns of data
* @param N: number of rows of data
* @param sample: whether to evaluate sample mean or not. In other words,
* whether
* to normalize the output using N-1 or N, for true or false, respectively
* @param rowMajor: whether the input data is row or col major
* @param stream: cuda stream
*/
template <typename Type, typename IdxType = int>
void mean(
Type* mu, const Type* data, IdxType D, IdxType N, bool sample, bool rowMajor, cudaStream_t stream)
void mean(Type* mu, const Type* data, IdxType D, IdxType N, bool rowMajor, cudaStream_t stream)
{
detail::mean(mu, data, D, N, sample, rowMajor, stream);
detail::mean(mu, data, D, N, rowMajor, stream);
}

/**
Expand All @@ -67,14 +63,11 @@ void mean(
* @param[in] handle the raft handle
* @param[in] data: the input matrix
* @param[out] mu: the output mean vector
* @param[in] sample: whether to evaluate sample mean or not. In other words, whether
* to normalize the output using N-1 or N, for true or false, respectively
*/
template <typename value_t, typename idx_t, typename layout_t>
void mean(raft::resources const& handle,
raft::device_matrix_view<const value_t, idx_t, layout_t> data,
raft::device_vector_view<value_t, idx_t> mu,
bool sample)
raft::device_vector_view<value_t, idx_t> mu)
{
static_assert(
std::is_same_v<layout_t, raft::row_major> || std::is_same_v<layout_t, raft::col_major>,
Expand All @@ -86,7 +79,6 @@ void mean(raft::resources const& handle,
data.data_handle(),
data.extent(1),
data.extent(0),
sample,
std::is_same_v<layout_t, raft::row_major>,
resource::get_cuda_stream(handle));
}
Expand Down
2 changes: 1 addition & 1 deletion cpp/test/neighbors/ann_ivf_flat.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -288,7 +288,7 @@ class AnnIVFFlatTest : public ::testing::TestWithParam<AnnIvfFlatInputs<IdxT>> {
(IdxT)ps.dim,
stream_);
raft::stats::mean<float, uint32_t>(
centroid.data(), cluster_data.data(), ps.dim, list_sizes[l], false, true, stream_);
centroid.data(), cluster_data.data(), ps.dim, list_sizes[l], true, stream_);
ASSERT_TRUE(raft::devArrMatch(index_2.centers().data_handle() + ps.dim * l,
centroid.data(),
ps.dim,
Expand Down
3 changes: 1 addition & 2 deletions cpp/test/random/rng.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -407,8 +407,7 @@ TEST(Rng, MeanError)
RngState r(seed, rtype);
normal(handle, r, data.data(), len, 3.3f, 0.23f);
// uniform(r, data, len, -1.0, 2.0);
raft::stats::mean(
mean_result.data(), data.data(), num_samples, num_experiments, false, false, stream);
raft::stats::mean(mean_result.data(), data.data(), num_samples, num_experiments, false, stream);
raft::stats::stddev(std_result.data(),
data.data(),
mean_result.data(),
Expand Down
4 changes: 2 additions & 2 deletions cpp/test/stats/cov.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -72,7 +72,7 @@ class CovTest : public ::testing::TestWithParam<CovInputs<T>> {
cov_act.resize(cols * cols, stream);

normal(handle, r, data.data(), len, params.mean, var);
raft::stats::mean(mean_act.data(), data.data(), cols, rows, false, params.rowMajor, stream);
raft::stats::mean(mean_act.data(), data.data(), cols, rows, params.rowMajor, stream);
if (params.rowMajor) {
using layout = raft::row_major;
cov(handle,
Expand Down Expand Up @@ -102,7 +102,7 @@ class CovTest : public ::testing::TestWithParam<CovInputs<T>> {
raft::update_device(data_cm.data(), data_h, 6, stream);
raft::update_device(cov_cm_ref.data(), cov_cm_ref_h, 4, stream);

raft::stats::mean(mean_cm.data(), data_cm.data(), 2, 3, false, false, stream);
raft::stats::mean(mean_cm.data(), data_cm.data(), 2, 3, false, stream);
cov(handle, cov_cm.data(), data_cm.data(), mean_cm.data(), 2, 3, true, false, true, stream);
}

Expand Down
121 changes: 49 additions & 72 deletions cpp/test/stats/mean.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,7 +33,7 @@ template <typename T>
struct MeanInputs {
T tolerance, mean;
int rows, cols;
bool sample, rowMajor;
bool rowMajor;
unsigned long long int seed;
T stddev = (T)1.0;
};
Expand All @@ -42,7 +42,7 @@ template <typename T>
::std::ostream& operator<<(::std::ostream& os, const MeanInputs<T>& dims)
{
return os << "{ " << dims.tolerance << ", " << dims.rows << ", " << dims.cols << ", "
<< dims.sample << ", " << dims.rowMajor << ", " << dims.stddev << "}" << std::endl;
<< ", " << dims.rowMajor << ", " << dims.stddev << "}" << std::endl;
}

template <typename T>
Expand Down Expand Up @@ -74,14 +74,12 @@ class MeanTest : public ::testing::TestWithParam<MeanInputs<T>> {
using layout = raft::row_major;
mean(handle,
raft::make_device_matrix_view<const T, int, layout>(data, rows, cols),
raft::make_device_vector_view<T, int>(mean_act.data(), cols),
params.sample);
raft::make_device_vector_view<T, int>(mean_act.data(), cols));
} else {
using layout = raft::col_major;
mean(handle,
raft::make_device_matrix_view<const T, int, layout>(data, rows, cols),
raft::make_device_vector_view<T, int>(mean_act.data(), cols),
params.sample);
raft::make_device_vector_view<T, int>(mean_act.data(), cols));
}
}

Expand All @@ -98,72 +96,51 @@ class MeanTest : public ::testing::TestWithParam<MeanInputs<T>> {
// measured mean (of a normal distribution) will fall outside of an epsilon of
// 0.15 only 4/10000 times. (epsilon of 0.1 will fail 30/100 times)
const std::vector<MeanInputs<float>> inputsf = {
{0.15f, 1.f, 1024, 32, true, false, 1234ULL},
{0.15f, 1.f, 1024, 64, true, false, 1234ULL},
{0.15f, 1.f, 1024, 128, true, false, 1234ULL},
{0.15f, 1.f, 1024, 256, true, false, 1234ULL},
{0.15f, -1.f, 1024, 32, false, false, 1234ULL},
{0.15f, -1.f, 1024, 64, false, false, 1234ULL},
{0.15f, -1.f, 1024, 128, false, false, 1234ULL},
{0.15f, -1.f, 1024, 256, false, false, 1234ULL},
{0.15f, 1.f, 1024, 32, true, true, 1234ULL},
{0.15f, 1.f, 1024, 64, true, true, 1234ULL},
{0.15f, 1.f, 1024, 128, true, true, 1234ULL},
{0.15f, 1.f, 1024, 256, true, true, 1234ULL},
{0.15f, -1.f, 1024, 32, false, true, 1234ULL},
{0.15f, -1.f, 1024, 64, false, true, 1234ULL},
{0.15f, -1.f, 1024, 128, false, true, 1234ULL},
{0.15f, -1.f, 1024, 256, false, true, 1234ULL},
{0.15f, -1.f, 1030, 1, false, false, 1234ULL},
{0.15f, -1.f, 1030, 60, true, false, 1234ULL},
{2.0f, -1.f, 31, 120, false, false, 1234ULL},
{2.0f, -1.f, 1, 130, false, false, 1234ULL},
{0.15f, -1.f, 1030, 1, false, true, 1234ULL},
{0.15f, -1.f, 1030, 60, true, true, 1234ULL},
{2.0f, -1.f, 31, 120, false, true, 1234ULL},
{2.0f, -1.f, 1, 130, false, true, 1234ULL},
{2.0f, -1.f, 1, 1, false, false, 1234ULL},
{2.0f, -1.f, 1, 1, false, true, 1234ULL},
{2.0f, -1.f, 7, 23, false, false, 1234ULL},
{2.0f, -1.f, 7, 23, false, true, 1234ULL},
{2.0f, -1.f, 17, 5, false, false, 1234ULL},
{2.0f, -1.f, 17, 5, false, true, 1234ULL},
{0.0001f, 0.1f, 1 << 27, 2, false, false, 1234ULL, 0.0001f},
{0.0001f, 0.1f, 1 << 27, 2, false, true, 1234ULL, 0.0001f}};

const std::vector<MeanInputs<double>> inputsd = {
{0.15, 1.0, 1024, 32, true, false, 1234ULL},
{0.15, 1.0, 1024, 64, true, false, 1234ULL},
{0.15, 1.0, 1024, 128, true, false, 1234ULL},
{0.15, 1.0, 1024, 256, true, false, 1234ULL},
{0.15, -1.0, 1024, 32, false, false, 1234ULL},
{0.15, -1.0, 1024, 64, false, false, 1234ULL},
{0.15, -1.0, 1024, 128, false, false, 1234ULL},
{0.15, -1.0, 1024, 256, false, false, 1234ULL},
{0.15, 1.0, 1024, 32, true, true, 1234ULL},
{0.15, 1.0, 1024, 64, true, true, 1234ULL},
{0.15, 1.0, 1024, 128, true, true, 1234ULL},
{0.15, 1.0, 1024, 256, true, true, 1234ULL},
{0.15, -1.0, 1024, 32, false, true, 1234ULL},
{0.15, -1.0, 1024, 64, false, true, 1234ULL},
{0.15, -1.0, 1024, 128, false, true, 1234ULL},
{0.15, -1.0, 1024, 256, false, true, 1234ULL},
{0.15, -1.0, 1030, 1, false, false, 1234ULL},
{0.15, -1.0, 1030, 60, true, false, 1234ULL},
{2.0, -1.0, 31, 120, false, false, 1234ULL},
{2.0, -1.0, 1, 130, false, false, 1234ULL},
{0.15, -1.0, 1030, 1, false, true, 1234ULL},
{0.15, -1.0, 1030, 60, true, true, 1234ULL},
{2.0, -1.0, 31, 120, false, true, 1234ULL},
{2.0, -1.0, 1, 130, false, true, 1234ULL},
{2.0, -1.0, 1, 1, false, false, 1234ULL},
{2.0, -1.0, 1, 1, false, true, 1234ULL},
{2.0, -1.0, 7, 23, false, false, 1234ULL},
{2.0, -1.0, 7, 23, false, true, 1234ULL},
{2.0, -1.0, 17, 5, false, false, 1234ULL},
{2.0, -1.0, 17, 5, false, true, 1234ULL},
{1e-8, 1e-1, 1 << 27, 2, false, false, 1234ULL, 0.0001},
{1e-8, 1e-1, 1 << 27, 2, false, true, 1234ULL, 0.0001}};
{0.15f, -1.f, 1024, 32, false, 1234ULL},
{0.15f, -1.f, 1024, 64, false, 1234ULL},
{0.15f, -1.f, 1024, 128, false, 1234ULL},
{0.15f, -1.f, 1024, 256, false, 1234ULL},
{0.15f, -1.f, 1024, 32, true, 1234ULL},
{0.15f, -1.f, 1024, 64, true, 1234ULL},
{0.15f, -1.f, 1024, 128, true, 1234ULL},
{0.15f, -1.f, 1024, 256, true, 1234ULL},
{0.15f, -1.f, 1030, 1, false, 1234ULL},
{2.0f, -1.f, 31, 120, false, 1234ULL},
{2.0f, -1.f, 1, 130, false, 1234ULL},
{0.15f, -1.f, 1030, 1, true, 1234ULL},
{2.0f, -1.f, 31, 120, true, 1234ULL},
{2.0f, -1.f, 1, 130, true, 1234ULL},
{2.0f, -1.f, 1, 1, false, 1234ULL},
{2.0f, -1.f, 1, 1, true, 1234ULL},
{2.0f, -1.f, 7, 23, false, 1234ULL},
{2.0f, -1.f, 7, 23, true, 1234ULL},
{2.0f, -1.f, 17, 5, false, 1234ULL},
{2.0f, -1.f, 17, 5, true, 1234ULL},
{0.0001f, 0.1f, 1 << 27, 2, false, 1234ULL, 0.0001f},
{0.0001f, 0.1f, 1 << 27, 2, true, 1234ULL, 0.0001f}};

const std::vector<MeanInputs<double>> inputsd = {{0.15, -1.0, 1024, 32, false, 1234ULL},
{0.15, -1.0, 1024, 64, false, 1234ULL},
{0.15, -1.0, 1024, 128, false, 1234ULL},
{0.15, -1.0, 1024, 256, false, 1234ULL},
{0.15, -1.0, 1024, 32, true, 1234ULL},
{0.15, -1.0, 1024, 64, true, 1234ULL},
{0.15, -1.0, 1024, 128, true, 1234ULL},
{0.15, -1.0, 1024, 256, true, 1234ULL},
{0.15, -1.0, 1030, 1, false, 1234ULL},
{2.0, -1.0, 31, 120, false, 1234ULL},
{2.0, -1.0, 1, 130, false, 1234ULL},
{0.15, -1.0, 1030, 1, true, 1234ULL},
{2.0, -1.0, 31, 120, true, 1234ULL},
{2.0, -1.0, 1, 130, true, 1234ULL},
{2.0, -1.0, 1, 1, false, 1234ULL},
{2.0, -1.0, 1, 1, true, 1234ULL},
{2.0, -1.0, 7, 23, false, 1234ULL},
{2.0, -1.0, 7, 23, true, 1234ULL},
{2.0, -1.0, 17, 5, false, 1234ULL},
{2.0, -1.0, 17, 5, true, 1234ULL},
{1e-8, 1e-1, 1 << 27, 2, false, 1234ULL, 0.0001},
{1e-8, 1e-1, 1 << 27, 2, true, 1234ULL, 0.0001}};

typedef MeanTest<float> MeanTestF;
TEST_P(MeanTestF, Result)
Expand Down
Loading
Loading