Skip to content

Batchnorm training mode support in a minimal build #17103

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

Merged
merged 2 commits into from
Aug 11, 2023
Merged

Batchnorm training mode support in a minimal build #17103

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
20ff81f
Batchnorm training mode support in a minimal build
baijumeswani Aug 10, 2023
84b97e4
Use enable_training_ops instead of enable_training_core
baijumeswani Aug 10, 2023
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
Next Next commit
Batchnorm training mode support in a minimal build
  • Loading branch information
@baijumeswani
baijumeswani committed Aug 10, 2023
commit 20ff81f00ba0e5f100e50d0225f73e346ea32eed
14 changes: 5 additions & 9 deletions onnxruntime/core/providers/cpu/nn/batch_norm.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,10 +29,6 @@

namespace onnxruntime {

#if !defined(ORT_MINIMAL_BUILD)
#define BATCHNORM_INCLUDE_TRAINING_SUPPORT
#endif

template <typename T>
class BatchNorm : public OpKernel {
public:
Expand All @@ -51,7 +47,7 @@ class BatchNorm : public OpKernel {
}

if (is_train_) {
#if defined(BATCHNORM_INCLUDE_TRAINING_SUPPORT)
#ifdef ENABLE_TRAINING_CORE
momentum_ = op_kernel_info.GetAttrOrDefault<float>("momentum", 0.9f);
ORT_ENFORCE(is_spatial_, "Training mode only supports spatial BN");
#else
Expand Down Expand Up @@ -88,7 +84,7 @@ class BatchNorm : public OpKernel {
// calculate sample_size (including all channels)
size_t sample_size_incl_all_channels = sample_size * C;

#if defined(BATCHNORM_INCLUDE_TRAINING_SUPPORT)
#ifdef ENABLE_TRAINING_CORE
AllocatorPtr alloc;
ORT_RETURN_IF_ERROR(p_op_kernel_context->GetTempSpaceAllocator(&alloc));

Expand All @@ -115,7 +111,7 @@ class BatchNorm : public OpKernel {
ConstEigenVectorArrayMap<T> scale_arr(scale->Data<T>(), is_spatial_ ? C : sample_size_incl_all_channels);
ConstEigenVectorArrayMap<T> bias_arr(B->Data<T>(), is_spatial_ ? C : sample_size_incl_all_channels);

#if defined(BATCHNORM_INCLUDE_TRAINING_SUPPORT)
#ifdef ENABLE_TRAINING_CORE
// Note that we only support spatial BN for training
if (is_train_) {
EigenVectorArrayMap<T> saved_mean_arr(saved_mean->MutableData<T>(), C);
Expand Down Expand Up @@ -166,7 +162,7 @@ class BatchNorm : public OpKernel {
ConstEigenVectorArrayMap<T> var_arr(var->Data<T>(), is_spatial_ ? C : sample_size_incl_all_channels);
inv_std = (var_arr + epsilon_).sqrt().inverse();
} else {
#if defined(BATCHNORM_INCLUDE_TRAINING_SUPPORT)
#ifdef ENABLE_TRAINING_CORE
EigenVectorArrayMap<T> saved_inv_std_arr(saved_inv_std->MutableData<T>(), C);
saved_inv_std_arr = (saved_inv_std_arr + epsilon_).inverse().sqrt();
inv_std = saved_inv_std_arr;
Expand All @@ -175,7 +171,7 @@ class BatchNorm : public OpKernel {

// If we're training, do batch normalization based on computation from this batch
ConstEigenVectorArrayMap<T> mean_arr(
#if defined(BATCHNORM_INCLUDE_TRAINING_SUPPORT)
#ifdef ENABLE_TRAINING_CORE
!is_train_ ? mean->Data<T>() : saved_mean->Data<T>(),
#else
mean->Data<T>(),
Expand Down
5 changes: 2 additions & 3 deletions onnxruntime/test/providers/cpu/nn/batch_norm_op_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,7 +2,6 @@
// Licensed under the MIT License.

#include "core/framework/tensor.h"
#include "core/providers/cpu/nn/batch_norm.h" // for BATCHNORM_INCLUDE_TRAINING_SUPPORT
#include "core/session/inference_session.h"
#include "test/common/dnnl_op_test_utils.h"
#include "test/providers/provider_test_utils.h"
Expand Down Expand Up @@ -847,7 +846,7 @@ TEST(BatchNormTest, BatchNorm2d_bfloat16) {
#endif // USE_DNNL

// TODO fix flaky test for CUDA
#ifdef BATCHNORM_INCLUDE_TRAINING_SUPPORT
#ifdef ENABLE_TRAINING_CORE
TEST(BatchNormTest, ForwardTrainingTestWithSavedOutputsOpset9) {
// TODO: Unskip when fixed #41968513
if (DefaultDmlExecutionProvider().get() != nullptr) {
Expand Down Expand Up @@ -937,7 +936,7 @@ TEST(BatchNormTest, ForwardTrainingTestOpset15) {
{kCudaExecutionProvider, kRocmExecutionProvider,
kTensorrtExecutionProvider, kOpenVINOExecutionProvider, kDnnlExecutionProvider});
}
#endif // BATCHNORM_INCLUDE_TRAINING_SUPPORT
#endif // ENABLE_TRAINING_CORE

} // namespace test
} // namespace onnxruntime