saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-06-18 01:54:05 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Update test to use 128 bytes for initializer so it can be allocated externally.

Browse source
This commit is contained in:
Scott McKay 2025-01-07 18:54:38 +10:00
parent d360f76626
commit 0dcf0864d3
2 changed files with 68 additions and 56 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

35
onnxruntime/core/graph/graph.cc
View file

@ -4097,7 +4097,7 @@ ONNX_NAMESPACE::GraphProto Graph::ToGraphProto() const {
// This is used for constructing full path for external data
// if it exists
auto add_initializer = [](TensorList& output_initializers, const TensorProto& initializer) -> Status {
auto add_initializer = [](TensorList& output_initializers, const TensorProto& initializer) -> void {
TensorProto& output = *output_initializers.Add();
output = initializer;
@ -4108,7 +4108,7 @@ ONNX_NAMESPACE::GraphProto Graph::ToGraphProto() const {
onnxruntime::FileOffsetType file_offset;
SafeInt<size_t> tensor_byte_size;
ORT_RETURN_IF_ERROR(utils::GetExternalDataInfo(initializer, ignored, location, file_offset, tensor_byte_size));
ORT_THROW_IF_ERROR(utils::GetExternalDataInfo(initializer, ignored, location, file_offset, tensor_byte_size));
if (location == onnxruntime::utils::kTensorProtoMemoryAddressTag) {
// file_offset is address
@ -4119,8 +4119,6 @@ ONNX_NAMESPACE::GraphProto Graph::ToGraphProto() const {
output.set_raw_data(data, tensor_byte_size);
}
}
return Status::OK();
};
auto* mutable_initializers = result.mutable_initializer();
@ -4129,27 +4127,24 @@ ONNX_NAMESPACE::GraphProto Graph::ToGraphProto() const {
const auto& model_path = ModelPath();
// We want to make sure that sparse initializers do not appear
// as dense duplicates within the initializers list.
if (!sparse_tensor_names_.empty()) {
const auto sparse_end = sparse_tensor_names_.end();
for (const auto& initializer : graph_proto_->initializer()) {
if (sparse_end == sparse_tensor_names_.find(initializer.name())) {
add_initializer(*mutable_initializers, initializer);
} else {
auto& sparse_initializer = *result.add_sparse_initializer();
auto status = utils::DenseTensorToSparseTensorProto(initializer, model_path, sparse_initializer);
ORT_ENFORCE(status.IsOK(), "Failed to convert dense initializer to sparse");
}
}
} else
#else
{
for (const auto& initializer : graph_proto_->initializer()) {
const bool has_sparse_initializers = !sparse_tensor_names_.empty();
const auto sparse_end = sparse_tensor_names_.end();
for (const auto& initializer : graph_proto_->initializer()) {
if (!has_sparse_initializers || sparse_end == sparse_tensor_names_.find(initializer.name())) {
add_initializer(*mutable_initializers, initializer);
} else {
auto& sparse_initializer = *result.add_sparse_initializer();
auto status = utils::DenseTensorToSparseTensorProto(initializer, model_path, sparse_initializer);
ORT_ENFORCE(status.IsOK(), "Failed to convert dense initializer to sparse");
}
}
#else
for (const auto& initializer : graph_proto_->initializer()) {
add_initializer(*mutable_initializers, initializer);
}
#endif
return result;
return result;
}
Status Graph::AddExternalInitializersToGraphProtoImpl(

89
onnxruntime/test/shared_lib/test_model_builder_api.cc
View file

@ -33,7 +33,7 @@ Ort::Session CreateSession(Ort::Env& env,
: default_session_options;
// Set this to save the model if you want to debug.
// session_options.SetOptimizedModelFilePath(ORT_TSTR("model_builder_output.onnx"));
session_options.SetOptimizedModelFilePath(ORT_TSTR("model_builder_output.onnx"));
Ort::Session session(env, graph_api_model, session_options);
@ -141,14 +141,14 @@ TEST(ModelBuilderAPITest, Basic_CApi) {
Ort::ThrowOnError(model_builder_api.CreateGraph(&graph));
//
// Create OrtModel with a Gemm. X input is 3x2, Y input is 2x3, Z output is 3x3.
// Create OrtModel with a Gemm. X input is 3x2, Y input is 2x8, Z output is 3x8.
// X is model input. Y is initializer.
// Set the alpha attribute of the Gemm node to 2.0 to test attribute handling.
//
// model input
OrtTensorTypeAndShapeInfo* tensor_type_info = nullptr;
std::vector<int64_t> input_dims = {3, 2};
std::vector<int64_t> input_dims = {3, 4};
// can use api.SetSymbolicDimensions to set symbolic dimensions.
// the input array should have the same rank as the call to SetDimensions.
// e.g. call SetDimensions with {-1, 3, 2} and SetSymbolicDimensions with {"N", nullptr, nullptr} to create
@ -170,7 +170,7 @@ TEST(ModelBuilderAPITest, Basic_CApi) {
// model outputs
OrtTypeInfo* output_type_info = nullptr;
std::vector<int64_t> output_dims = {3, 3};
std::vector<int64_t> output_dims = {3, 8};
Ort::ThrowOnError(api.CreateTensorTypeAndShapeInfo(&tensor_type_info));
Ort::ThrowOnError(api.SetTensorElementType(tensor_type_info, ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT));
@ -209,11 +209,13 @@ TEST(ModelBuilderAPITest, Basic_CApi) {
node = nullptr; // graph now owns node
// Y input
std::vector<int64_t> y_dims = {2, 3};
deleter.weights.emplace_back(
std::make_unique<std::vector<float>>(std::initializer_list<float>{1.0f, 2.0f, 3.0f,
4.0f, 5.0f, 6.0f}));
// As it's 128 bytes it could either be allocated using CreateTensorAsOrtValue or use existing memory.
// Under 128 bytes must use CreateTensorAsOrtValue.
std::vector<int64_t> y_dims = {4, 8};
deleter.weights.emplace_back(std::make_unique<std::vector<float>>(32));
auto& y_values = *deleter.weights.back();
std::iota(y_values.begin(), y_values.end(), 1.0f);
// create an initializer for the Y input. add to `weights` so the memory remains valid.
OrtValue* y_tensor = nullptr;
@ -228,18 +230,24 @@ TEST(ModelBuilderAPITest, Basic_CApi) {
y_tensor = nullptr; // graph now owns
if (use_constant_node) {
// Test that a Constant node is converted to an intializer
// Test that a Constant node is converted to an initializer
// create Constant node that is used as the Max in a Clip to limit the output
OrtOpAttr* value_attr = nullptr;
float max = 60.0f;
Ort::ThrowOnError(api.CreateOpAttr("value", &max, sizeof(max), ORT_OP_ATTR_FLOAT, &value_attr));
node = CreateNode(model_builder_api, "Constant", "clip_max", {}, {"max"}, {value_attr});
// create Constant nodes for min/max to limit output range
OrtOpAttr* min_attr = nullptr;
float min = 400.0f;
Ort::ThrowOnError(api.CreateOpAttr("value", &min, sizeof(min), ORT_OP_ATTR_FLOAT, &min_attr));
node = CreateNode(model_builder_api, "Constant", "clip_min", {}, {"min"}, {min_attr});
Ort::ThrowOnError(model_builder_api.AddNodeToGraph(graph, node));
node = nullptr; // graph now owns node
node = CreateNode(model_builder_api, "Clip", "Clip1", {gemm_output_name.c_str(), "", "max"}, {"Z"});
OrtOpAttr* max_attr = nullptr;
float max = 900.0f;
Ort::ThrowOnError(api.CreateOpAttr("value", &max, sizeof(max), ORT_OP_ATTR_FLOAT, &max_attr));
node = CreateNode(model_builder_api, "Constant", "clip_max", {}, {"max"}, {max_attr});
Ort::ThrowOnError(model_builder_api.AddNodeToGraph(graph, node));
node = nullptr; // graph now owns node
node = CreateNode(model_builder_api, "Clip", "Clip1", {gemm_output_name.c_str(), "min", "max"}, {"Z"});
Ort::ThrowOnError(model_builder_api.AddNodeToGraph(graph, node));
node = nullptr; // graph now owns node
}
@ -261,22 +269,25 @@ TEST(ModelBuilderAPITest, Basic_CApi) {
std::vector<Input<float>> inputs(1);
auto& input = inputs[0];
input.name = "X";
input.dims = {3, 2};
input.values = {1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f};
input.dims = {3, 4};
input.values = {1.0f, 2.0f, 3.0f, 4.0f,
8.0f, 7.0f, 6.0f, 5.0f,
9.0f, 3.0f, 5.0f, 7.0f};
std::vector<int64_t> expected_dims = {3, 3};
std::vector<int64_t> expected_dims = {3, 8};
ModelBuilderAPI::Model cxx_model(model);
auto session = CreateSession(*ort_env, cxx_model);
std::vector<float> expected_output;
if (use_constant_node) {
expected_output = {18.0f, 24.0f, 30.0f,
38.0f, 52.0f, 60.0f, // clipped
58.0f, 60.0f, 60.0f}; // clipped
// clipped with max 160
expected_output = {400.0f, 400.0f, 400.0f, 400.0f, 420.0f, 440.0f, 460.0f, 480.0f,
596.0f, 648.0f, 700.0f, 752.0f, 804.0f, 856.0f, 900.0f, 900.0f,
592.0f, 640.0f, 688.0f, 736.0f, 784.0f, 832.0f, 880.0f, 900.0f};
} else {
expected_output = {18.0f, 24.0f, 30.0f,
38.0f, 52.0f, 66.0f,
58.0f, 80.0f, 102.0f};
expected_output = {340.0f, 360.0f, 380.0f, 400.0f, 420.0f, 440.0f, 460.0f, 480.0f,
596.0f, 648.0f, 700.0f, 752.0f, 804.0f, 856.0f, 908.0f, 960.0f,
592.0f, 640.0f, 688.0f, 736.0f, 784.0f, 832.0f, 880.0f, 928.0f};
}
TestInference<float>(session, inputs, "Z", expected_dims, expected_output);
@ -306,7 +317,7 @@ TEST(ModelBuilderAPITest, Basic_CxxApi) {
std::vector<ModelBuilderAPI::ValueInfo> graph_outputs;
// model input. it's {3, 2} but use a symbolic dim to test that works.
std::vector<int64_t> input_dims({-1, 2});
std::vector<int64_t> input_dims({-1, 4});
std::vector<std::string> input_symbolic_dims({"multiple_of_3", ""});
TensorTypeAndShapeInfo input_tensor_info(ONNXTensorElementDataType::ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT,
input_dims,
@ -315,7 +326,7 @@ TEST(ModelBuilderAPITest, Basic_CxxApi) {
graph_inputs.emplace_back("X", input_type_info.GetConst());
// model outputs
std::vector<int64_t> output_dims = {-1, 3};
std::vector<int64_t> output_dims = {-1, 8};
std::vector<std::string> output_symbolic_dims({"multiple_of_3", ""});
TensorTypeAndShapeInfo output_tensor_info(ONNXTensorElementDataType::ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT,
output_dims,
@ -340,10 +351,14 @@ TEST(ModelBuilderAPITest, Basic_CxxApi) {
// create an initializer for the Y input.
// add to `weights` so it remains valid for the lifetime of the session and we can avoid copying the data.
std::vector<int64_t> y_dims = {2, 3};
weights.emplace_back(std::make_unique<std::vector<float>>(std::initializer_list<float>{1.0f, 2.0f, 3.0f,
4.0f, 5.0f, 6.0f}));
// As it's 128 bytes it could either be allocated using CreateTensorAsOrtValue or use existing memory.
// Under 128 bytes must use CreateTensorAsOrtValue.
std::vector<int64_t> y_dims = {4, 8};
weights.emplace_back(std::make_unique<std::vector<float>>(32));
auto& y_values = *weights.back();
std::iota(y_values.begin(), y_values.end(), 1.0f);
auto info = Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
// if you use this API the initializer data MUST remain valid for the lifetime of the InferenceSession
@ -357,16 +372,18 @@ TEST(ModelBuilderAPITest, Basic_CxxApi) {
std::vector<Input<float>> inputs(1);
auto& input = inputs[0];
input.name = "X";
input.dims = {3, 2};
input.values = {1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f};
input.dims = {3, 4};
input.values = {1.0f, 2.0f, 3.0f, 4.0f,
8.0f, 7.0f, 6.0f, 5.0f,
9.0f, 3.0f, 5.0f, 7.0f};
std::vector<int64_t> expected_dims = {3, 3};
std::vector<int64_t> expected_dims = {3, 8};
auto session = CreateSession(*ort_env, model);
TestInference<float>(session, inputs, "Z", expected_dims,
{18.0f, 24.0f, 30.0f,
38.0f, 52.0f, 66.0f,
58.0f, 80.0f, 102.0f});
{340.0f, 360.0f, 380.0f, 400.0f, 420.0f, 440.0f, 460.0f, 480.0f,
596.0f, 648.0f, 700.0f, 752.0f, 804.0f, 856.0f, 908.0f, 960.0f,
592.0f, 640.0f, 688.0f, 736.0f, 784.0f, 832.0f, 880.0f, 928.0f});
}
TEST(ModelBuilderAPITest, BasicModelEdit_CxxApi) {