talksik/mediapipe
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Project import generated by Copybara.

Browse Source 
GitOrigin-RevId: bbbbcb4f5174dea33525729ede47c770069157cd
This commit is contained in:
MediaPipe Team
2021-10-18 17:00:29 -04:00
committed by chuoling
parent 33d683c671
commit 1faeaae7e5
75 changed files with 1944 additions and 560 deletions
Download Patch File Download Diff File Expand all files Collapse all files
mediapipe/calculators/tensor/inference_calculator.proto
+22 -1
View File
@@ -34,7 +34,6 @@ option java_outer_classname = "InferenceCalculatorProto";
// }
// }
// }
//
message InferenceCalculatorOptions {
extend mediapipe.CalculatorOptions {
optional InferenceCalculatorOptions ext = 336783863;
@@ -69,8 +68,30 @@ message InferenceCalculatorOptions {
// Load pre-compiled serialized binary cache to accelerate init process.
// Only available for OpenCL delegate on Android.
// Kernel caching will only be enabled if this path is set.
//
// NOTE: binary cache usage may be skipped if valid serialized model,
// specified by "serialized_model_dir", exists.
//
// TODO: update to cached_kernel_dir
optional string cached_kernel_path = 2;
// A dir to load from and save to a pre-compiled serialized model used to
// accelerate init process.
//
// NOTE: available for OpenCL delegate on Android only when
// "use_advanced_gpu_api" is set to true and "model_token" is set
// properly.
//
// NOTE: serialized model takes precedence over binary cache
// specified by "cached_kernel_path", which still can be used if
// serialized model is invalid or missing.
optional string serialized_model_dir = 7;
// Unique token identifying the model. Used in conjunction with
// "serialized_model_dir". It is the caller's responsibility to ensure
// there is no clash of the tokens.
optional string model_token = 8;
// Encapsulated compilation/runtime tradeoffs.
enum InferenceUsage {
UNSPECIFIED = 0;
mediapipe/calculators/tensor/inference_calculator_gl.cc
+43 -14
View File
@@ -20,6 +20,7 @@
#include "absl/memory/memory.h"
#include "absl/status/status.h"
#include "mediapipe/calculators/tensor/inference_calculator.h"
#include "mediapipe/framework/deps/file_path.h"
#include "mediapipe/util/tflite/config.h"
#if MEDIAPIPE_TFLITE_GL_INFERENCE
@@ -49,8 +50,8 @@ class InferenceCalculatorGlImpl
absl::Status Close(CalculatorContext* cc) override;
private:
absl::Status ReadKernelsFromFile();
absl::Status WriteKernelsToFile();
absl::Status ReadGpuCaches();
absl::Status SaveGpuCaches();
absl::Status LoadModel(CalculatorContext* cc);
absl::Status LoadDelegate(CalculatorContext* cc);
absl::Status LoadDelegateAndAllocateTensors(CalculatorContext* cc);
@@ -82,6 +83,8 @@ class InferenceCalculatorGlImpl
bool use_kernel_caching_ = false;
std::string cached_kernel_filename_;
bool use_serialized_model_ = false;
std::string serialized_model_path_;
};
absl::Status InferenceCalculatorGlImpl::UpdateContract(CalculatorContract* cc) {
@@ -114,6 +117,9 @@ absl::Status InferenceCalculatorGlImpl::Open(CalculatorContext* cc) {
tflite_gpu_runner_usage_ = delegate.gpu().usage();
use_kernel_caching_ =
use_advanced_gpu_api_ && delegate.gpu().has_cached_kernel_path();
use_serialized_model_ = use_advanced_gpu_api_ &&
delegate.gpu().has_serialized_model_dir() &&
delegate.gpu().has_model_token();
use_gpu_delegate_ = !use_advanced_gpu_api_;
if (use_kernel_caching_) {
@@ -123,6 +129,12 @@ absl::Status InferenceCalculatorGlImpl::Open(CalculatorContext* cc) {
".ker";
#endif // MEDIAPIPE_ANDROID
}
if (use_serialized_model_) {
#ifdef MEDIAPIPE_ANDROID
serialized_model_path_ = mediapipe::file::JoinPath(
delegate.gpu().serialized_model_dir(), delegate.gpu().model_token());
#endif // MEDIAPIPE_ANDROID
}
// When use_advanced_gpu_api_, model loading is handled in InitTFLiteGPURunner
// for everything.
@@ -210,7 +222,7 @@ absl::Status InferenceCalculatorGlImpl::Process(CalculatorContext* cc) {
return absl::OkStatus();
}
absl::Status InferenceCalculatorGlImpl::WriteKernelsToFile() {
absl::Status InferenceCalculatorGlImpl::SaveGpuCaches() {
#ifdef MEDIAPIPE_ANDROID
if (use_kernel_caching_) {
// Save kernel file.
@@ -220,12 +232,22 @@ absl::Status InferenceCalculatorGlImpl::WriteKernelsToFile() {
MP_RETURN_IF_ERROR(
mediapipe::file::SetContents(cached_kernel_filename_, cache_str));
}
if (use_serialized_model_) {
// Save serialized model file.
ASSIGN_OR_RETURN(std::vector<uint8_t> serialized_model_vec,
tflite_gpu_runner_->GetSerializedModel());
absl::string_view serialized_model(
reinterpret_cast<char*>(serialized_model_vec.data()),
serialized_model_vec.size());
MP_RETURN_IF_ERROR(
mediapipe::file::SetContents(serialized_model_path_, serialized_model));
}
#endif // MEDIAPIPE_ANDROID
return absl::OkStatus();
}
absl::Status InferenceCalculatorGlImpl::Close(CalculatorContext* cc) {
MP_RETURN_IF_ERROR(WriteKernelsToFile());
MP_RETURN_IF_ERROR(SaveGpuCaches());
if (use_gpu_delegate_) {
MP_RETURN_IF_ERROR(gpu_helper_.RunInGlContext([this]() -> Status {
gpu_buffers_in_.clear();
@@ -239,17 +261,24 @@ absl::Status InferenceCalculatorGlImpl::Close(CalculatorContext* cc) {
return absl::OkStatus();
}
absl::Status InferenceCalculatorGlImpl::ReadKernelsFromFile() {
absl::Status InferenceCalculatorGlImpl::ReadGpuCaches() {
#ifdef MEDIAPIPE_ANDROID
if (use_kernel_caching_) {
if (use_kernel_caching_ && File::Exists(cached_kernel_filename_)) {
// Load pre-compiled kernel file.
if (mediapipe::File::Exists(cached_kernel_filename_)) {
std::string cache_str;
MP_RETURN_IF_ERROR(
mediapipe::file::GetContents(cached_kernel_filename_, &cache_str));
std::vector<uint8_t> cache_vec(cache_str.begin(), cache_str.end());
tflite_gpu_runner_->SetSerializedBinaryCache(std::move(cache_vec));
}
std::string cache_str;
MP_RETURN_IF_ERROR(
mediapipe::file::GetContents(cached_kernel_filename_, &cache_str));
std::vector<uint8_t> cache_vec(cache_str.begin(), cache_str.end());
tflite_gpu_runner_->SetSerializedBinaryCache(std::move(cache_vec));
}
if (use_serialized_model_ && File::Exists(serialized_model_path_)) {
// Load serialized model file.
std::string serialized_model_str;
MP_RETURN_IF_ERROR(
file::GetContents(serialized_model_path_, &serialized_model_str));
std::vector<uint8_t> serialized_model_vec(serialized_model_str.begin(),
serialized_model_str.end());
tflite_gpu_runner_->SetSerializedModel(std::move(serialized_model_vec));
}
#endif // MEDIAPIPE_ANDROID
return absl::OkStatus();
@@ -313,7 +342,7 @@ absl::Status InferenceCalculatorGlImpl::InitTFLiteGPURunner(
tflite_gpu_runner_->GetOutputShapes()[i].c};
}
MP_RETURN_IF_ERROR(ReadKernelsFromFile());
MP_RETURN_IF_ERROR(ReadGpuCaches());
MP_RETURN_IF_ERROR(tflite_gpu_runner_->Build());