[onert-micro] Enable Training

onert-micro/onert-micro/CMakeLists.txt

@@ -13,6 +13,13 @@ else()
     set(KERNEL_REGISTER_FILE ${OM_KERNELS_BUILD_LIST})
 endif()
 
+# if no path is specified, use classic to KernelsToBuild list, else use generated one
+if (NOT OM_KERNELS_TRAIN_LIST)
+    set(KERNEL_TRAIN_REGISTER_FILE "${OM_PAL_DIR}/KernelsToTrain.lst")
+else()
+    set(KERNEL_TRAIN_REGISTER_FILE ${OM_KERNELS_TRAIN_LIST})
+endif()
+
 set(PASS_REGISTER_FILE "${OM_INCLUDE_DIR}/optimize/BuildPass.lst")
 set(CUSTOM_KERNEL_REGISTER_FILE "${OM_PAL_DIR}/CustomKernelsToBuild.lst")
 

onert-micro/onert-micro/include/OMConfig.h

@@ -78,6 +78,7 @@ struct OMTrainingContext
   float epsilon = 10e-8;
   uint32_t num_step = 0;
   uint32_t num_epoch = 0;
+  uint32_t epochs = 0;
 };
 
 /*

onert-micro/onert-micro/include/pal/mcu/KernelsToTrain.lst

@@ -0,0 +1,5 @@
+REGISTER_TRAIN_KERNEL(FULLY_CONNECTED, FullyConnected)
+REGISTER_TRAIN_KERNEL(SOFTMAX, Softmax)
+REGISTER_TRAIN_KERNEL(RESHAPE, Reshape)
+REGISTER_TRAIN_KERNEL(CONV_2D, Conv2D)
+REGISTER_TRAIN_KERNEL(MAX_POOL_2D, MaxPool2D)

onert-micro/onert-micro/src/CMakeLists.txt

@@ -10,9 +10,14 @@ set(OM_INCLUDE_IMPORT_DIR "${OM_INCLUDE_DIR}/import")
 #define optimize path
 set(OM_SOURCE_OPTIMIZE_DIR "${OM_SOURCE_DIR}/optimize")
 set(OM_INCLUDE_OPTIMIZE_DIR "${OM_INCLUDE_DIR}/optimize")
+#define train path
+set(OM_SOURCE_TRAIN_DIR "${OM_SOURCE_DIR}/train")
+set(OM_INCLUDE_TRAIN_DIR "${OM_INCLUDE_DIR}/train")
 
 #OM_Interpreter lib binary name
 set(OM_INTERPRETER_LIB "onert_micro_interpreter")
+#OM_Training_Interpreter lib binary name
+set(OM_TRAINING_INTERPRETER_LIB "onert_micro_training_interpreter")
 #Core lib binary name
 set(OM_CORE_LIB "onert_micro_core${OM_SUFFIX}")
 #Execute lib binary name
@@ -23,6 +28,8 @@ set(OM_IMPORT_LIB "onert_micro_import${OM_SUFFIX}")
 set(OM_OPTIMIZE_LIB "onert_micro_optimize${OM_SUFFIX}")
 #PAL lib binary name
 set(OM_PAL_LIB "onert_micro_pal${OM_SUFFIX}")
+#Train lib binary name
+set(OM_TRAIN_LIB "onert_micro_train${OM_SUFFIX}")
 
 message(STATUS "ONERT MICRO BEGIN")
 
@@ -45,4 +52,20 @@ add_library(${OM_INTERPRETER_LIB} STATIC OMInterpreter.cpp)
 target_include_directories(${OM_INTERPRETER_LIB} PUBLIC "${OM_INCLUDE_DIR}")
 target_link_libraries(${OM_INTERPRETER_LIB} PUBLIC ${OM_CORE_LIB})
 
+# Training part
+message (STATUS "ONERT MICRO TRAINING BEGIN")
+
+#build train lib
+add_subdirectory(${OM_SOURCE_TRAIN_DIR})
+
+target_link_libraries(${OM_CORE_LIB} PUBLIC ${OM_TRAIN_LIB})
+
+add_library(${OM_TRAINING_INTERPRETER_LIB} STATIC OMTrainingInterpreter.cpp)
+
+target_include_directories(${OM_TRAINING_INTERPRETER_LIB} PUBLIC "${OM_INCLUDE_DIR}")
+target_link_libraries(${OM_TRAINING_INTERPRETER_LIB} PUBLIC ${OM_CORE_LIB})
+target_link_libraries(${OM_TRAINING_INTERPRETER_LIB} PUBLIC ${OM_TRAIN_LIB})
+
+message (STATUS "ONERT MICRO TRAINING END")
+
 message(STATUS "ONERT MICRO FINISHED")

onert-micro/onert-micro/src/core/CMakeLists.txt

@@ -4,10 +4,15 @@ set(SOURCES
         OMKernelType.cpp
         OMRuntimeContext.cpp
         OMRuntimeStorage.cpp
+        OMTrainingRuntimeModule.cpp
         OMRuntimeGraph.cpp
         OMRuntimeModule.cpp
         OMUtils.cpp
         OMDataType.cpp
+        train/OMTrainingHandler.cpp
+        train/OMTrainingStorage.cpp
+        train/OMCheckpointSaver.cpp
+        train/OMCheckpointLoader.cpp
         memory/OMMemoryManager.cpp
         memory/OMRuntimeAllocator.cpp
         reader/OMCircleReader.cpp

onert-micro/onert-micro/src/core/memory/OMRuntimeAllocator.cpp

@@ -94,6 +94,9 @@ OMStatus OMRuntimeAllocator::deallocate(size_t kernel_index, OMRuntimeStorage *s
   {
     uint8_t *allocated_data = nullptr;
     OMStatus status = storage->getDataByTensorIndex(&allocated_data, tensor_index);
+    // To continue deallocate due to current tensor is not saved in storage
+    if (allocated_data == nullptr)
+      continue;
     if (status != Ok)
       return status;