需要在APP里使用tensorflow lite來運行PC端訓練的model.tlite，又想apk的體積最小，嘗試了如下方法：

1. 在gradle里配置

implementation("org.tensorflow:tensorflow-lite:2.16.1")

這樣會引入tensorflow.jar，最終apk的size增加大約2.2M

2. 根據tensorflow官方的優化編譯教程

https://www.tensorflow.org/lite/android/lite_build?spm=5176.28103460.0.0.73711db8niy7UE&hl=zh-cn

針對我們的模型,構建出針對性的TensorFlow Lite AAR，最后集成到apk里，體積增加約1.5M

分析TensorFlow Lite AAR的實現，發現其本質還是通過JNI調用了libtensorflowlite.so，

而這個libtensorflowlite.so，包含了tensorflow lite幾乎所有核心framework代碼，因此肯定很大。

3. 因為我們僅需要用到tensorflow lite里model 初始化，interpreter推理等基礎功能，并不需要tensorflow lite里的其他功能，因此，想要最小，直接在我們的JNI文件里，集成tensorflow lite相關類的源碼進行編譯，應該就能使得體積增加最小化了。

把我們JNI文件依賴的類，比如

#include "tensorflow/lite/interpreter.h"
#include "tensorflow/lite/model.h"

等.h和.cc引入我們的JNI里，一起編譯就行了。

一開始是在android studio里，導入tensorflow lite的源碼， 修改CmakeLists.txt，嘗試編譯出可以獨立運行的JNI so， 但是總是失敗。

最后，把JNI文件，放到tensorflow lite的源碼目錄里，利用tensorflow的編譯工具bazel，編譯成功。然后把生成的milc_jni.so放到app的jniLibs里，成功:

a. 在tensorflow/lite/下創建milc_jni/這個目錄，目錄下創建BUILD，milc_jni.cc,?custom_op_resolver.h和custom_op_resolver.cc

b. 根據我們的模型文件model.tflite里用到的算子，比如，我只用了FULLY_CONNECTED，RELU， LOGISTIC這3個算子，定制精簡算子的Resolver類

custom_op_resolver.h

#ifndef TENSORFLOW_LITE_CUSTOM_OP_RESOLVER_H_
#define TENSORFLOW_LITE_CUSTOM_OP_RESOLVER_H_#include "tensorflow/lite/mutable_op_resolver.h"namespace tflite {class MinimalOpResolver : public MutableOpResolver {public:MinimalOpResolver();
};}  // namespace tflite#endif  // TENSORFLOW_LITE_CUSTOM_OP_RESOLVER_H_

custom_op_resolver.cc

#include "tensorflow/lite/milc_jni/custom_op_resolver.h"
#include "tensorflow/lite/kernels/builtin_op_kernels.h"namespace tflite {
MinimalOpResolver::MinimalOpResolver() {// 使用 kernels::builtin:: 命名空間下的注冊函數AddBuiltin(BuiltinOperator_FULLY_CONNECTED, tflite::ops::builtin::Register_FULLY_CONNECTED());AddBuiltin(BuiltinOperator_RELU, tflite::ops::builtin::Register_RELU());AddBuiltin(BuiltinOperator_LOGISTIC, tflite::ops::builtin::Register_LOGISTIC());
}
}  // namespace tflite

c. 創建JNI文件milc_jni.cc

#include <jni.h>
#include "tensorflow/lite/interpreter.h"
#include "tensorflow/lite/model.h"
#include "tensorflow/lite/milc_jni/custom_op_resolver.h"
#include <android/log.h>#define LOG_TAG "TensorFlowLiteJNI"
#define LOGI(...) __android_log_print(ANDROID_LOG_INFO, LOG_TAG, __VA_ARGS__)
#define LOGE(...) __android_log_print(ANDROID_LOG_ERROR, LOG_TAG, __VA_ARGS__)// 移除所有日志輸出
//#define LOGI(...)
//#define LOGE(...)extern "C" JNIEXPORT jfloat JNICALL
Java_com_xm_j_milc_predictJNI(JNIEnv* env, jobject /* this */, jstring modelPath, jfloatArray inputArray) {const char* modelPathStr = env->GetStringUTFChars(modelPath, nullptr);// 獲取輸入數組jfloat* inputElements = env->GetFloatArrayElements(inputArray, nullptr);jsize inputLength = env->GetArrayLength(inputArray);if (inputLength != 31) {LOGE("Input array length must be 31");env->ReleaseStringUTFChars(modelPath, modelPathStr);env->ReleaseFloatArrayElements(inputArray, inputElements, JNI_ABORT);return -1.0;}// 加載 TensorFlow Lite 模型std::unique_ptr<tflite::FlatBufferModel> model = tflite::FlatBufferModel::BuildFromFile(modelPathStr);if (!model) {LOGE("Failed to load model from %s", modelPathStr);env->ReleaseStringUTFChars(modelPath, modelPathStr);env->ReleaseFloatArrayElements(inputArray, inputElements, JNI_ABORT);return -1.0;}// 創建解釋器//tflite::ops::builtin::BuiltinOpResolver resolver;tflite::MinimalOpResolver resolver;std::unique_ptr<tflite::Interpreter> interpreter;tflite::InterpreterBuilder(*model, resolver)(&interpreter);if (!interpreter) {LOGE("Failed to create interpreter");env->ReleaseStringUTFChars(modelPath, modelPathStr);env->ReleaseFloatArrayElements(inputArray, inputElements, JNI_ABORT);return -1.0;}// 分配張量if (interpreter->AllocateTensors() != kTfLiteOk) {LOGE("Failed to allocate tensors");env->ReleaseStringUTFChars(modelPath, modelPathStr);env->ReleaseFloatArrayElements(inputArray, inputElements, JNI_ABORT);return -1.0;}// 設置輸入float* input = interpreter->typed_input_tensor<float>(0);for (int i = 0; i < inputLength; ++i) {input[i] = inputElements[i];}// 運行推理if (interpreter->Invoke() != kTfLiteOk) {LOGE("Failed to invoke interpreter");env->ReleaseStringUTFChars(modelPath, modelPathStr);env->ReleaseFloatArrayElements(inputArray, inputElements, JNI_ABORT);return -1.0;}// 獲取輸出// 5. 獲取輸出結果float* outputTensor = interpreter->typed_output_tensor<float>(0);// 釋放資源env->ReleaseStringUTFChars(modelPath, modelPathStr);env->ReleaseFloatArrayElements(inputArray, inputElements, JNI_ABORT);return outputTensor[0];  // 直接返回標量值
}

d. 創建BUILD文件

# 自定義操作解析器（僅包含必要算子）
cc_library(name = "custom_op_resolver",srcs = ["custom_op_resolver.cc"],hdrs = ["custom_op_resolver.h"],deps = ["//tensorflow/lite/kernels:builtin_ops",],
)cc_binary(name = "milc_jni.so",srcs = ["milc_jni.cc"],linkshared = True,linkstatic = True,  # 靜態鏈接所有依賴deps = [":custom_op_resolver","//tensorflow/lite:framework","//tensorflow/lite/kernels:builtin_ops","@flatbuffers//:flatbuffers",],copts = ["-Oz","-flto=thin","-ffunction-sections","-fdata-sections","-fvisibility=hidden","-fvisibility-inlines-hidden","-DFLATBUFFERS_RELEASE","-DTF_LITE_STRIP_ERROR_STRINGS=1","-DNDEBUG","-DFORCE_MINIMAL_LOGGING","-fno-exceptions","-fno-rtti","-fno-unwind-tables","-fno-asynchronous-unwind-tables","-ffreestanding",],linkopts = ["-flto=thin","-Wl,--gc-sections","-Wl,--exclude-libs,ALL","-s", "-Wl,--as-needed","-Wl,-z,norelro","-Wl,--build-id=none",  # 移除構建ID"-Wl,--strip-all",  # 徹底去除符號"-nostdlib","-lc","-Wl,--hash-style=gnu",  # 更小的哈希表"-Wl,--compress-debug-sections=zlib",  # 壓縮調試節],features = ["-layering_check",],
)

e. 在tensorflow的源碼目錄里，初始化好環境，AndroidNDK之類的，然后執行編譯

bazel build -c opt   --config=android_arm64   --copt="-DFORCE_DISABLE_ALL_OPS"   --linkopt="-Wl,--gc-sections"   --linkopt="-Wl,--exclude-libs,ALL"   --linkopt="-s" --define=tflite_with_xnnpack=false  --copt="-Os" --copt="-fomit-frame-pointer"   --copt="-ffunction-sections"   --copt="-fdata-sections"   --copt="-fvisibility=hidden"   --copt="-g0" --copt="-DFLATBUFFERS_RELEASE"  //tensorflow/lite/milc_jni:milc_jni.so

然后，就會生成一個milc_jni.so，大約500K，它是可以獨立運行的，不用依賴libtensorflowlite.so，因此，APK的size，也就只會增加約500K。

f.針對生成的milc_jni.so，進一步壓縮優化

sudo apt-get install upx
upx --android-shlib --best --lzma milc_jni.so -o milc_jni_upx.so

最終的milc_jni_upx.so大約200K，因此，APK的size，也就只會增加約200K。