Merge branch 'mace-api' into 'master'

Add version api.

See merge request !176

mace/BUILD

@@ -32,6 +32,14 @@ config_setting(
     visibility = ["//visibility:public"],
 )
 
+config_setting(
+    name = "enable_neon",
+    define_values = {
+        "neon": "true",
+    },
+    visibility = ["//visibility:public"],
+)
+
 config_setting(
     name = "embed_binary_program",
     define_values = {

mace/codegen/BUILD

@@ -15,6 +15,7 @@ cc_library(
     deps = [
         "//mace/core",
         "//mace/ops",
+        '//mace/codegen:version_lib',
     ] + if_embed_binary_program([
         '//mace/codegen:generated_opencl_lib',
         '//mace/codegen:generated_tuning_lib',
@@ -34,3 +35,13 @@ cc_library(
     copts = ["-std=c++11", "-D_GLIBCXX_USE_C99_MATH_TR1"],
     linkstatic = 1,
 )
+
+cc_library(
+    name = "version_lib",
+    srcs = glob(["version/*.cc"]),
+    copts = ["-std=c++11", "-D_GLIBCXX_USE_C99_MATH_TR1"],
+    linkstatic = 1,
+    deps = [
+        "//mace/core",
+    ]
+)

mace/core/BUILD

@@ -35,7 +35,7 @@ cc_library(
 cc_library(
     name = "core",
     srcs = glob(["*.cc"]),
-    hdrs = glob(["*.h"]),
+    hdrs = glob(["*.h", "public/*.h"]),
     copts = ["-std=c++11", "-D_GLIBCXX_USE_C99_MATH_TR1"],
     linkopts = if_android(["-pie"]),
     deps = [

mace/core/allocator.h

@@ -9,7 +9,7 @@
 #include <malloc.h>
 #include "mace/core/common.h"
 #include "mace/core/registry.h"
-#include "mace/core/mace.h"
+#include "mace/core/public/mace.h"
 #include "mace/core/types.h"
 
 namespace mace {

mace/core/arg_helper.h

@@ -8,7 +8,7 @@
 #include <map>
 
 #include "mace/core/common.h"
-#include "mace/core/mace.h"
+#include "mace/core/public/mace.h"
 
 namespace mace {
 

mace/core/mace.cc

@@ -2,7 +2,7 @@
 // Copyright (c) 2017 XiaoMi All rights reserved.
 //
 
-#include "mace/core/mace.h"
+#include "mace/core/public/mace.h"
 #include "mace/core/types.h"
 #include "mace/core/net.h"
 #include "mace/core/workspace.h"
@@ -495,8 +495,7 @@ MaceEngine::MaceEngine(const NetDef *net_def, DeviceType device_type):
   ws_->CreateTensor("mace_input_node:0", GetDeviceAllocator(device_type_), DT_FLOAT);
   net_ = std::move(CreateNet(*net_def, ws_.get(), device_type));
 }
-MaceEngine::~MaceEngine() {
-}
+MaceEngine::~MaceEngine(){}
 const float *MaceEngine::Run(const float *input,
                              const std::vector<index_t> &input_shape,
                              std::vector<int64_t> &output_shape) {

mace/core/net.h

@@ -6,7 +6,7 @@
 #define MACE_CORE_NET_H_
 
 #include "mace/core/common.h"
-#include "mace/core/mace.h"
+#include "mace/core/public/mace.h"
 
 namespace mace {
 

mace/core/operator.h

@@ -11,7 +11,7 @@
 #include "mace/core/registry.h"
 #include "mace/core/tensor.h"
 #include "mace/core/workspace.h"
-#include "mace/core/mace.h"
+#include "mace/core/public/mace.h"
 
 namespace mace {
 

mace/core/public/version.h

+//
+// Copyright (c) 2017 XiaoMi All rights reserved.
+//
+
+#ifndef MACE_CORE_PUBLIC_VERSION_H_
+#define MACE_CORE_PUBLIC_VERSION_H_
+
+#define MACE_MAJOR_VERSION 0
+#define MACE_MINOR_VERSION 1
+#define MACE_PATCH_VERSION 0
+
+// MACE_VERSION_SUFFIX is non-empty for pre-releases (e.g. "-alpha", "-alpha.1",
+// "-beta", "-rc", "-rc.1")
+#define MACE_VERSION_SUFFIX ""
+
+#define MACE_STR_HELPER(x) #x
+#define MACE_STR(x) MACE_STR_HELPER(x)
+
+// e.g. "0.5.0" or "0.6.0-alpha".
+#define MACE_VERSION_STRING                                            \
+  (MACE_STR(MACE_MAJOR_VERSION) "." MACE_STR(MACE_MINOR_VERSION) "." MACE_STR( \
+      MACE_PATCH_VERSION) MACE_VERSION_SUFFIX)
+
+extern const char *MaceVersion();
+extern const char *MaceGitVersion();
+#endif //  MACE_CORE_PUBLIC_VERSION_H_

mace/core/serializer.h

@@ -7,7 +7,7 @@
 
 #include "mace/core/common.h"
 #include "mace/core/tensor.h"
-#include "mace/core/mace.h"
+#include "mace/core/public/mace.h"
 
 namespace mace {
 

mace/core/tensor.h

@@ -9,7 +9,7 @@
 #include "mace/core/common.h"
 #include "mace/utils/logging.h"
 #include "mace/core/types.h"
-#include "mace/core/mace.h"
+#include "mace/core/public/mace.h"
 
 namespace mace {
 

mace/core/types.h

@@ -6,7 +6,7 @@
 #define MACE_CORE_TYPES_H_
 
 #include "mace/core/common.h"
-#include "mace/core/mace.h"
+#include "mace/core/public/mace.h"
 #include "mace/core/half.h"
 
 

mace/core/workspace.h

@@ -7,7 +7,7 @@
 
 #include "mace/core/common.h"
 #include "mace/core/tensor.h"
-#include "mace/core/mace.h"
+#include "mace/core/public/mace.h"
 
 namespace mace {
 

mace/examples/BUILD

 # Examples
-load("//mace:mace.bzl", "if_android", "if_embed_binary_program")
+load("//mace:mace.bzl", "if_android", "if_enable_neon", "if_embed_binary_program")
 
 cc_binary(
     name = "helloworld",
@@ -7,7 +7,7 @@ cc_binary(
         "helloworld.cc",
     ],
     copts = ["-std=c++11", "-D_GLIBCXX_USE_C99_MATH_TR1"],
-    linkopts = ["-fopenmp"],
+    linkopts = if_enable_neon(["-fopenmp"]),
     deps = [
         "//mace/core",
         "//mace/ops",
@@ -20,7 +20,7 @@ cc_test(
     testonly = 1,
     srcs = ["benchmark_example.cc"],
     copts = ["-std=c++11", "-D_GLIBCXX_USE_C99_MATH_TR1"],
-    linkopts = ["-fopenmp"],
+    linkopts = if_enable_neon(["-fopenmp"]),
     linkstatic = 1,
     deps = [
         "//mace/core",
@@ -32,7 +32,7 @@ cc_binary(
     name = "mace_run",
     srcs = ["mace_run.cc"],
     copts = ["-std=c++11", "-D_GLIBCXX_USE_C99_MATH_TR1"],
-    linkopts = ["-fopenmp"],
+    linkopts = if_enable_neon(["-fopenmp"]),
     linkstatic = 1,
     deps = [
         "//mace/codegen:generated_models_lib",

mace/examples/mace_run.cc

@@ -17,10 +17,12 @@
 #include <iostream>
 #include <cstdlib>
 #include "mace/utils/command_line_flags.h"
-#include "mace/core/mace.h"
 #include "mace/utils/logging.h"
 #include "mace/utils/env_time.h"
 
+#include "mace/core/public/mace.h"
+#include "mace/core/public/version.h"
+
 using namespace std;
 using namespace mace;
 
@@ -82,7 +84,9 @@ int main(int argc, char **argv) {
     return -1;
   }
 
-  VLOG(0) << "model: " << model_file << std::endl
+  VLOG(0) << "mace version: " << MaceVersion() << std::endl
+          << "mace git version: " << MaceGitVersion() << std::endl
+          << "model: " << model_file << std::endl
           << "input: " << input_node << std::endl
           << "output: " << output_node << std::endl
           << "input_shape: " << input_shape << std::endl

mace/kernels/BUILD

@@ -7,22 +7,22 @@ package(
 
 licenses(["notice"])  # Apache 2.0
 
-load("//mace:mace.bzl", "if_android", "if_android_arm64", "if_embed_binary_program")
+load("//mace:mace.bzl", "if_android", "if_enable_neon", "if_embed_binary_program")
 
 cc_library(
     name = "kernels",
-    srcs = glob(["*.cc", "opencl/*.cc"]) + if_android_arm64(glob([
+    srcs = glob(["*.cc", "opencl/*.cc"]) + if_enable_neon(glob([
         "neon/*.cc",
     ])),
-    hdrs = glob(["*.h", "opencl/*.h"]) + if_android_arm64(glob([
+    hdrs = glob(["*.h", "opencl/*.h"]) + if_enable_neon(glob([
         "neon/*.h",
     ])),
     copts = [
         "-std=c++11",
-        "-fopenmp",
         "-D_GLIBCXX_USE_C99_MATH_TR1",
     ] +
-    if_embed_binary_program(["-DMACE_EMBED_BINARY_PROGRAM"]),
+    if_embed_binary_program(["-DMACE_EMBED_BINARY_PROGRAM"]) +
+    if_enable_neon(["-fopenmp"]),
     linkopts = if_android(["-lm"]),
     deps = [
         "//mace/core",

mace/kernels/batch_norm.h

@@ -7,7 +7,7 @@
 
 #include "mace/core/future.h"
 #include "mace/core/tensor.h"
-#include "mace/core/mace.h"
+#include "mace/core/public/mace.h"
 
 namespace mace {
 namespace kernels {

mace/kernels/bias_add.h

@@ -7,7 +7,7 @@
 
 #include "mace/core/future.h"
 #include "mace/core/tensor.h"
-#include "mace/core/mace.h"
+#include "mace/core/public/mace.h"
 
 namespace mace {
 namespace kernels {

mace/kernels/concat.h

@@ -8,7 +8,7 @@
 #include "mace/core/common.h"
 #include "mace/core/future.h"
 #include "mace/core/types.h"
-#include "mace/core/mace.h"
+#include "mace/core/public/mace.h"
 #include "mace/core/tensor.h"
 
 namespace mace {

mace/kernels/depthwise_conv2d.h

@@ -8,7 +8,7 @@
 #include "mace/core/future.h"
 #include "mace/core/common.h"
 #include "mace/kernels/conv_pool_2d_util.h"
-#include "mace/core/mace.h"
+#include "mace/core/public/mace.h"
 
 namespace mace {
 namespace kernels {

mace/kernels/space_to_batch.h

@@ -7,7 +7,7 @@
 
 #include "mace/core/future.h"
 #include "mace/core/tensor.h"
-#include "mace/core/mace.h"
+#include "mace/core/public/mace.h"
 
 namespace mace {
 namespace kernels {

mace/mace.bzl

@@ -35,3 +35,9 @@ def if_embed_binary_program(a):
       "//mace:embed_binary_program": a,
       "//conditions:default": [],
   })
+
+def if_enable_neon(a):
+  return select({
+      "//mace:enable_neon": a,
+      "//conditions:default": [],
+  })

mace/ops/BUILD

@@ -7,7 +7,7 @@ package(
 
 licenses(["notice"])  # Apache 2.0
 
-load("//mace:mace.bzl", "if_android")
+load("//mace:mace.bzl", "if_android", "if_enable_neon")
 
 cc_library(
     name = "test",
@@ -34,7 +34,8 @@ cc_library(
         ["*.h"],
         exclude = ["ops_test_util.h"],
     ),
-    copts = ["-std=c++11", "-D_GLIBCXX_USE_C99_MATH_TR1"],
+    copts = ["-std=c++11", "-D_GLIBCXX_USE_C99_MATH_TR1"] +
+            if_enable_neon(["-DMACE_ENABLE_NEON"]),
     deps = [
         "//mace/kernels",
     ],

mace/ops/addn.cc

@@ -11,12 +11,12 @@ REGISTER_CPU_OPERATOR(OpKeyBuilder("AddN")
                              .Build(),
                       AddNOp<DeviceType::CPU, float>);
 
-#if __ARM_NEON
+#if MACE_ENABLE_NEON
 REGISTER_NEON_OPERATOR(OpKeyBuilder("AddN")
                              .TypeConstraint<float>("T")
                              .Build(),
                        AddNOp<DeviceType::NEON, float>);
-#endif  // __ARM_NEON
+#endif  // MACE_ENABLE_NEON
 
 REGISTER_OPENCL_OPERATOR(OpKeyBuilder("AddN")
                              .TypeConstraint<float>("T")

mace/ops/batch_norm.cc

@@ -11,12 +11,12 @@ REGISTER_CPU_OPERATOR(OpKeyBuilder("BatchNorm")
                              .Build(),
                       BatchNormOp<DeviceType::CPU, float>);
 
-#if __ARM_NEON
+#if MACE_ENABLE_NEON
 REGISTER_NEON_OPERATOR(OpKeyBuilder("BatchNorm")
                              .TypeConstraint<float>("T")
                              .Build(),
                        BatchNormOp<DeviceType::NEON, float>);
-#endif  // __ARM_NEON
+#endif  // MACE_ENABLE_NEON
 
 REGISTER_OPENCL_OPERATOR(OpKeyBuilder("BatchNorm")
                              .TypeConstraint<float>("T")

mace/ops/concat.h

@@ -7,7 +7,6 @@
 
 #include "mace/core/operator.h"
 #include "mace/kernels/concat.h"
-#include "mace/core/mace.h"
 namespace mace {
 
 template <DeviceType D, typename T>

mace/ops/conv_2d.cc

@@ -16,12 +16,12 @@ REGISTER_CPU_OPERATOR(OpKeyBuilder("Conv2D")
                           .Build(),
                       Conv2dOp<DeviceType::CPU, half>);
 
-#if __ARM_NEON
+#if MACE_ENABLE_NEON
 REGISTER_NEON_OPERATOR(OpKeyBuilder("Conv2D")
                              .TypeConstraint<float>("T")
                              .Build(),
                        Conv2dOp<DeviceType::NEON, float>);
-#endif  // __ARM_NEON
+#endif  // MACE_ENABLE_NEON
 
 REGISTER_OPENCL_OPERATOR(OpKeyBuilder("Conv2D")
                              .TypeConstraint<float>("T")

mace/ops/depthwise_conv2d.cc

@@ -11,12 +11,12 @@ REGISTER_CPU_OPERATOR(OpKeyBuilder("DepthwiseConv2d")
                              .Build(),
                       DepthwiseConv2dOp<DeviceType::CPU, float>);
 
-#if __ARM_NEON
+#if MACE_ENABLE_NEON
 REGISTER_NEON_OPERATOR(OpKeyBuilder("DepthwiseConv2d")
                              .TypeConstraint<float>("T")
                              .Build(),
                        DepthwiseConv2dOp<DeviceType::NEON, float>);
-#endif  // __ARM_NEON
+#endif  // MACE_ENABLE_NEON
 
 REGISTER_OPENCL_OPERATOR(OpKeyBuilder("DepthwiseConv2d")
                              .TypeConstraint<float>("T")

mace/ops/global_avg_pooling.cc

@@ -11,11 +11,11 @@ REGISTER_CPU_OPERATOR(OpKeyBuilder("GlobalAvgPooling")
                              .Build(),
                       GlobalAvgPoolingOp<DeviceType::CPU, float>);
 
-#if __ARM_NEON
+#if MACE_ENABLE_NEON
 REGISTER_NEON_OPERATOR(OpKeyBuilder("GlobalAvgPooling")
                              .TypeConstraint<float>("T")
                              .Build(),
                        GlobalAvgPoolingOp<DeviceType::NEON, float>);
-#endif  // __ARM_NEON
+#endif  // MACE_ENABLE_NEON
 
 }  //  namespace mace

mace/ops/global_avg_pooling.h

-//
+//DMACE_ENABLE_NEON
 // Copyright (c) 2017 XiaoMi All rights reserved.
 //
 

mace/ops/ops_test_util.h

@@ -11,6 +11,7 @@
 #include "mace/core/common.h"
 #include "mace/core/net.h"
 #include "mace/core/tensor.h"
+#include "mace/core/workspace.h"
 #include "mace/core/runtime/opencl/opencl_runtime.h"
 #include "mace/kernels/opencl/helper.h"
 #include "mace/utils/utils.h"

mace/ops/pooling.cc

@@ -15,12 +15,12 @@ REGISTER_CPU_OPERATOR(OpKeyBuilder("Pooling")
                           .Build(),
                       PoolingOp<DeviceType::CPU, half>);
 
-#if __ARM_NEON
+#if MACE_ENABLE_NEON
 REGISTER_NEON_OPERATOR(OpKeyBuilder("Pooling")
                              .TypeConstraint<float>("T")
                              .Build(),
                        PoolingOp<DeviceType::NEON, float>);
-#endif  // __ARM_NEON
+#endif  // MACE_ENABLE_NEON
 
 REGISTER_OPENCL_OPERATOR(OpKeyBuilder("Pooling")
                              .TypeConstraint<float>("T")

mace/ops/relu.cc

@@ -11,12 +11,12 @@ REGISTER_CPU_OPERATOR(OpKeyBuilder("Relu")
                           .Build(),
                       ReluOp<DeviceType::CPU, float>);
 
-#if __ARM_NEON
+#if MACE_ENABLE_NEON
 REGISTER_NEON_OPERATOR(OpKeyBuilder("Relu")
                              .TypeConstraint<float>("T")
                              .Build(),
                        ReluOp<DeviceType::NEON, float>);
-#endif  // __ARM_NEON
+#endif  // MACE_ENABLE_NEON
 
 REGISTER_OPENCL_OPERATOR(OpKeyBuilder("Relu")
                              .TypeConstraint<float>("T")

mace/ops/resize_bilinear.cc

@@ -11,12 +11,12 @@ REGISTER_CPU_OPERATOR(OpKeyBuilder("ResizeBilinear")
                           .Build(),
                       ResizeBilinearOp<DeviceType::CPU, float>);
 
-#if __ARM_NEON
+#if MACE_ENABLE_NEON
 REGISTER_NEON_OPERATOR(OpKeyBuilder("ResizeBilinear")
                              .TypeConstraint<float>("T")
                              .Build(),
                        ResizeBilinearOp<DeviceType::NEON, float>);
-#endif  // __ARM_NEON
+#endif  // MACE_ENABLE_NEON
 
 REGISTER_OPENCL_OPERATOR(OpKeyBuilder("ResizeBilinear")
                              .TypeConstraint<float>("T")

mace/python/tools/model.template

@@ -5,7 +5,7 @@
 
 {% if mode == 0 %}
 #include <vector>
-#include "mace/core/mace.h"
+#include "mace/core/public/mace.h"
 
 namespace {{tag}}{
 
@@ -24,7 +24,7 @@ void Create{{tensor.name}}(std::vector<mace::TensorProto> &tensors) {
 {% elif mode == 1 %}
 #include <vector>
 #include <string>
-#include "mace/core/mace.h"
+#include "mace/core/public/mace.h"
 
 namespace {
 static void UpdateOp(mace::OperatorDef &op,
@@ -95,7 +95,7 @@ void CreateOperator{{i}}(mace::OperatorDef &op) {
 {% else %}
 #include <vector>
 #include <string>
-#include "mace/core/mace.h"
+#include "mace/core/public/mace.h"
 
 namespace {{tag}} {
 

mace/tools/git/gen_version_source.sh

+#!/usr/bin/env bash
+#
+# Copyright (c) 2017 XiaoMi All rights reserved.
+#
+
+OUTPUT_FILENAME=$1
+if [[ -z "${OUTPUT_FILENAME}}" ]]; then
+  echo "Usage: $0 <filename>"
+  exit 1
+fi
+
+DATE_STR=$(date +%Y%m%d)
+GIT_VERSION=$(git describe --long --tags)
+if [[ $? != 0 ]]; then
+  GIT_VERSION=unknown
+else
+  GIT_VERSION=${GIT_VERSION}-${DATE_STR}
+fi
+
+cat <<EOF > ${OUTPUT_FILENAME}
+#include "mace/core/public/version.h"
+const char *MaceVersion() { return MACE_VERSION_STRING; }
+const char *MaceGitVersion() { return "${GIT_VERSION}"; }
+EOF

tools/bazel-adb-run.sh

@@ -27,7 +27,11 @@ PROFILINE="--define profiling=true"
 BRANCH=$(git symbolic-ref --short HEAD)
 COMMIT_ID=$(git rev-parse --short HEAD)
 
-bazel build -c opt $STRIP --verbose_failures $BAZEL_TARGET --crosstool_top=//external:android/crosstool --host_crosstool_top=@bazel_tools//tools/cpp:toolchain --cpu=$ANDROID_ABI
+bazel build -c opt $STRIP --verbose_failures $BAZEL_TARGET \
+   --crosstool_top=//external:android/crosstool \
+   --host_crosstool_top=@bazel_tools//tools/cpp:toolchain  \
+   --cpu=$ANDROID_ABI \
+   --define neon=true
 
 if [ $? -ne 0 ]; then
   exit 1

tools/validate_gcn.sh

@@ -27,6 +27,7 @@ CL_CODEGEN_DIR=${CODEGEN_DIR}/opencl
 CL_BIN_DIR=${CODEGEN_DIR}/opencl_bin
 TUNING_CODEGEN_DIR=${CODEGEN_DIR}/tuning
 TUNING_OR_NOT=${3:-0}
+VERSION_SOURCE_PATH=${CODEGEN_DIR}/version
 
 build_and_run()
 {
@@ -50,7 +51,6 @@ build_and_run()
   adb push ${MODEL_DIR}/${INPUT_FILE_NAME} ${PHONE_DATA_DIR}
   adb push bazel-bin/mace/examples/mace_run ${PHONE_DATA_DIR}
 
-  num_threads=${1:-4}
   if [[ "${TUNING_OR_NOT}" != "0" && "$EMBED_OPENCL_BINARY" != true ]];then
     tuning_flag=1
     round=0 # only warm up
@@ -63,7 +63,6 @@ build_and_run()
     MACE_CPP_MIN_VLOG_LEVEL=0 \
     MACE_RUN_PARAMETER_PATH=${PHONE_DATA_DIR}/mace_run.config \
     MACE_KERNEL_PATH=$KERNEL_DIR \
-    OMP_NUM_THREADS=$num_threads \
     ${PHONE_DATA_DIR}/mace_run \
     --model=${PHONE_DATA_DIR}/${MACE_MODEL_NAME} \
     --input=mace_input_node \
@@ -95,10 +94,15 @@ bazel-bin/mace/python/tools/tf_converter --input=${TF_MODEL_FILE_PATH} \
                                          --model_tag=${MODEL_TAG} \
                                          --confuse=True
 
-echo "Step 3: Run model on the phone with files"
+echo "Step 3: Generate version source"
+rm -rf ${VERSION_SOURCE_PATH}
+mkdir -p ${VERSION_SOURCE_PATH}
+bash mace/tools/git/gen_version_source.sh ${VERSION_SOURCE_PATH}/version.cc
+
+echo "Step 4: Run model on the phone with files"
 build_and_run false
 
-echo "Step 4: Generate OpenCL binary program and config code"
+echo "Step 5: Generate OpenCL binary program and config code"
 rm -rf ${CL_BIN_DIR}
 adb pull ${KERNEL_DIR} ${CL_BIN_DIR}
 rm -rf ${CL_CODEGEN_DIR}
@@ -106,20 +110,20 @@ mkdir -p ${CL_CODEGEN_DIR}
 python mace/python/tools/opencl_codegen.py \
   --cl_binary_dir=${CL_BIN_DIR} --output_path=${CL_CODEGEN_DIR}/opencl_compiled_program.cc
 
-echo "Step 5: Generate tuning source file"
+echo "Step 6: Generate tuning source file"
 adb pull ${PHONE_DATA_DIR}/mace_run.config ${CL_BIN_DIR}
 mkdir -p ${TUNING_CODEGEN_DIR}
 python mace/python/tools/binary_codegen.py \
   --binary_file=${CL_BIN_DIR}/mace_run.config --output_path=${TUNING_CODEGEN_DIR}/tuning_params.cc
 
-echo "Step 6: Run model on the phone using binary"
+echo "Step 7: Run model on the phone using binary"
 build_and_run true
 
-echo "Step 7: Pull the mace run result."
+echo "Step 8: Pull the mace run result."
 rm -rf ${MODEL_DIR}/${OUTPUT_FILE_NAME}
 adb </dev/null pull ${PHONE_DATA_DIR}/${OUTPUT_FILE_NAME} ${MODEL_DIR}
 
-echo "Step 8: Validate the result"
+echo "Step 9: Validate the result"
 python tools/validate.py --model_file ${TF_MODEL_FILE_PATH} \
     --input_file ${MODEL_DIR}/${INPUT_FILE_NAME} \
     --mace_out_file ${MODEL_DIR}/${OUTPUT_FILE_NAME} \