Mobile AI Compute Engine Documentation

Welcome to Mobile AI Compute Engine documentation.

The main documentation is organized into the following sections:

Introduction

MACE (Mobile AI Compute Engine) is a deep learning inference framework optimized for mobile heterogeneous computing platforms. MACE provides tools and documents to help users to deploy deep learning models to mobile phones, tablets, personal computers and IoT devices.

Architecture

The following figure shows the overall architecture.

MACE Model

MACE defines a customized model format which is similar to Caffe2. The MACE model can be converted from exported models by TensorFlow and Caffe.

MACE Interpreter

Mace Interpreter mainly parses the NN graph and manages the tensors in the graph.

Runtime

CPU/GPU/DSP runtime correspond to the Ops for different devices.

Workflow

The following figure shows the basic work flow of MACE.

1. Configure model deployment file

Model deploy configuration file (.yml) describes the information of the model and library, MACE will build the library based on the file.

2. Build libraries

Build MACE dynamic or static libraries.

3. Convert model

Convert TensorFlow or Caffe model to MACE model.

4.1. Deploy

Integrate the MACE library into your application and run with MACE API.

4.2. Run (CLI)

MACE provides mace_run command line tool, which could be used to run model and validate model correctness against original TensorFlow or Caffe results.

4.3. Benchmark

MACE provides benchmark tool to get the Op level profiling result of the model.

简介

Mobile AI Compute Engine (MACE) 是一个专为移动端异构计算设备优化的深度学习前向预测框架。 MACE覆盖了常见的移动端计算设备（CPU，GPU和DSP），并且提供了完整的工具链和文档，用户借助MACE能够 很方便地在移动端部署深度学习模型。MACE已经在小米内部广泛使用并且被充分验证具有业界领先的性能和稳定性。

框架

下图描述了MACE的基本框架。

MACE Model

MACE定义了自有的模型格式（类似于Caffe2），通过MACE提供的工具可以将Caffe和TensorFlow的模型 转为MACE模型。

MACE Interpreter

MACE Interpreter主要负责解析运行神经网络图（DAG）并管理网络中的Tensors。

Runtime

CPU/GPU/DSP Runtime对应于各个计算设备的算子实现。

使用流程

下图描述了MACE使用的基本流程。

1. 配置模型部署文件(.yml)

模型部署文件详细描述了需要部署的模型以及生成库的信息，MACE根据该文件最终生成对应的库文件。

2. 编译MACE库

编译MACE的静态库或者动态库。

3. 转换模型

将TensorFlow 或者 Caffe的模型转为MACE的模型。

4.1. 部署

根据不同使用目的集成Build阶段生成的库文件，然后调用MACE相应的接口执行模型。

4.2. 命令行运行

MACE提供了命令行工具，可以在命令行运行模型，可以用来测试模型运行时间，内存占用和正确性。

4.3. Benchmark

MACE提供了命令行benchmark工具，可以细粒度的查看模型中所涉及的所有算子的运行时间。

Environment requirement

MACE requires the following dependencies:

Required dependencies

Software	Installation command	Tested version
Python		2.7
Bazel	bazel installation guide	0.13.0
CMake	apt-get install cmake	>= 3.11.3
Jinja2	pip install -I jinja2==2.10	2.10
PyYaml	pip install -I pyyaml==3.12	3.12.0
sh	pip install -I sh==1.12.14	1.12.14
Numpy	pip install -I numpy==1.14.0	Required by model validation

Optional dependencies

Software	Installation command	Remark
Android NDK	NDK installation guide	Required by Android build, r15b, r15c, r16b, r17b
ADB	apt-get install android-tools-adb	Required by Android run, >= 1.0.32
TensorFlow	pip install -I tensorflow==1.6.0	Required by TensorFlow model
Docker	docker installation guide	Required by docker mode for Caffe model
Scipy	pip install -I scipy==1.0.0	Required by model validation
FileLock	pip install -I filelock==3.0.0	Required by run on Android

Note

• For Android build, ANDROID_NDK_HOME must be confifigured by using export ANDROID_NDK_HOME=/path/to/ndk
• It will link libc++ instead of gnustl if NDK version >= r17b and bazel version >= 0.13.0, please refer to NDK cpp-support.

Using docker

Pull or build docker image

MACE provides docker images with dependencies installed and also Dockerfiles for images building, you can pull the existing ones directly or build them from the Dockerfiles. In most cases, the lite edition image can satisfy developer's basic needs.

Note

It's highly recommended to pull built images.

• lite edition docker image.

# Pull lite edition docker image
docker pull registry.cn-hangzhou.aliyuncs.com/xiaomimace/mace-dev-lite
# Build lite edition docker image
docker build -t registry.cn-hangzhou.aliyuncs.com/xiaomimace/mace-dev-lite ./docker/mace-dev-lite

• full edition docker image (which contains multiple NDK versions and other dev tools).

# Pull full edition docker image
docker pull registry.cn-hangzhou.aliyuncs.com/xiaomimace/mace-dev
# Build full edition docker image
docker build -t registry.cn-hangzhou.aliyuncs.com/xiaomimace/mace-dev ./docker/mace-dev

Note

We will show steps with lite edition later.

Using the image

Create container with the following command

# Create a container named `mace-dev`
docker run -it --privileged -d --name mace-dev \
           -v /dev/bus/usb:/dev/bus/usb --net=host \
           -v /local/path:/container/path \
           registry.cn-hangzhou.aliyuncs.com/xiaomimace/mace-dev-lite
# Execute an interactive bash shell on the container
docker exec -it mace-dev /bin/bash

Manual setup

The setup steps are based on Ubuntu, you can change the commands correspondingly for other systems. For the detailed installation dependencies, please refer to Environment requirement.

Install Bazel

Recommend bazel with version larger than 0.13.0 (Refer to Bazel documentation).

export BAZEL_VERSION=0.13.1
mkdir /bazel && \
    cd /bazel && \
    wget https://github.com/bazelbuild/bazel/releases/download/$BAZEL_VERSION/bazel-$BAZEL_VERSION-installer-linux-x86_64.sh && \
    chmod +x bazel-*.sh && \
    ./bazel-$BAZEL_VERSION-installer-linux-x86_64.sh && \
    cd / && \
    rm -f /bazel/bazel-$BAZEL_VERSION-installer-linux-x86_64.sh

Install Android NDK

The recommended Android NDK versions includes r15b, r15c and r16b (Refers to NDK installation guide).

# Download NDK r15c
cd /opt/ && \
    wget -q https://dl.google.com/android/repository/android-ndk-r15c-linux-x86_64.zip && \
    unzip -q android-ndk-r15c-linux-x86_64.zip && \
    rm -f android-ndk-r15c-linux-x86_64.zip

export ANDROID_NDK_VERSION=r15c
export ANDROID_NDK=/opt/android-ndk-${ANDROID_NDK_VERSION}
export ANDROID_NDK_HOME=${ANDROID_NDK}

# add to PATH
export PATH=${PATH}:${ANDROID_NDK_HOME}

Install extra tools

apt-get install -y --no-install-recommends \
    cmake \
    android-tools-adb
pip install -i http://pypi.douban.com/simple/ --trusted-host pypi.douban.com setuptools
pip install -i http://pypi.douban.com/simple/ --trusted-host pypi.douban.com \
    "numpy>=1.14.0" \
    scipy \
    jinja2 \
    pyyaml \
    sh==1.12.14 \
    pycodestyle==2.4.0 \
    filelock

Install TensorFlow (Optional)

pip install -i http://pypi.douban.com/simple/ --trusted-host pypi.douban.com tensorflow==1.6.0

Install Caffe (Optional)

Please follow the installation instruction of Caffe.

Basic usage

Build and run an example model

At first, make sure the environment has been set up correctly already (refer to Environment requirement).

The followings are instructions about how to quickly build and run a provided model in MACE Model Zoo.

Here we use the mobilenet-v2 model as an example.

Commands

1. Pull MACE project.

git clone https://github.com/XiaoMi/mace.git
git fetch --all --tags --prune

# Checkout the latest tag (i.e. release version)
tag_name=`git describe --abbrev=0 --tags`
git checkout tags/${tag_name}

Note

It's highly recommanded to use a release version instead of master branch.

2. Pull MACE Model Zoo project.

git clone https://github.com/XiaoMi/mace-models.git

3. Build a generic MACE library.

cd path/to/mace
# Build library
# output lib path: builds/lib
bash tools/build-standalone-lib.sh

Note

• Libraries in builds/lib/armeabi-v7a/cpu_gpu/ means it can run on cpu or gpu devices.
• The results in builds/lib/armeabi-v7a/cpu_gpu_dsp/ need HVX supported.

4. Convert the pre-trained mobilenet-v2 model to MACE format model.

cd path/to/mace
# Build library
python tools/converter.py convert --config=/path/to/mace-models/mobilenet-v2/mobilenet-v2.yml

5. Run the model.

Note

If you want to run on device/phone, please plug in at least one device/phone.

# Run example
python tools/converter.py run --config=/path/to/mace-models/mobilenet-v2/mobilenet-v2.yml --example

# Test model run time
python tools/converter.py run --config=/path/to/mace-models/mobilenet-v2/mobilenet-v2.yml --round=100

# Validate the correctness by comparing the results against the
# original model and framework, measured with cosine distance for similarity.
python tools/converter.py run --config=/path/to/mace-models/mobilenet-v2/mobilenet-v2.yml --validate

Build your own model

This part will show you how to use your own pre-trained model in MACE.

1. Prepare your model

MACE now supports models from TensorFlow and Caffe (more frameworks will be supported).

• TensorFlow

  Prepare your pre-trained TensorFlow model.pb file.

  Use Graph Transform Tool to optimize your model for inference. This tool will improve the efficiency of inference by making several optimizations like operators folding, redundant node removal etc. We strongly recommend MACE users to use it before building.

  Usage for CPU/GPU,

  # CPU/GPU:
  ./transform_graph \
      --in_graph=/path/to/your/tf_model.pb \
      --out_graph=/path/to/your/output/tf_model_opt.pb \
      --inputs='input node name' \
      --outputs='output node name' \
      --transforms='strip_unused_nodes(type=float, shape="1,64,64,3")
          strip_unused_nodes(type=float, shape="1,64,64,3")
          remove_nodes(op=Identity, op=CheckNumerics)
          fold_constants(ignore_errors=true)
          flatten_atrous_conv
          fold_batch_norms
          fold_old_batch_norms
          remove_control_dependencies
          strip_unused_nodes
          sort_by_execution_order'
  
   Usage for DSP,
  

  # DSP:
  ./transform_graph \
      --in_graph=/path/to/your/tf_model.pb \
      --out_graph=/path/to/your/output/tf_model_opt.pb \
      --inputs='input node name' \
      --outputs='output node name' \
      --transforms='strip_unused_nodes(type=float, shape="1,64,64,3")
          strip_unused_nodes(type=float, shape="1,64,64,3")
          remove_nodes(op=Identity, op=CheckNumerics)
          fold_constants(ignore_errors=true)
          fold_batch_norms
          fold_old_batch_norms
          backport_concatv2
          quantize_weights(minimum_size=2)
          quantize_nodes
          remove_control_dependencies
          strip_unused_nodes
          sort_by_execution_order'
  

• Caffe

  Caffe 1.0+ models are supported in MACE converter tool.

  If your model is from lower version Caffe, you need to upgrade it by using the Caffe built-in tool before converting.

  # Upgrade prototxt
  $CAFFE_ROOT/build/tools/upgrade_net_proto_text MODEL.prototxt MODEL.new.prototxt
  
  # Upgrade caffemodel
  $CAFFE_ROOT/build/tools/upgrade_net_proto_binary MODEL.caffemodel MODEL.new.caffemodel
  

2. Create a deployment file for your model

When converting a model or building a library, MACE needs to read a YAML file which is called model deployment file here.

A model deployment file contains all the information of your model(s) and building options. There are several example deployment files in MACE Model Zoo project.

The following shows two basic usage of deployment files for TensorFlow and Caffe models. Modify one of them and use it for your own case.

• TensorFlow

  # The name of library
  library_name: mobilenet
  target_abis: [arm64-v8a]
  model_graph_format: file
  model_data_format: file
  models:
    mobilenet_v1: # model tag, which will be used in model loading and must be specific.
      platform: tensorflow
      # path to your tensorflow model's pb file. Support local path, http:// and https://
      model_file_path: https://cnbj1.fds.api.xiaomi.com/mace/miai-models/mobilenet-v1/mobilenet-v1-1.0.pb
      # sha256_checksum of your model's pb file.
      # use this command to get the sha256_checksum: sha256sum path/to/your/pb/file
      model_sha256_checksum: 71b10f540ece33c49a7b51f5d4095fc9bd78ce46ebf0300487b2ee23d71294e6
      # define your model's interface
      # if there multiple inputs or outputs, write like blow:
      # subgraphs:
      # - input_tensors:
      #     - input0
      #     - input1
      #   input_shapes:
      #     - 1,224,224,3
      #     - 1,224,224,3
      #    output_tensors:
      #      - output0
      #      - output1
      #    output_shapes:
      #      - 1,1001
      #      - 1,1001
      subgraphs:
        - input_tensors:
            - input
          input_shapes:
            - 1,224,224,3
          output_tensors:
            - MobilenetV1/Predictions/Reshape_1
          output_shapes:
            - 1,1001
      # cpu, gpu or cpu+gpu
      runtime: cpu+gpu
      winograd: 0
  

• Caffe

  # The name of library
  library_name: squeezenet-v10
  target_abis: [arm64-v8a]
  model_graph_format: file
  model_data_format: file
  models:
    squeezenet-v10: # model tag, which will be used in model loading and must be specific.
      platform: caffe
      # support local path, http:// and https://
      model_file_path: https://cnbj1.fds.api.xiaomi.com/mace/miai-models/squeezenet/squeezenet-v1.0.prototxt
      weight_file_path: https://cnbj1.fds.api.xiaomi.com/mace/miai-models/squeezenet/squeezenet-v1.0.caffemodel
      # sha256_checksum of your model's graph and data files.
      # get the sha256_checksum: sha256sum path/to/your/file
      model_sha256_checksum: db680cf18bb0387ded9c8e9401b1bbcf5dc09bf704ef1e3d3dbd1937e772cae0
      weight_sha256_checksum: 9ff8035aada1f9ffa880b35252680d971434b141ec9fbacbe88309f0f9a675ce
      # define your model's interface
      # if there multiple inputs or outputs, write like blow:
      # subgraphs:
      # - input_tensors:
      #     - input0
      #     - input1
      #   input_shapes:
      #     - 1,224,224,3
      #     - 1,224,224,3
      #    output_tensors:
      #      - output0
      #      - output1
      #    output_shapes:
      #      - 1,1001
      #      - 1,1001
      subgraphs:
        - input_tensors:
            - data
          input_shapes:
            - 1,227,227,3
          output_tensors:
            - prob
          output_shapes:
            - 1,1,1,1000
      runtime: cpu+gpu
      winograd: 0
  

More details about model deployment file are in Advanced usage.

3. Convert your model

When the deployment file is ready, you can use MACE converter tool to convert your model(s).

python tools/converter.py convert --config=/path/to/your/model_deployment_file.yml

This command will download or load your pre-trained model and convert it to a MACE model proto file and weights data file. The generated model files will be stored in build/${library_name}/model folder.

Warning

Please set model_graph_format: file and model_data_format: file in your deployment file before converting. The usage of model_graph_format: code will be demonstrated in Advanced usage.

4. Build MACE into a library

You could Download the prebuilt MACE Library from Github MACE release page.

Or use bazel to build MACE source code into a library.

cd path/to/mace
# Build library
# output lib path: builds/lib
bash tools/build-standalone-lib.sh

The above command will generate dynamic library builds/lib/${ABI}/${DEVICES}/libmace.so and static library builds/lib/${ABI}/${DEVICES}/libmace.a.

Warning

Please verify that the target_abis param in the above command and your deployment file are the same.

5. Run your model

With the converted model, the static or shared library and header files, you can use the following commands to run and validate your model.

Warning

If you want to run on device/phone, please plug in at least one device/phone.

• run

  run the model.

  # Test model run time
  python tools/converter.py run --config=/path/to/your/model_deployment_file.yml --round=100
  
  # Validate the correctness by comparing the results against the
  # original model and framework, measured with cosine distance for similarity.
  python tools/converter.py run --config=/path/to/your/model_deployment_file.yml --validate
  

• benchmark

  benchmark and profile the model.

  # Benchmark model, get detailed statistics of each Op.
  python tools/converter.py benchmark --config=/path/to/your/model_deployment_file.yml
  

6. Deploy your model into applications

You could run model on CPU, GPU and DSP (based on the runtime in your model deployment file). However, there are some differences in different devices.

• CPU

  Almost all of mobile SoCs use ARM-based CPU architecture, so your model could run on different SoCs in theory.

• GPU

  Although most GPUs use OpenCL standard, but there are some SoCs not fully complying with the standard, or the GPU is too low-level to use. So you should have some fallback strategies when the GPU run failed.

• DSP

  MACE only support Qualcomm DSP.

In the converting and building steps, you've got the static/shared library, model files and header files.

${library_name} is the name you defined in the first line of your deployment YAML file.

Note

When linking generated libmace.a into shared library, version script is helpful for reducing a specified set of symbols to local scope.

• The generated static library files are organized as follows,

builds
├── include
│   └── mace
│       └── public
│           ├── mace.h
│           └── mace_runtime.h
├── lib
│   ├── arm64-v8a
│   │   └── cpu_gpu
│   │       ├── libmace.a
│   │       └── libmace.so
│   ├── armeabi-v7a
│   │   ├── cpu_gpu
│   │   │   ├── libmace.a
│   │   │   └── libmace.so
│   │   └── cpu_gpu_dsp
│   │       ├── libhexagon_controller.so
│   │       ├── libmace.a
│   │       └── libmace.so
│   └── linux-x86-64
│       ├── libmace.a
│       └── libmace.so
└── mobilenet-v1
    ├── model
    │   ├── mobilenet_v1.data
    │   └── mobilenet_v1.pb
    └── _tmp
        └── arm64-v8a
            └── mace_run_static

Please refer to mace/examples/example.ccfor full usage. The following list the key steps.

// Include the headers
#include "mace/public/mace.h"
#include "mace/public/mace_runtime.h"

// 0. Set compiled OpenCL kernel cache, this is used to reduce the
// initialization time since the compiling is too slow. It's suggested
// to set this even when pre-compiled OpenCL program file is provided
// because the OpenCL version upgrade may also leads to kernel
// recompilations.
const std::string file_path ="path/to/opencl_cache_file";
std::shared_ptr<KVStorageFactory> storage_factory(
    new FileStorageFactory(file_path));
ConfigKVStorageFactory(storage_factory);

// 1. Declare the device type (must be same with ``runtime`` in configuration file)
DeviceType device_type = DeviceType::GPU;

// 2. Define the input and output tensor names.
std::vector<std::string> input_names = {...};
std::vector<std::string> output_names = {...};

// 3. Create MaceEngine instance
std::shared_ptr<mace::MaceEngine> engine;
MaceStatus create_engine_status;

// Create Engine from model file
create_engine_status =
    CreateMaceEngineFromProto(model_pb_data,
                              model_data_file.c_str(),
                              input_names,
                              output_names,
                              device_type,
                              &engine);
if (create_engine_status != MaceStatus::MACE_SUCCESS) {
  // fall back to other strategy.
}

// 4. Create Input and Output tensor buffers
std::map<std::string, mace::MaceTensor> inputs;
std::map<std::string, mace::MaceTensor> outputs;
for (size_t i = 0; i < input_count; ++i) {
  // Allocate input and output
  int64_t input_size =
      std::accumulate(input_shapes[i].begin(), input_shapes[i].end(), 1,
                      std::multiplies<int64_t>());
  auto buffer_in = std::shared_ptr<float>(new float[input_size],
                                          std::default_delete<float[]>());
  // Load input here
  // ...

  inputs[input_names[i]] = mace::MaceTensor(input_shapes[i], buffer_in);
}

for (size_t i = 0; i < output_count; ++i) {
  int64_t output_size =
      std::accumulate(output_shapes[i].begin(), output_shapes[i].end(), 1,
                      std::multiplies<int64_t>());
  auto buffer_out = std::shared_ptr<float>(new float[output_size],
                                           std::default_delete<float[]>());
  outputs[output_names[i]] = mace::MaceTensor(output_shapes[i], buffer_out);
}

// 5. Run the model
MaceStatus status = engine.Run(inputs, &outputs);

More details are in Advanced usage.

Advanced usage

This part contains the full usage of MACE.

Overview

As mentioned in the previous part, a model deployment file defines a case of model deployment. The building process includes parsing model deployment file, converting models, building MACE core library and packing generated model libraries.

Deployment file

One deployment file will generate one library normally, but if more than one ABIs are specified, one library will be generated for each ABI. A deployment file can also contain multiple models. For example, an AI camera application may contain face recognition, object recognition, and voice recognition models, all of which can be defined in one deployment file.

• Example

  Here is an example deployment file with two models.

  # The name of library
  library_name: mobile_squeeze
  # host, armeabi-v7a or arm64-v8a
  target_abis: [arm64-v8a]
  # The build mode for model(s).
  # 'code' for transferring model(s) into cpp code, 'file' for keeping model(s) in protobuf file(s) (.pb).
  model_graph_format: code
  # 'code' for transferring model data(s) into cpp code, 'file' for keeping model data(s) in file(s) (.data).
  model_data_format: code
  # One yaml config file can contain multi models' deployment info.
  models:
    mobilenet_v1:
        platform: tensorflow
        model_file_path: https://cnbj1.fds.api.xiaomi.com/mace/miai-models/mobilenet-v1/mobilenet-v1-1.0.pb
        model_sha256_checksum: 71b10f540ece33c49a7b51f5d4095fc9bd78ce46ebf0300487b2ee23d71294e6
        subgraphs:
          - input_tensors:
              - input
            input_shapes:
              - 1,224,224,3
            output_tensors:
              - MobilenetV1/Predictions/Reshape_1
            output_shapes:
              - 1,1001
            validation_inputs_data:
              - https://cnbj1.fds.api.xiaomi.com/mace/inputs/dog.npy
        runtime: cpu+gpu
        limit_opencl_kernel_time: 0
        obfuscate: 0
        winograd: 0
    squeezenet_v11:
        platform: caffe
        model_file_path: http://cnbj1-inner-fds.api.xiaomi.net/mace/mace-models/squeezenet/SqueezeNet_v1.1/model.prototxt
        weight_file_path: http://cnbj1-inner-fds.api.xiaomi.net/mace/mace-models/squeezenet/SqueezeNet_v1.1/weight.caffemodel
        model_sha256_checksum: 625c952063da1569e22d2f499dc454952244d42cd8feca61f05502566e70ae1c
        weight_sha256_checksum: 72b912ace512e8621f8ff168a7d72af55910d3c7c9445af8dfbff4c2ee960142
        subgraphs:
          - input_tensors:
              - data
            input_shapes:
              - 1,227,227,3
            output_tensors:
              - prob
            output_shapes:
              - 1,1,1,1000
        runtime: cpu+gpu
        limit_opencl_kernel_time: 0
        obfuscate: 0
        winograd: 0
  

• Configurations

Options	Usage
library_name	Library name.
target_abis	The target ABI(s) to build, could be 'host', 'armeabi-v7a' or 'arm64-v8a'. If more than one ABIs will be used, separate them by commas.
target_socs	[optional] Build for specific SoCs.
model_graph_format	model graph format, could be 'file' or 'code'. 'file' for converting model graph to ProtoBuf file(.pb) and 'code' for converting model graph to c++ code.
model_data_format	model data format, could be 'file' or 'code'. 'file' for converting model weight to data file(.data) and 'code' for converting model weight to c++ code.
model_name	model name should be unique if there are more than one models. LIMIT: if build_type is code, model_name will be used in c++ code so that model_name must comply with c++ name specification.
platform	The source framework, tensorflow or caffe.
model_file_path	The path of your model file which can be local path or remote URL.
model_sha256_checksum	The SHA256 checksum of the model file.
weight_file_path	[optional] The path of Caffe model weights file.
weight_sha256_checksum	[optional] The SHA256 checksum of Caffe model weights file.
subgraphs	subgraphs key. DO NOT EDIT
input_tensors	The input tensor name(s) (tensorflow) or top name(s) of inputs' layer (caffe). If there are more than one tensors, use one line for a tensor.
output_tensors	The output tensor name(s) (tensorflow) or top name(s) of outputs' layer (caffe). If there are more than one tensors, use one line for a tensor.
input_shapes	The shapes of the input tensors, in NHWC order.
output_shapes	The shapes of the output tensors, in NHWC order.
input_ranges	The numerical range of the input tensors' data, default [-1, 1]. It is only for test.
validation_inputs_data	[optional] Specify Numpy validation inputs. When not provided, [-1, 1] random values will be used.
runtime	The running device, one of [cpu, gpu, dsp, cpu_gpu]. cpu_gpu contains CPU and GPU model definition so you can run the model on both CPU and GPU.
data_type	[optional] The data type used for specified runtime. [fp16_fp32, fp32_fp32] for GPU, default is fp16_fp32, [fp32] for CPU and [uint8] for DSP.
limit_opencl_kernel_time	[optional] Whether splitting the OpenCL kernel within 1 ms to keep UI responsiveness, default is 0.
obfuscate	[optional] Whether to obfuscate the model operator name, default to 0.
winograd	[optional] Which type winograd to use, could be [0, 2, 4]. 0 for disable winograd, 2 and 4 for enable winograd, 4 may be faster than 2 but may take more memory.

Note

Some command tools:

# Get device's soc info.
adb shell getprop | grep platform

# command for generating sha256_sum
sha256sum /path/to/your/file

Advanced usage

There are two common advanced use cases:

• converting model to C++ code.
• tuning GPU kernels for a specific SoC.

• Convert model(s) to C++ code

  Warning

  If you want to use this case, you can just use static mace library.

  • 1. Change the model deployment file(.yml)

    If you want to protect your model, you can convert model to C++ code. there are also two cases:

    • convert model graph to code and model weight to file with below model configuration.

    model_graph_format: code
    model_data_format: file
    

    • convert both model graph and model weight to code with below model configuration.

    model_graph_format: code
    model_data_format: code
    

    Note

    Another model protection method is using obfuscate to obfuscate names of model's operators.

  • 2. Convert model(s) to code

    python tools/converter.py convert --config=/path/to/model_deployment_file.yml
    

    The command will generate ${library_name}.a in builds/${library_name}/model directory and ** .h * in builds/${library_name}/include like the following dir-tree.

    # model_graph_format: code
    # model_data_format: file
    
    builds
      ├── include
      │   └── mace
      │       └── public
      │           ├── mace_engine_factory.h
      │           └── mobilenet_v1.h
      └── model
          ├── mobilenet-v1.a
          └── mobilenet_v1.data
    
    # model_graph_format: code
    # model_data_format: code
    
    builds
      ├── include
      │   └── mace
      │       └── public
      │           ├── mace_engine_factory.h
      │           └── mobilenet_v1.h
      └── model
          └── mobilenet-v1.a
    

  • 3. Deployment

    • Link libmace.a and ${library_name}.a to your target.
    • Refer to mace/examples/example.ccfor full usage. The following list the key steps.

    // Include the headers
    #include "mace/public/mace.h"
    #include "mace/public/mace_runtime.h"
    // If the model_graph_format is code
    #include "mace/public/${model_name}.h"
    #include "mace/public/mace_engine_factory.h"
    
    // ... Same with the code in basic usage
    
    // 4. Create MaceEngine instance
    std::shared_ptr<mace::MaceEngine> engine;
    MaceStatus create_engine_status;
    // Create Engine from compiled code
    create_engine_status =
        CreateMaceEngineFromCode(model_name.c_str(),
                                 model_data_file, // empty string if model_data_format is code
                                 input_names,
                                 output_names,
                                 device_type,
                                 &engine);
    if (create_engine_status != MaceStatus::MACE_SUCCESS) {
      // Report error
    }
    
    // ... Same with the code in basic usage
    

• Tuning for specific SoC's GPU

  If you want to use the GPU of a specific device, you can just specify the target_socs in your YAML file and then tune the MACE lib for it (OpenCL kernels), which may get 1~10% performance improvement.

  • 1. Change the model deployment file(.yml)

    Specify target_socs in your model deployment file(.yml):

    target_socs: [sdm845]
    

    Note

    Get device's soc info: adb shell getprop | grep platform

  • 2. Convert model(s)

    python tools/converter.py convert --config=/path/to/model_deployment_file.yml
    

  • 3. Tuning

    The tools/converter.py will enable automatic tuning for GPU kernels. This usually takes some time to finish depending on the complexity of your model.

    Note

    You should plug in device(s) with the specific SoC(s).

    python tools/converter.py run --config=/path/to/model_deployment_file.yml --validate
    

    The command will generate two files in builds/${library_name}/opencl, like the following dir-tree.

    builds
    └── mobilenet-v2
        ├── model
        │   ├── mobilenet_v2.data
        │   └── mobilenet_v2.pb
        └── opencl
            └── arm64-v8a
               ├── moblinet-v2_compiled_opencl_kernel.MiNote3.sdm660.bin
               └── moblinet-v2_tuned_opencl_parameter.MiNote3.sdm660.bin
    

    • mobilenet-v2-gpu_compiled_opencl_kernel.MI6.msm8998.bin stands for the OpenCL binaries used for your models, which could accelerate the initialization stage. Details please refer to OpenCL Specification.
    • mobilenet-v2-tuned_opencl_parameter.MI6.msm8998.bin stands for the tuned OpenCL parameters for the SoC.

  • 4. Deployment

    • Change the names of files generated above for not collision and push them to your own device's directory.
    • Use like the previous procedure, below lists the key steps differently.

    // Include the headers
    #include "mace/public/mace.h"
    #include "mace/public/mace_runtime.h"
    
    // 0. Set pre-compiled OpenCL binary program file paths and OpenCL parameters file path when available
    if (device_type == DeviceType::GPU) {
      mace::SetOpenCLBinaryPaths(path/to/opencl_binary_paths);
      mace::SetOpenCLParameterPath(path/to/opencl_parameter_file);
    }
    
    // ... Same with the code in basic usage.
    

Useful Commands

• run the model

# Test model run time
python tools/converter.py run --config=/path/to/model_deployment_file.yml --round=100

# Validate the correctness by comparing the results against the
# original model and framework, measured with cosine distance for similarity.
python tools/converter.py run --config=/path/to/model_deployment_file.yml --validate

# Check the memory usage of the model(**Just keep only one model in deployment file**)
python tools/converter.py run --config=/path/to/model_deployment_file.yml --round=10000 &
sleep 5
adb shell dumpsys meminfo | grep mace_run
kill %1

Warning

run rely on convert command, you should convert before run.

• benchmark and profile model

# Benchmark model, get detailed statistics of each Op.
python tools/converter.py benchmark --config=/path/to/model_deployment_file.yml

Warning

benchmark rely on convert command, you should benchmark after convert.

Common arguments

option	type	default	commands	explanation
--omp_num_threads	int	-1	run/benchmark	number of threads
--cpu_affinity_policy	int	1	run/benchmark	0:AFFINITY_NONE/1:AFFINITY_BIG_ONLY/2:AFFINITY_LITTLE_ONLY
--gpu_perf_hint	int	3	run/benchmark	0:DEFAULT/1:LOW/2:NORMAL/3:HIGH
--gpu_perf_hint	int	3	run/benchmark	0:DEFAULT/1:LOW/2:NORMAL/3:HIGH
--gpu_priority_hint	int	3	run/benchmark	0:DEFAULT/1:LOW/2:NORMAL/3:HIGH

Use -h to get detailed help.

python tools/converter.py -h
python tools/converter.py build -h
python tools/converter.py run -h
python tools/converter.py benchmark -h

Reduce Library Size

• dynamic library

  The generated dynamic library by script tools/build-standalone-lib.sh is about 1.6M for armeabi-v7a and 2.1M for arm64-v8a. It can be reduced by modifying some build options.

  • If the models don't need to run on device dsp, change the build option --define hexagon=true to false. And the library will be decreased about 100KB.
  • Futher more, if only cpu device needed, change --define opencl=true to false. This way will reduce half of library size to about 700KB for armeabi-v7a and 1000KB for arm64-v8a

• static library

  • The methods in dynamic library can be useful for static library too. In additional, the static library may also contain model graph and model datas if the configs model_graph_format and model_data_format in deployment file are set to code.
  • It is recommended to use version script and strip feature when linking mace static library. The effect is remarkable.

Operator lists

Operator	Supported	Remark
AVERAGE_POOL_2D	Y
ARGMAX	Y	Only CPU and TensorFlow is supported.
BATCH_NORM	Y	Fusion with activation is supported.
BATCH_TO_SPACE_ND	Y
BIAS_ADD	Y
CAST	Y	Only CPU and TensorFlow model is supported.
CHANNEL_SHUFFLE	Y
CONCATENATION	Y	Only support channel axis concatenation.
CONV_2D	Y	Fusion with BN and activation layer is supported.
CROP	Y	Only Caffe's crop layer is supported (in GPU, offset on channel-dim should be dividable by 4).
DECONV_2D	Y	Supports Caffe's Deconvolution and TensorFlow's tf.layers.conv2d_transpose.
DEPTHWISE_CONV_2D	Y	Only multiplier = 1 is supported; Fusion is supported.
DEPTH_TO_SPACE	Y
DEQUANTIZE	Y	Model quantization will be supported later.
ELEMENT_WISE	Y	ADD/MUL/DIV/MIN/MAX/NEG/ABS/SQR_DIFF/POW/RSQRT/EQUAL
EMBEDDING_LOOKUP	Y	Only support channel axis concatenation.
FULLY_CONNECTED	Y
GROUP_CONV_2D		Caffe model with group count = channel count is supported.
IDENTITY	Y	Only TensorFlow model is supported.
LOCAL_RESPONSE_NORMALIZATION	Y
LOGISTIC	Y
LSTM
MATMUL	Y	Only CPU is supported.
MAX_POOL_2D	Y
PAD	Y
PSROI_ALIGN	Y
PRELU	Y	Only Caffe model is supported
REDUCE_MEAN	Y	Only TensorFlow model is supported. For GPU only H + W axis reduce is supported.
RELU	Y
RELU1	Y
RELU6	Y
RELUX	Y
RESHAPE	Y	Limited support: GPU is full supported, for CPU only supports softmax-like usage.
RESIZE_BILINEAR	Y
RNN
RPN_PROPOSAL_LAYER	Y
SHAPE	Y	Only CPU and TensorFlow is supported.
STACK	Y	Only CPU and TensorFlow is supported.
STRIDEDSLICE	Y	Only CPU and TensorFlow is supported.
SLICE	Y	In TensorFlow, this op is equivalent to SPLIT; Only support channel axis slice.
SOFTMAX	Y
SPACE_TO_BATCH_ND	Y
SPACE_TO_DEPTH	Y
SQEEZE	Y	Only CPU and TensorFlow is supported.
TANH	Y
TRANSPOSE	Y	Only CPU and TensorFlow is supported.

Contributing guide

License

The source file should contain a license header. See the existing files as the example.

Python coding style

Changes to Python code should conform to PEP8 Style Guide for Python Code.

You can use pycodestyle to check the style.

C++ coding style

Changes to C++ code should conform to Google C++ Style Guide.

You can use cpplint to check the style and use clang-format to format the code:

clang-format -style="{BasedOnStyle: google,            \
                      DerivePointerAlignment: false,   \
                      PointerAlignment: Right,         \
                      BinPackParameters: false}" $file

C++ logging guideline

VLOG is used for verbose logging, which is configured by environment variable MACE_CPP_MIN_VLOG_LEVEL. The guideline of VLOG level is as follows:

0. Ad hoc debug logging, should only be added in test or temporary ad hoc
   debugging
1. Important network level Debug/Latency trace log (Op run should never
   generate level 1 vlog)
2. Important op level Latency trace log
3. Unimportant Debug/Latency trace log
4. Verbose Debug/Latency trace log

C++ marco

C++ macros should start with MACE_, except for most common ones like LOG and VLOG.

Adding a new Op

You can create a custom op if it is not supported yet.

To add a custom op, you need to follow these steps:

Define the Op class

Define the new Op class in mace/ops/my_custom_op.h.

#ifndef MACE_OPS_MY_CUSTOM_OP_H_
#define MACE_OPS_MY_CUSTOM_OP_H_

#include "mace/core/operator.h"
#include "mace/kernels/my_custom_op.h"

namespace mace {
namespace ops {

template <DeviceType D, typename T>
class MyCustomOp : public Operator<D, T> {
 public:
  MyCustomOp(const OperatorDef &op_def, Workspace *ws)
      : Operator<D, T>(op_def, ws),
        functor_() {}

  bool Run(StatsFuture *future) override {
    const Tensor *input = this->Input(INPUT);
    Tensor *output = this->Output(OUTPUT);
   
    functor_(input, output, future);
    return true;
  }

 protected:
  OP_INPUT_TAGS(INPUT);
  OP_OUTPUT_TAGS(OUTPUT);

 private:
  kernels::MyCustomOpFunctor<D, T> functor_;
};

}  // namespace ops
}  // namespace mace

#endif  // MACE_OPS_MY_CUSTOM_OP_H_

Register the new Op

Define the Ops registering function in mace/ops/my_custom_op.cc.

#include "mace/ops/my_custom_op.h"

namespace mace {
namespace ops {

void Register_My_Custom_Op(OperatorRegistryBase *op_registry) {
  REGISTER_OPERATOR(op_registry, OpKeyBuilder("my_custom_op")
                                     .Device(DeviceType::CPU)
                                     .TypeConstraint<float>("T")
                                     .Build(),
                    Custom_Op<DeviceType::CPU, float>);

  REGISTER_OPERATOR(op_registry, OpKeyBuilder("my_custom_op")
                                     .Device(DeviceType::OPENCL)
                                     .TypeConstraint<float>("T")
                                     .Build(),
                    Custom_Op<DeviceType::OPENCL, float>);

  REGISTER_OPERATOR(op_registry, OpKeyBuilder("my_custom_op")
                                     .Device(DeviceType::OPENCL)
                                     .TypeConstraint<half>("T")
                                     .Build(),
                    Custom_Op<DeviceType::OPENCL, half>);
}

}  // namespace ops
}  // namespace mace

And then register the new Op in mace/ops/ops_register.cc.

#include "mace/ops/ops_register.h"

namespace mace {

namespace ops {
// Keep in lexicographical order

...

extern void Register_My_Custom_Op(OperatorRegistryBase *op_registry);

...

}  // namespace ops


OperatorRegistry::OperatorRegistry() : OperatorRegistryBase() {
  // Keep in lexicographical order

  ...

  ops::Register_My_Custom_Op(this);

  ...

}

}  // namespace mace

Implement the Op kernel code

You need to implement the CPU kernel in a mace/kernels/my_custom_op.h and optionally OpenCL kernel in mace/kernels/kernels/my_custom_op_opencl.cc and mace/kernels/kernels/cl/my_custom_op.cl. You can also optimize the CPU kernel with NEON.

Add test and benchmark

It's strongly recommended to add unit tests and micro benchmarks for your new Op. If you wish to contribute back, it's required.

Add Op in model converter

You need to add this new Op in the model converter.

Document the new Op

Finally, add an entry in operator table in the document.

How to run tests

To run tests, you need to first cross compile the code, push the binary into the device and then execute the binary. To automate this process, MACE provides tools/bazel_adb_run.py tool.

You need to make sure your device has been connected to your dev pc before running tests.

Run unit tests

MACE use gtest for unit tests.

• Run all unit tests defined in a Bazel target, for example, run ops_test:

  python tools/bazel_adb_run.py --target="//mace/ops:ops_test" \
                                --run_target=True
  

• Run unit tests with gtest filter, for example, run Conv2dOpTest unit tests:

  python tools/bazel_adb_run.py --target="//mace/ops:ops_test" \
                                --run_target=True \
                                --args="--gtest_filter=Conv2dOpTest*"
  

Run micro benchmarks

MACE provides a micro benchmark framework for performance tuning.

• Run all micro benchmarks defined in a Bazel target, for example, run all ops_benchmark micro benchmarks:

  python tools/bazel_adb_run.py --target="//mace/ops:ops_benchmark" \
                                --run_target=True
  

• Run micro benchmarks with regex filter, for example, run all CONV_2D GPU micro benchmarks:

  python tools/bazel_adb_run.py --target="//mace/ops:ops_benchmark" \
                                --run_target=True \
                                --args="--filter=MACE_BM_CONV_2D_.*_GPU"
  

Memory layout

CPU runtime memory layout

The CPU tensor buffer is organized in the following order:

Tensor type	Buffer
Intermediate input/output	NCHW
Convolution Filter	OIHW
Depthwise Convolution Filter	MIHW
1-D Argument, length = W	W

GPU runtime memory layout

GPU runtime implementation base on OpenCL, which uses 2D image with CL_RGBA channel order as the tensor storage. This requires OpenCL 1.2 and above.

The way of mapping the Tensor data to OpenCL 2D image (RGBA) is critical for kernel performance.

In CL_RGBA channel order, each 2D image pixel contains 4 data items. The following tables describe the mapping from different type of tensors to 2D RGBA Image.

Input/Output Tensor

The Input/Output Tensor is stored in NHWC format:

Tensor type	Buffer	Image size [width, height]	Explanation
Channel-Major Input/Output	NHWC	[W * (C+3)/4, N * H]	Default Input/Output format
Height-Major Input/Output	NHWC	[W * C, N * (H+3)/4]	WinogradTransform and MatMul output format
Width-Major Input/Output	NHWC	[(W+3)/4 * C, N * H]	Unused now

Each Pixel of Image contains 4 elements. The below table list the coordination relation between Image and Buffer.

Tensor type	Pixel coordinate relationship	Explanation
Channel-Major Input/Output	P[i, j] = {E[n, h, w, c] | (n=j/H, h=j%H, w=i%W, c=[i/W * 4 + k])}	k=[0, 4)
Height-Major Input/Output	P[i, j] = {E[n, h, w, c] | (n=j%N, h=[j/H*4 + k], w=i%W, c=i/W)}	k=[0, 4)
Width-Major Input/Output	P[i, j] = {E[n, h, w, c] | (n=j/H, h=j%H, w=[i%W*4 + k], c=i/W)}	k=[0, 4)

Filter Tensor

Tensor	Buffer	Image size [width, height]	Explanation
Convolution Filter	OIHW	[I, (O+3)/4 * W * H]	Convolution filter format，There is no difference compared to [H*W*I, (O+3)/4]
Depthwise Convlution Filter	MIHW	[H * W * M, (I+3)/4]	Depthwise-Convolution filter format

Each Pixel of Image contains 4 elements. The below table list the coordination relation between Image and Buffer.

Tensor type	Pixel coordinate relationship	Explanation
Convolution Filter	P[m, n] = {E[o, i, h, w] | (o=[n/HW*4+k], i=m, h=T/W, w=T%W)}	HW= H * W, T=n%HW, k=[0, 4)
Depthwise Convlution Filter	P[m, n] = {E[0, i, h, w] | (i=[n*4+k], h=m/W, w=m%W)}	only support multiplier == 1, k=[0, 4)

1-D Argument Tensor

Tensor type	Buffer	Image size [width, height]	Explanation
1-D Argument	W	[(W+3)/4, 1]	1D argument format, e.g. Bias

Each Pixel of Image contains 4 elements. The below table list the coordination relation between Image and Buffer.

Tensor type	Pixel coordinate relationship	Explanation
1-D Argument	P[i, 0] = {E[w] | w=i*4+k}	k=[0, 4)

Frequently asked questions

Does the tensor data consume extra memory when compiled into C++ code?

When compiled into C++ code, the tensor data will be mmaped by the system loader. For the CPU runtime, the tensor data are used without memory copy. For the GPU and DSP runtime, the tensor data are used once during model initialization. The operating system is free to swap the pages out, however, it still consumes virtual memory addresses. So generally speaking, it takes no extra physical memory. If you are short of virtual memory space (this should be very rare), you can use the option to load the tensor data from data file (can be manually unmapped after initialization) instead of compiled code.

Why is the generated static library file size so huge?

The static library is simply an archive of a set of object files which are intermediate and contain much extra information, please check whether the final binary file size is as expected.

Why is the generated binary file (including shared library) size so huge?

When compiling the model into C++ code, the final binary may contains extra debug symbols, they usually take a lot of space. Try to strip the shared library or binary and make sure you are following best practices to reduce the size of an ELF binary, including disabling C++ exception, disabling RTTI, avoiding C++ iostream, hidden internal functions etc. In most cases, the expected overhead should be less than {model weights size in float32}/2 + 3MB.

OpenCL allocator failed with CL_OUT_OF_RESOURCES

OpenCL runtime usually requires continuous virtual memory for its image buffer, the error will occur when the OpenCL driver can't find the continuous space due to high memory usage or fragmentation. Several solutions can be tried:

• Change the model by reducing its memory usage
• Split the Op with the biggest single memory buffer
• Change from armeabi-v7a to arm64-v8a to expand the virtual address space
• Reduce the memory consumption of other modules of the same process

Why is the performance worse than the official result for the same model?

The power options may not set properly, see mace/public/mace_runtime.h for details.

Why is the UI getting poor responsiveness when running model with GPU runtime?

Try to set limit_opencl_kernel_time to 1. If still not resolved, try to modify the source code to use even smaller time intervals or changed to CPU or DSP runtime.

Why is MACE not working on DSP?

Running models on Hexagon DSP need a few prerequisites for DSP developers:

• You need make sure SOCs of your phone is manufactured by Qualcomm and has HVX supported.
• You need a phone that disables secure boot (once enabled, cannot be reversed, so you probably can only get that type phones from manufacturers)
• You need sign your phone by using testsig provided by Qualcomm. (Download Qualcomm Hexagon SDK first, plugin your phone to PC, run scripts/testsig.py)
• You need install Hexagon nnlib backend by following nnlib README (https://github.com/XiaoMi/nnlib).

Then, there you go. You can run Mace on Hexagon DSP.