On-Device Inference

Overview

MindSpore Lite is a lightweight deep neural network inference engine that provides the inference function for models on the device side. Models can be trained by MindSpore or imported from third-party, such as TensorFlow Lite, ONNX and Caffe. This tutorial describes how to use and compile MindSpore Lite for its own model.

Figure 1 On-device inference frame diagram

Mindspore Lite’s framework consists of frontend, IR, backend, Lite, RT and micro.

Frontend: It used for model generation. Users can use the model building interface to build models, or transform third-party models into mindspore models.
IR: It includes the tensor definition, operator prototype definition and graph definition of mindspore. The back-end optimization is based on IR.
Backend: Graph optimization and high level optimization are independent of hardware, such as operator fusion and constant folding, while low level optimization is related to hardware; quantization includes weight quantization, activation value quantization and other post training quantization methods.
Lite RT: In the inference runtime, session provides the external interface, kernel registry is operator registry, scheduler is operator heterogeneous scheduler and executor is operator executor. Lite RT and shares with Micro the underlying infrastructure layers such as operator library, memory allocation, runtime thread pool and parallel primitives.
Micro: Code Gen generates .c files according to the model, and infrastructure such as the underlying operator library is shared with Lite RT.

Compilation Method

You need to compile the MindSpore Lite by yourself. This section describes how to perform cross compilation in the Ubuntu environment.

The environment requirements are as follows:

Hardware requirements
- Memory: 1 GB or above
- Hard disk space: 10 GB or above
System requirements
- System is limited on Linux: Ubuntu = 18.04.02LTS
Software dependencies
- cmake >= 3.14.1
- GCC >= 5.4
- Android_NDK r20b
Using MindSpore Lite converter tool needs to add more dependencies：
- autoconf >= 2.69
- libtool >= 2.4.6
- libressl >= 3.1.3
- automake >= 1.11.6
- libevent >= 2.0
- m4 >= 1.4.18
- openssl >= 1.1.1

The compilation procedure is as follows:

Download source code from the code repository.

git clone https://gitee.com/mindspore/mindspore.git -b r0.7

Run the following command in the root directory of the source code to compile MindSpore Lite.
- Compile converter tool:
```
bash build.sh -I x86_64
```
- Compile inference platform:
  
  Set path of ANDROID_NDK:
```
export ANDROID_NDK={$NDK_PATH}/android-ndk-r20b 
```
  According to the phone，you can choosearm64:
```
bash build.sh -I arm64
```
  orarm32:
```
bash build.sh -I arm32
```
Go to the mindspore/output directory of the source code to obtain the compilation result. Unzip mindspore-lite-0.7.0-converter-ubuntu.tar.gz to get the result mindspore-lite-0.7.0 after building.
```
tar -xvf mindspore-lite-0.7.0-converter-ubuntu.tar.gz
```

Use of On-Device Inference

When MindSpore is used to perform model inference in the APK project of an app, preprocessing input is required before model inference. For example, before an image is converted into the tensor format required by MindSpore inference, the image needs to be resized. After MindSpore completes model inference, postprocess the model inference result and sends the processed output to the app.

This section describes how to use MindSpore to perform model inference. The setup of an APK project and pre- and post-processing of model inference are not described here.

To perform on-device model inference using MindSpore, perform the following steps.

Generating an On-Device Model File

After training is complete, load the generated checkpoint file to the defined network.

param_dict = load_checkpoint(ckpt_file_name=ckpt_file_path)
load_param_into_net(net, param_dict)

Call the export API to export the .mindir model file on the device.

export(net, input_data, file_name="./lenet.mindir", file_format='MINDIR')

Take the LeNet network as an example. The generated on-device model file is lenet.mindir. The complete sample code lenet.py is as follows:

import os
import numpy as np
import mindspore.nn as nn
import mindspore.ops.operations as P
import mindspore.context as context
from mindspore.common.tensor import Tensor
from mindspore.train.serialization import export, load_checkpoint, load_param_into_net

class LeNet(nn.Cell):
    def __init__(self):
        super(LeNet, self).__init__()
        self.relu = P.ReLU()
        self.batch_size = 32
        self.conv1 = nn.Conv2d(1, 6, kernel_size=5, stride=1, padding=0, has_bias=False, pad_mode='valid')
        self.conv2 = nn.Conv2d(6, 16, kernel_size=5, stride=1, padding=0, has_bias=False, pad_mode='valid')
        self.pool = nn.MaxPool2d(kernel_size=2, stride=2)
        self.reshape = P.Reshape()
        self.fc1 = nn.Dense(400, 120)
        self.fc2 = nn.Dense(120, 84)
        self.fc3 = nn.Dense(84, 10)
        
    def construct(self, input_x):
        output = self.conv1(input_x)
        output = self.relu(output)
        output = self.pool(output)
        output = self.conv2(output)
        output = self.relu(output)
        output = self.pool(output)
        output = self.reshape(output, (self.batch_size, -1))
        output = self.fc1(output)
        output = self.relu(output)
        output = self.fc2(output)
        output = self.relu(output)
        output = self.fc3(output)
        return output
        
if __name__ == '__main__':
    context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
    seed = 0
    np.random.seed(seed)
    origin_data = np.random.uniform(low=0, high=255, size=(32, 1, 32, 32)).astype(np.float32)
    origin_data.tofile("lenet.bin")
    input_data = Tensor(origin_data)
    net = LeNet()
    ckpt_file_path = "path_to/lenet.ckpt"

    is_ckpt_exist = os.path.exists(ckpt_file_path)
    if is_ckpt_exist:
        param_dict = load_checkpoint(ckpt_file_name=ckpt_file_path)
        load_param_into_net(net, param_dict)
        export(net, input_data, file_name="./lenet.mindir", file_format='MINDIR')
        print("export model success.")
    else:
        print("checkpoint file does not exist.")

In mindspore/output/mindspore-lite-0.7.0/converter directory, calling MindSpore convert tool named converter_lite, convert model file (.mindir) to on_device inference model file (.ms).
```
./converter_lite --fmk=MS --modelFile=./lenet.mindir --outputFile=lenet
```
Result:
```
INFO [converter/converter.cc:146] Runconverter] CONVERTER RESULT: SUCCESS!
```
This means that the model has been successfully converted to the mindspore on_device inference model.

Implementing On-Device Inference

Use the .ms model file and image data as input to create a session and implement inference on the device.

Take the lenet.ms model as an example to implement on-device inference, the specific steps are:

Load the .ms model file to the memory buffer. The ReadFile function needs to be implemented by users, according to the C++ tutorial.
Call the CreateSession API to get a session.
Call the CompileGraph API in previous Session and transport model.
Call the GetInputs API to get input tensor, then set graph information as data, data will be used in to perform model inference.
Call the RunGraph API in the Session to perform inference.
Call the GetOutputs API to obtain the output.

The complete sample code is as follows:

#include <iostream>
#include <fstream>
#include "schema/model_generated.h"
#include "include/model.h"
#include "include/lite_session.h"
#include "include/errorcode.h"
#include "ir/dtype/type_id.h"


char *ReadFile(const char *file, size_t *size) {
  if (file == nullptr) {
    std::cerr << "file is nullptr" << std::endl;
    return nullptr;
  }
  if (size == nullptr) {
    std::cerr << "size is nullptr" << std::endl;
    return nullptr;
  }
  std::ifstream ifs(file);
  if (!ifs.good()) {
    std::cerr << "file: " << file << " is not exist" << std::endl;
    return nullptr;
  }

  if (!ifs.is_open()) {
    std::cerr << "file: " << file << " open failed" << std::endl;
    return nullptr;
  }

  ifs.seekg(0, std::ios::end);
  *size = ifs.tellg();
  std::unique_ptr<char> buf(new char[*size]);

  ifs.seekg(0, std::ios::beg);
  ifs.read(buf.get(), *size);
  ifs.close();

  return buf.release();
}

int main(int argc, const char **argv) {
  size_t model_size;
  std::string model_path = "./lenet.ms";

// 1. Read File
  auto model_buf = ReadFile(model_path.c_str(), &model_size);
  if (model_buf == nullptr) {
    std::cerr << "ReadFile return nullptr" << std::endl;
    return -1;
  }

// 2. Import Model
  auto model = mindspore::lite::Model::Import(model_buf, model_size);
  if (model == nullptr) {
    std::cerr << "Import model failed" << std::endl;
    delete[](model_buf);
    return -1;
  }
  delete[](model_buf);
  auto context = new mindspore::lite::Context;
  context->cpuBindMode = mindspore::lite::NO_BIND;
  context->deviceCtx.type = mindspore::lite::DT_CPU;
  context->threadNum = 4;

// 3. Create Session
  auto session = mindspore::session::LiteSession::CreateSession(context);
  if (session == nullptr) {
    std::cerr << "CreateSession failed" << std::endl;
    return -1;
  }
  delete context;
  auto ret = session->CompileGraph(model.get());
  if (ret != mindspore::lite::RET_OK) {
    std::cerr << "CompileGraph failed" << std::endl;
    delete session;
    return -1;
  }

// 4. Get Inputs
  auto inputs = session->GetInputs();
  if (inputs.size() != 1) {
    std::cerr << "Lenet should has only one input" << std::endl;
    delete session;
    return -1;
  }
  auto in_tensor = inputs.front();
  if (in_tensor == nullptr) {
    std::cerr << "in_tensor is nullptr" << std::endl;
    delete session;
    return -1;
  }
  size_t data_size;
  std::string data_path = "./data.bin";
  auto input_buf = ReadFile(data_path.c_str(), &data_size);
  if (input_buf == nullptr) {
    std::cerr << "ReadFile return nullptr" << std::endl;
    delete session;
    return -1;
  }
  if (in_tensor->Size()!=data_size) {
    std::cerr << "Input data size is not suit for model input" << std::endl;
    delete[](input_buf);
    delete session;
    return -1;
  }
  auto *in_data = in_tensor->MutableData();
  if (in_data == nullptr) {
    std::cerr << "Data of in_tensor is nullptr" << std::endl;
    delete[](input_buf);
    delete session;
    return -1;     
  }
  memcpy(in_data, input_buf, data_size);
  delete[](input_buf);

// 5. Run Graph
  ret = session->RunGraph();
  if (ret != mindspore::lite::RET_OK) {
    std::cerr << "RunGraph failed" << std::endl;
    delete session;
    return -1;
  } 

// 6. Get Outputs
  auto outputs = session->GetOutputs();
  if (outputs.size()!= 1) {
    std::cerr << "Lenet should has only one output" << std::endl;
    delete session;
    return -1;     
  }
  auto out_tensor = outputs.front();
  if (out_tensor == nullptr) {
    std::cerr << "out_tensor is nullptr" << std::endl;
    delete session;
    return -1;     
  }
  if (out_tensor->data_type()!=mindspore::TypeId::kNumberTypeFloat32) {
    std::cerr << "Output of lenet should in float32" << std::endl;
    delete session;
    return -1;     
  }
  auto *out_data = reinterpret_cast<float *>(out_tensor->MutableData());
  if (out_data == nullptr) {
    std::cerr << "Data of out_tensor is nullptr" << std::endl;
    delete session;
    return -1;     
  }
  std::cout << "Output data: "; 
  for (size_t i = 0; i < 10 & i < out_tensor->ElementsNum(); i++) {
  std::cout << " " << out_data[i];
  }
  std::cout << std::endl;
  delete session;
  return 0;
}