MindSpore-based Inference Service Deployment

Linux Ascend Environmental Setup Enterprise Expert

Overview

MindSpore Serving is a lightweight and high-performance service module that helps MindSpore developers efficiently deploy online inference services in the production environment. After completing model training using MindSpore, you can export the MindSpore model and use MindSpore Serving to create an inference service for the model. Currently, only Ascend 910 is supported.

Starting Serving

After MindSpore is installed using pip, the Serving executable program is stored in /{your python path}/lib/python3.7/site-packages/mindspore/ms_serving. Run the following command to start Serving:

ms_serving [--help] [--model_path=<MODEL_PATH>] [--model_name=<MODEL_NAME>] [--port=<PORT1>] 
                    [--rest_api_port=<PORT2>] [--device_id=<DEVICE_ID>]

Parameters are described as follows:

Parameter	Attribute	Function	Parameter Type	Default Value	Value Range
`--help`	Optional	Displays the help information about the startup command.	-	-	-
`--model_path=<MODEL_PATH>`	Mandatory	Path for storing the model to be loaded.	String	Null	-
`--model_name=<MODEL_NAME>`	Mandatory	Name of the model file to be loaded.	String	Null	-
`--port=<PORT>`	Optional	Specifies the external Serving gRPC port number.	Integer	5500	1–65535
`--rest_api_port=<PORT2>`	Specifies the external Serving REST API port number.	Integer	5500	1–65535
`--device_id=<DEVICE_ID>`	Optional	Specifies device ID to be used.	Integer	0	0 to 7

Before running the startup command, add the path /{your python path}/lib:/{your python path}/lib/python3.7/site-packages/mindspore/lib to the environment variable LD_LIBRARY_PATH. port and rest_ api_port cannot be the same.

Application Example

The following uses a simple network as an example to describe how to use MindSpore Serving.

Exporting Model

Before exporting the model, you need to configure MindSpore base environment.

Use add_model.py to build a network with only the Add operator and export the MindSpore inference deployment model.

python add_model.py

Execute the script to generate the tensor_add.mindir file. The input of the model is two one-dimensional tensors with shape [2,2], and the output is the sum of the two input tensors.

Starting Serving Inference

ms_serving --model_path={model directory} --model_name=tensor_add.mindir

If the server prints the MS Serving Listening on 0.0.0.0:5500 log, the Serving has loaded the inference model.

Client Samples

Python Client Sample

Before running the client sample, add the path /{your python path}/lib/python3.7/site-packages/mindspore/ to the environment variable PYTHONPATH.

Obtain ms_client.py and start the Python client.

python ms_client.py

If the following information is displayed, the Serving has correctly executed the inference of the Add network.

ms client received:
[[2. 2.]
 [2. 2.]]

C++ Client Sample

Obtain an executable client sample program.

Download the MindSpore source code. You can use either of the following methods to compile and obtain the client sample program:
- When MindSpore is compiled using the source code, the Serving C++ client sample program is generated. You can find the ms_client executable program in the build/mindspore/serving/example/cpp_client directory.
- Independent compilation
  
  Preinstall gRPC.
  
  Run the following command in the MindSpore source code path to compile a client sample program:
```
cd mindspore/serving/example/cpp_client
mkdir build && cd build
cmake -D GRPC_PATH={grpc_install_dir} ..
make
```
  In the preceding command, {grpc_install_dir} indicates the gRPC installation path. Replace it with the actual gRPC installation path.
Start the client.

Execute ms_client to send an inference request to the Serving.
```
./ms_client --target=localhost:5500
```
If the following information is displayed, the Serving has correctly executed the inference of the Add network.
```
Compute [[1, 2], [3, 4]] + [[1, 2], [3, 4]]
Add result is 2 4 6 8
client received: RPC OK
```

The client code consists of the following parts:

Implement the client based on MSService::Stub and create a client instance.

class MSClient {
 public:
  explicit MSClient(std::shared_ptr<Channel> channel) :  stub_(MSService::NewStub(channel)) {}
 private:
  std::unique_ptr<MSService::Stub> stub_;
};

MSClient client(grpc::CreateChannel(target_str, grpc::InsecureChannelCredentials()));

Build the request input parameter Request, output parameter Reply, and gRPC client Context based on the actual network input.

PredictRequest request;
PredictReply reply;
ClientContext context;

//construct tensor
Tensor data;

//set shape
TensorShape shape;
shape.add_dims(4);
*data.mutable_tensor_shape() = shape;

//set type
data.set_tensor_type(ms_serving::MS_FLOAT32);
std::vector<float> input_data{1, 2, 3, 4};

//set datas
data.set_data(input_data.data(), input_data.size());

//add tensor to request
*request.add_data() = data;
*request.add_data() = data;

Call the gRPC API to communicate with the Serving that has been started, and obtain the return value.
```
Status status = stub_->Predict(&context, request, &reply);
```

For details about the complete code, see ms_client.

REST API Client Sample

Send data in the form of data: data field: flatten each input data of network model into one-dimensional data. Suppose the network model has n inputs, and the final data structure is a two-dimensional list of 1 * n. As in this example, flatten the model input data [[1.0, 2.0], [3.0, 4.0]] and [[1.0, 2.0], [3.0, 4.0]] to form [[1.0, 2.0, 3.0, 4.0], [1.0, 2.0, 3.0, 4.0]].
```
curl -X POST -d '{"data": [[1.0, 2.0, 3.0, 4.0], [1.0, 2.0, 3.0, 4.0]]}' http://127.0.0.1:5501
```
The following return values are displayed, indicating that the serving service has correctly executed the reasoning of the add network, and the output data structure is similar to that of the input:
```
{"data":[[2.0,4.0,6.0,8.0]]}
```
Send data in the form of tensor: tensor field: composed of each input of the network model, keeping the original shape of input. As in this example, the model input data [[1.0, 2.0], [3.0, 4.0]] and [[1.0, 2.0], [3.0, 4.0]] are combined into [[[1.0, 2.0], [3.0, 4.0]], [[1.0, 2.0], [3.0, 4.0]]].
```
curl -X POST -d '{"tensor": [[[1.0, 2.0], [3.0, 4.0]], [[1.0, 2.0], [3.0, 4.0]]]}' http://127.0.0.1:5501
```
The following return values are displayed, indicating that the serving service has correctly executed the reasoning of the add network, and the output data structure is similar to that of the input:
```
{"tensor":[[2.0,4.0], [6.0,8.0]]}
```

REST APICurrently only int32 and fp32 are supported as inputs.