Skip to content

S
Serving

PaddlePaddle / Serving
大约 2 年 前同步成功

通知 187
Star 833
Fork 253
S
Serving
提交 af65b28f 编写于 作者: B BohaoWu
浏览文件
操作

Merge branch 'develop' of https://github.com/PaddlePaddle/Serving into develop

上级 bcc0c749 7ea59bda
展开全部 显示空白变更内容
内联 并排
Showing with 5761 addition and 2841 deletion
CMakeLists.txt
浏览文件 @ af65b28f
......@@ -54,6 +54,7 @@ option(SERVER "Compile Paddle Serving Server" OFF)
option(APP "Compile Paddle Serving App package" OFF)
option(WITH_ELASTIC_CTR "Compile ELASITC-CTR solution" OFF)
option(PACK "Compile for whl" OFF)
option(WITH_TRT "Compile Paddle Serving with TRT" OFF)
set(WITH_MKLML ${WITH_MKL})
if (NOT DEFINED WITH_MKLDNN)
......
README.md
浏览文件 @ af65b28f
......@@ -45,32 +45,26 @@ nvidia-docker exec -it test bash
```
```shell
pip install paddle-serving-client==0.3.2
pip install paddle-serving-server==0.3.2 # CPU
pip install paddle-serving-server-gpu==0.3.2.post9 # GPU with CUDA9.0
pip install paddle-serving-server-gpu==0.3.2.post10 # GPU with CUDA9.0
pip install paddle-serving-client==0.4.0
pip install paddle-serving-server==0.4.0 # CPU
pip install paddle-serving-server-gpu==0.4.0.post9 # GPU with CUDA9.0
pip install paddle-serving-server-gpu==0.4.0.post10 # GPU with CUDA10.0
pip install paddle-serving-server-gpu==0.4.0.trt # GPU with CUDA10.1+TensorRT
```
You may need to use a domestic mirror source (in China, you can use the Tsinghua mirror source, add `-i https://pypi.tuna.tsinghua.edu.cn/simple` to pip command) to speed up the download.
If you need install modules compiled with develop branch, please download packages from [latest packages list](./doc/LATEST_PACKAGES.md) and install with `pip install` command.
Packages of paddle-serving-server and paddle-serving-server-gpu support Centos 6/7 and Ubuntu 16/18.
Packages of paddle-serving-server and paddle-serving-server-gpu support Centos 6/7, Ubuntu 16/18, Windows 10.
Packages of paddle-serving-client and paddle-serving-app support Linux and Windows, but paddle-serving-client only support python2.7/3.6/3.7.
Packages of paddle-serving-client and paddle-serving-app support Linux and Windows, but paddle-serving-client only support python2.7/3.5/3.6/3.7.
Recommended to install paddle >= 1.8.2.
Recommended to install paddle >= 1.8.4.
<h2 align="center"> Pre-built services with Paddle Serving</h2>
For **Windows Users**, please read the document [Paddle Serving for Windows Users](./doc/WINDOWS_TUTORIAL.md)
<h3 align="center">Latest release</h4>
<p align="center">
<a href="https://github.com/PaddlePaddle/Serving/tree/develop/python/examples/ocr">Optical Character Recognition</a>
<br>
<a href="https://github.com/PaddlePaddle/Serving/tree/develop/python/examples/faster_rcnn_model">Object Detection</a>
<br>
<a href="https://github.com/PaddlePaddle/Serving/tree/develop/python/examples/deeplabv3">Image Segmentation</a>
<p>
<h2 align="center"> Pre-built services with Paddle Serving</h2>
<h3 align="center">Chinese Word Segmentation</h4>
......@@ -111,11 +105,11 @@ tar -xzf uci_housing.tar.gz
Paddle Serving provides HTTP and RPC based service for users to access
### HTTP service
### RPC service
Paddle Serving provides a built-in python module called `paddle_serving_server.serve` that can start a RPC service or a http service with one-line command. If we specify the argument `--name uci`, it means that we will have a HTTP service with a url of `$IP:$PORT/uci/prediction`
A user can also start a RPC service with `paddle_serving_server.serve`. RPC service is usually faster than HTTP service, although a user needs to do some coding based on Paddle Serving's python client API. Note that we do not specify `--name` here.
``` shell
python -m paddle_serving_server.serve --model uci_housing_model --thread 10 --port 9292 --name uci
python -m paddle_serving_server.serve --model uci_housing_model --thread 10 --port 9292
```
<center>
......@@ -123,41 +117,63 @@ python -m paddle_serving_server.serve --model uci_housing_model --thread 10 --po
|--------------|------|-----------|--------------------------------|
| `thread` | int | `4` | Concurrency of current service |
| `port` | int | `9292` | Exposed port of current service to users|
| `name` | str | `""` | Service name, can be used to generate HTTP request url |
| `model` | str | `""` | Path of paddle model directory to be served |
| `mem_optim_off` | - | - | Disable memory / graphic memory optimization |
| `ir_optim` | - | - | Enable analysis and optimization of calculation graph |
| `use_mkl` (Only for cpu version) | - | - | Run inference with MKL |
| `use_trt` (Only for trt version) | - | - | Run inference with TensorRT |
Here, we use `curl` to send a HTTP POST request to the service we just started. Users can use any python library to send HTTP POST as well, e.g, [requests](https://requests.readthedocs.io/en/master/).
</center>
``` shell
curl -H "Content-Type:application/json" -X POST -d '{"feed":[{"x": [0.0137, -0.1136, 0.2553, -0.0692, 0.0582, -0.0727, -0.1583, -0.0584, 0.6283, 0.4919, 0.1856, 0.0795, -0.0332]}], "fetch":["price"]}' http://127.0.0.1:9292/uci/prediction
```
### RPC service
A user can also start a RPC service with `paddle_serving_server.serve`. RPC service is usually faster than HTTP service, although a user needs to do some coding based on Paddle Serving's python client API. Note that we do not specify `--name` here.
``` shell
python -m paddle_serving_server.serve --model uci_housing_model --thread 10 --port 9292
```
``` python
```python
# A user can visit rpc service through paddle_serving_client API
from paddle_serving_client import Client
import numpy as np
client = Client()
client.load_client_config("uci_housing_client/serving_client_conf.prototxt")
client.connect(["127.0.0.1:9292"])
data = [0.0137, -0.1136, 0.2553, -0.0692, 0.0582, -0.0727,
-0.1583, -0.0584, 0.6283, 0.4919, 0.1856, 0.0795, -0.0332]
fetch_map = client.predict(feed={"x": data}, fetch=["price"])
fetch_map = client.predict(feed={"x": np.array(data).reshape(1,13,1)}, fetch=["price"])
print(fetch_map)
```
Here, `client.predict` function has two arguments. `feed` is a `python dict` with model input variable alias name and values. `fetch` assigns the prediction variables to be returned from servers. In the example, the name of `"x"` and `"price"` are assigned when the servable model is saved during training.
### WEB service
Users can also put the data format processing logic on the server side, so that they can directly use curl to access the service, refer to the following case whose path is `python/examples/fit_a_line`
```python
from paddle_serving_server.web_service import WebService
import numpy as np
class UciService(WebService):
def preprocess(self, feed=[], fetch=[]):
feed_batch = []
is_batch = True
new_data = np.zeros((len(feed), 1, 13)).astype("float32")
for i, ins in enumerate(feed):
nums = np.array(ins["x"]).reshape(1, 1, 13)
new_data[i] = nums
feed = {"x": new_data}
return feed, fetch, is_batch
uci_service = UciService(name="uci")
uci_service.load_model_config("uci_housing_model")
uci_service.prepare_server(workdir="workdir", port=9292)
uci_service.run_rpc_service()
uci_service.run_web_service()
```
for client side,
```
curl -H "Content-Type:application/json" -X POST -d '{"feed":[{"x": [0.0137, -0.1136, 0.2553, -0.0692, 0.0582, -0.0727, -0.1583, -0.0584, 0.6283, 0.4919, 0.1856, 0.0795, -0.0332]}], "fetch":["price"]}' http://127.0.0.1:9292/uci/prediction
```
the response is
```
{"result":{"price":[[18.901151657104492]]}}
```
<h2 align="center">Some Key Features of Paddle Serving</h2>
- Integrate with Paddle training pipeline seamlessly, most paddle models can be deployed **with one line command**.
......@@ -212,6 +228,10 @@ To connect with other users and contributors, welcome to join our [Slack channel
If you want to contribute code to Paddle Serving, please reference [Contribution Guidelines](doc/CONTRIBUTE.md)
- Special Thanks to [@BeyondYourself](https://github.com/BeyondYourself) in complementing the gRPC tutorial, updating the FAQ doc and modifying the mdkir command
- Special Thanks to [@mcl-stone](https://github.com/mcl-stone) in updating faster_rcnn benchmark
- Special Thanks to [@cg82616424](https://github.com/cg82616424) in updating the unet benchmark and modifying resize comment error
### Feedback
For any feedback or to report a bug, please propose a [GitHub Issue](https://github.com/PaddlePaddle/Serving/issues).
......
README_CN.md
浏览文件 @ af65b28f
......@@ -47,21 +47,24 @@ nvidia-docker run -p 9292:9292 --name test -dit hub.baidubce.com/paddlepaddle/se
nvidia-docker exec -it test bash
```
```shell
pip install paddle-serving-client==0.3.2
pip install paddle-serving-server==0.3.2 # CPU
pip install paddle-serving-server-gpu==0.3.2.post9 # GPU with CUDA9.0
pip install paddle-serving-server-gpu==0.3.2.post10 # GPU with CUDA10.0
pip install paddle-serving-client==0.4.0
pip install paddle-serving-server==0.4.0 # CPU
pip install paddle-serving-server-gpu==0.4.0.post9 # GPU with CUDA9.0
pip install paddle-serving-server-gpu==0.4.0.post10 # GPU with CUDA10.0
pip install paddle-serving-server-gpu==0.4.0.trt # GPU with CUDA10.1+TensorRT
```
您可能需要使用国内镜像源（例如清华源, 在pip命令中添加`-i https://pypi.tuna.tsinghua.edu.cn/simple`）来加速下载。
如果需要使用develop分支编译的安装包，请从[最新安装包列表](./doc/LATEST_PACKAGES.md)中获取下载地址进行下载，使用`pip install`命令进行安装。
paddle-serving-server和paddle-serving-server-gpu安装包支持Centos 6/7和Ubuntu 16/18。
paddle-serving-server和paddle-serving-server-gpu安装包支持Centos 6/7, Ubuntu 16/18和Windows 10。
paddle-serving-client和paddle-serving-app安装包支持Linux和Windows，其中paddle-serving-client仅支持python2.7/3.5/3.6。
推荐安装1.8.2及以上版本的paddle
推荐安装1.8.4及以上版本的paddle
对于**Windows 10 用户**，请参考文档[Windows平台使用Paddle Serving指导](./doc/WINDOWS_TUTORIAL_CN.md)。
<h2 align="center"> Paddle Serving预装的服务 </h2>
......@@ -105,13 +108,12 @@ tar -xzf uci_housing.tar.gz
Paddle Serving 为用户提供了基于 HTTP 和 RPC 的服务
<h3 align="center">RPC服务</h3>
<h3 align="center">HTTP服务</h3>
Paddle Serving提供了一个名为`paddle_serving_server.serve`的内置python模块，可以使用单行命令启动RPC服务或HTTP服务。如果我们指定参数`--name uci`，则意味着我们将拥有一个HTTP服务，其URL为$IP:$PORT/uci/prediction`。
用户还可以使用`paddle_serving_server.serve`启动RPC服务。 尽管用户需要基于Paddle Serving的python客户端API进行一些开发，但是RPC服务通常比HTTP服务更快。需要指出的是这里我们没有指定`--name`。
``` shell
python -m paddle_serving_server.serve --model uci_housing_model --thread 10 --port 9292 --name uci
python -m paddle_serving_server.serve --model uci_housing_model --thread 10 --port 9292
```
<center>
......@@ -124,22 +126,12 @@ python -m paddle_serving_server.serve --model uci_housing_model --thread 10 --po
| `mem_optim_off` | - | - | Disable memory optimization |
| `ir_optim` | - | - | Enable analysis and optimization of calculation graph |
| `use_mkl` (Only for cpu version) | - | - | Run inference with MKL |
| `use_trt` (Only for trt version) | - | - | Run inference with TensorRT |
我们使用 `curl` 命令来发送HTTP POST请求给刚刚启动的服务。用户也可以调用python库来发送HTTP POST请求，请参考英文文档 [requests](https://requests.readthedocs.io/en/master/)。
我们使用 `curl` 命令来发送HTTP POST请求给刚刚启动的服务。用户也可以调用python库来发送HTTP POST请求，请参考英文文
档 [requests](https://requests.readthedocs.io/en/master/)。
</center>
``` shell
curl -H "Content-Type:application/json" -X POST -d '{"feed":[{"x": [0.0137, -0.1136, 0.2553, -0.0692, 0.0582, -0.0727, -0.1583, -0.0584, 0.6283, 0.4919, 0.1856, 0.0795, -0.0332]}], "fetch":["price"]}' http://127.0.0.1:9292/uci/prediction
```
<h3 align="center">RPC服务</h3>
用户还可以使用`paddle_serving_server.serve`启动RPC服务。 尽管用户需要基于Paddle Serving的python客户端API进行一些开发，但是RPC服务通常比HTTP服务更快。需要指出的是这里我们没有指定`--name`。
``` shell
python -m paddle_serving_server.serve --model uci_housing_model --thread 10 --port 9292
```
``` python
# A user can visit rpc service through paddle_serving_client API
from paddle_serving_client import Client
......@@ -149,12 +141,45 @@ client.load_client_config("uci_housing_client/serving_client_conf.prototxt")
client.connect(["127.0.0.1:9292"])
data = [0.0137, -0.1136, 0.2553, -0.0692, 0.0582, -0.0727,
-0.1583, -0.0584, 0.6283, 0.4919, 0.1856, 0.0795, -0.0332]
fetch_map = client.predict(feed={"x": data}, fetch=["price"])
fetch_map = client.predict(feed={"x": np.array(data).reshape(1,13,1)}, fetch=["price"])
print(fetch_map)
```
在这里，`client.predict`函数具有两个参数。 `feed`是带有模型输入变量别名和值的`python dict`。 `fetch`被要从服务器返回的预测变量赋值。 在该示例中，在训练过程中保存可服务模型时，被赋值的tensor名为`"x"`和`"price"`。
<h3 align="center">HTTP服务</h3>
用户也可以将数据格式处理逻辑放在服务器端进行，这样就可以直接用curl去访问服务，参考如下案例，在目录`python/examples/fit_a_line`
```python
from paddle_serving_server.web_service import WebService
import numpy as np
class UciService(WebService):
def preprocess(self, feed=[], fetch=[]):
feed_batch = []
is_batch = True
new_data = np.zeros((len(feed), 1, 13)).astype("float32")
for i, ins in enumerate(feed):
nums = np.array(ins["x"]).reshape(1, 1, 13)
new_data[i] = nums
feed = {"x": new_data}
return feed, fetch, is_batch
uci_service = UciService(name="uci")
uci_service.load_model_config("uci_housing_model")
uci_service.prepare_server(workdir="workdir", port=9292)
uci_service.run_rpc_service()
uci_service.run_web_service()
```
客户端输入
```
curl -H "Content-Type:application/json" -X POST -d '{"feed":[{"x": [0.0137, -0.1136, 0.2553, -0.0692, 0.0582, -0.0727, -0.1583, -0.0584, 0.6283, 0.4919, 0.1856, 0.0795, -0.0332]}], "fetch":["price"]}' http://127.0.0.1:9292/uci/prediction
```
返回结果
```
{"result":{"price":[[18.901151657104492]]}}
```
<h2 align="center">Paddle Serving的核心功能</h2>
- 与Paddle训练紧密连接，绝大部分Paddle模型可以 **一键部署**.
......@@ -207,6 +232,10 @@ print(fetch_map)
如果您想为Paddle Serving贡献代码，请参考 [Contribution Guidelines](doc/CONTRIBUTE.md)
- 特别感谢 [@BeyondYourself](https://github.com/BeyondYourself) 提供grpc教程，更新FAQ教程，整理文件目录。
- 特别感谢 [@mcl-stone](https://github.com/mcl-stone) 提供faster rcnn benchmark脚本
- 特别感谢 [@cg82616424](https://github.com/cg82616424) 提供unet benchmark脚本和修改部分注释错误
### 反馈
如有任何反馈或是bug，请在 [GitHub Issue](https://github.com/PaddlePaddle/Serving/issues)提交
......
cmake/paddlepaddle.cmake
浏览文件 @ af65b28f
......@@ -34,7 +34,11 @@ message( "WITH_GPU = ${WITH_GPU}")
SET(PADDLE_VERSION "1.8.4")
if (WITH_GPU)
SET(PADDLE_LIB_VERSION "${PADDLE_VERSION}-gpu-cuda${CUDA_VERSION_MAJOR}-cudnn7-avx-mkl")
if (WITH_TRT)
SET(PADDLE_LIB_VERSION "${PADDLE_VERSION}-gpu-cuda10.1-cudnn7.6-avx-mkl-trt6")
else()
SET(PADDLE_LIB_VERSION "${PADDLE_VERSION}-gpu-cuda10-cudnn7-avx-mkl")
endif()
else()
if (WITH_AVX)
if (WITH_MKLML)
......@@ -50,7 +54,23 @@ endif()
SET(PADDLE_LIB_PATH "http://paddle-inference-lib.bj.bcebos.com/${PADDLE_LIB_VERSION}/fluid_inference.tgz")
MESSAGE(STATUS "PADDLE_LIB_PATH=${PADDLE_LIB_PATH}")
if (WITH_GPU OR WITH_MKLML)
ExternalProject_Add(
if (WITH_TRT)
ExternalProject_Add(
"extern_paddle"
${EXTERNAL_PROJECT_LOG_ARGS}
URL "${PADDLE_LIB_PATH}"
PREFIX "${PADDLE_SOURCES_DIR}"
DOWNLOAD_DIR "${PADDLE_DOWNLOAD_DIR}"
CONFIGURE_COMMAND ""
BUILD_COMMAND ""
UPDATE_COMMAND ""
INSTALL_COMMAND
${CMAKE_COMMAND} -E copy_directory ${PADDLE_DOWNLOAD_DIR}/paddle/include ${PADDLE_INSTALL_DIR}/include &&
${CMAKE_COMMAND} -E copy_directory ${PADDLE_DOWNLOAD_DIR}/paddle/lib ${PADDLE_INSTALL_DIR}/lib &&
${CMAKE_COMMAND} -E copy_directory ${PADDLE_DOWNLOAD_DIR}/third_party ${PADDLE_INSTALL_DIR}/third_party
)
else()
ExternalProject_Add(
"extern_paddle"
${EXTERNAL_PROJECT_LOG_ARGS}
URL "${PADDLE_LIB_PATH}"
......@@ -64,7 +84,8 @@ ExternalProject_Add(
${CMAKE_COMMAND} -E copy_directory ${PADDLE_DOWNLOAD_DIR}/paddle/lib ${PADDLE_INSTALL_DIR}/lib &&
${CMAKE_COMMAND} -E copy_directory ${PADDLE_DOWNLOAD_DIR}/third_party ${PADDLE_INSTALL_DIR}/third_party &&
${CMAKE_COMMAND} -E copy ${PADDLE_INSTALL_DIR}/third_party/install/mkldnn/lib/libmkldnn.so.0 ${PADDLE_INSTALL_DIR}/third_party/install/mkldnn/lib/libmkldnn.so
)
)
endif()
else()
ExternalProject_Add(
"extern_paddle"
......@@ -92,8 +113,16 @@ LINK_DIRECTORIES(${PADDLE_INSTALL_DIR}/third_party/install/mkldnn/lib)
ADD_LIBRARY(openblas STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET openblas PROPERTY IMPORTED_LOCATION ${PADDLE_INSTALL_DIR}/third_party/install/openblas/lib/libopenblas.a)
ADD_LIBRARY(paddle_fluid STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET paddle_fluid PROPERTY IMPORTED_LOCATION ${PADDLE_INSTALL_DIR}/lib/libpaddle_fluid.a)
ADD_LIBRARY(paddle_fluid SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET paddle_fluid PROPERTY IMPORTED_LOCATION ${PADDLE_INSTALL_DIR}/lib/libpaddle_fluid.so)
if (WITH_TRT)
ADD_LIBRARY(nvinfer SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET nvinfer PROPERTY IMPORTED_LOCATION ${TENSORRT_ROOT}/lib/libnvinfer.so)
ADD_LIBRARY(nvinfer_plugin SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET nvinfer_plugin PROPERTY IMPORTED_LOCATION ${TENSORRT_ROOT}/lib/libnvinfer_plugin.so)
endif()
ADD_LIBRARY(xxhash STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET xxhash PROPERTY IMPORTED_LOCATION ${PADDLE_INSTALL_DIR}/third_party/install/xxhash/lib/libxxhash.a)
......@@ -102,3 +131,8 @@ LIST(APPEND external_project_dependencies paddle)
LIST(APPEND paddle_depend_libs
xxhash)
if(WITH_TRT)
LIST(APPEND paddle_depend_libs
nvinfer nvinfer_plugin)
endif()
core/configure/proto/server_configure.proto
浏览文件 @ af65b28f
......@@ -44,6 +44,7 @@ message EngineDesc {
optional bool static_optimization = 14;
optional bool force_update_static_cache = 15;
optional bool enable_ir_optimization = 16;
optional bool use_trt = 17;
};
// model_toolkit conf
......
core/cube/CMakeLists.txt
浏览文件 @ af65b28f
......@@ -12,8 +12,9 @@
# See the License for the specific language governing permissions and
# limitations under the License
#execute_process(COMMAND go env -w GO111MODULE=off)
add_subdirectory(cube-server)
add_subdirectory(cube-api)
add_subdirectory(cube-builder)
add_subdirectory(cube-transfer)
add_subdirectory(cube-agent)
#add_subdirectory(cube-transfer)
#add_subdirectory(cube-agent)
core/general-client/include/general_model.h
浏览文件 @ af65b28f
......@@ -218,25 +218,15 @@ class PredictorClient {
int destroy_predictor();
int batch_predict(
const std::vector<std::vector<std::vector<float>>>& float_feed_batch,
const std::vector<std::string>& float_feed_name,
const std::vector<std::vector<int>>& float_shape,
const std::vector<std::vector<std::vector<int64_t>>>& int_feed_batch,
const std::vector<std::string>& int_feed_name,
const std::vector<std::vector<int>>& int_shape,
const std::vector<std::string>& fetch_name,
PredictorRes& predict_res_batch, // NOLINT
const int& pid,
const uint64_t log_id);
int numpy_predict(
const std::vector<std::vector<py::array_t<float>>>& float_feed_batch,
const std::vector<std::string>& float_feed_name,
const std::vector<std::vector<int>>& float_shape,
const std::vector<std::vector<int>>& float_lod_slot_batch,
const std::vector<std::vector<py::array_t<int64_t>>>& int_feed_batch,
const std::vector<std::string>& int_feed_name,
const std::vector<std::vector<int>>& int_shape,
const std::vector<std::vector<int>>& int_lod_slot_batch,
const std::vector<std::string>& fetch_name,
PredictorRes& predict_res_batch, // NOLINT
const int& pid,
......
core/general-client/src/general_model.cpp
浏览文件 @ af65b28f
......@@ -137,227 +137,15 @@ int PredictorClient::create_predictor() {
return 0;
}
int PredictorClient::batch_predict(
const std::vector<std::vector<std::vector<float>>> &float_feed_batch,
const std::vector<std::string> &float_feed_name,
const std::vector<std::vector<int>> &float_shape,
const std::vector<std::vector<std::vector<int64_t>>> &int_feed_batch,
const std::vector<std::string> &int_feed_name,
const std::vector<std::vector<int>> &int_shape,
const std::vector<std::string> &fetch_name,
PredictorRes &predict_res_batch,
const int &pid,
const uint64_t log_id) {
int batch_size = std::max(float_feed_batch.size(), int_feed_batch.size());
predict_res_batch.clear();
Timer timeline;
int64_t preprocess_start = timeline.TimeStampUS();
int fetch_name_num = fetch_name.size();
_api.thrd_initialize();
std::string variant_tag;
_predictor = _api.fetch_predictor("general_model", &variant_tag);
predict_res_batch.set_variant_tag(variant_tag);
VLOG(2) << "fetch general model predictor done.";
VLOG(2) << "float feed name size: " << float_feed_name.size();
VLOG(2) << "int feed name size: " << int_feed_name.size();
VLOG(2) << "max body size : " << brpc::fLU64::FLAGS_max_body_size;
Request req;
req.set_log_id(log_id);
for (auto &name : fetch_name) {
req.add_fetch_var_names(name);
}
for (int bi = 0; bi < batch_size; bi++) {
VLOG(2) << "prepare batch " << bi;
std::vector<Tensor *> tensor_vec;
FeedInst *inst = req.add_insts();
std::vector<std::vector<float>> float_feed = float_feed_batch[bi];
std::vector<std::vector<int64_t>> int_feed = int_feed_batch[bi];
for (auto &name : float_feed_name) {
tensor_vec.push_back(inst->add_tensor_array());
}
for (auto &name : int_feed_name) {
tensor_vec.push_back(inst->add_tensor_array());
}
VLOG(2) << "batch [" << bi << "] int_feed_name and float_feed_name "
<< "prepared";
int vec_idx = 0;
VLOG(2) << "tensor_vec size " << tensor_vec.size() << " float shape "
<< float_shape.size();
for (auto &name : float_feed_name) {
int idx = _feed_name_to_idx[name];
Tensor *tensor = tensor_vec[idx];
VLOG(2) << "prepare float feed " << name << " shape size "
<< float_shape[vec_idx].size();
for (uint32_t j = 0; j < float_shape[vec_idx].size(); ++j) {
tensor->add_shape(float_shape[vec_idx][j]);
}
tensor->set_elem_type(1);
for (uint32_t j = 0; j < float_feed[vec_idx].size(); ++j) {
tensor->add_float_data(float_feed[vec_idx][j]);
}
vec_idx++;
}
VLOG(2) << "batch [" << bi << "] "
<< "float feed value prepared";
vec_idx = 0;
for (auto &name : int_feed_name) {
int idx = _feed_name_to_idx[name];
Tensor *tensor = tensor_vec[idx];
if (_type[idx] == 0) {
VLOG(2) << "prepare int64 feed " << name << " shape size "
<< int_shape[vec_idx].size();
VLOG(3) << "feed var name " << name << " index " << vec_idx
<< "first data " << int_feed[vec_idx][0];
for (uint32_t j = 0; j < int_feed[vec_idx].size(); ++j) {
tensor->add_int64_data(int_feed[vec_idx][j]);
}
} else if (_type[idx] == 2) {
VLOG(2) << "prepare int32 feed " << name << " shape size "
<< int_shape[vec_idx].size();
VLOG(3) << "feed var name " << name << " index " << vec_idx
<< "first data " << int32_t(int_feed[vec_idx][0]);
for (uint32_t j = 0; j < int_feed[vec_idx].size(); ++j) {
tensor->add_int_data(int32_t(int_feed[vec_idx][j]));
}
}
for (uint32_t j = 0; j < int_shape[vec_idx].size(); ++j) {
tensor->add_shape(int_shape[vec_idx][j]);
}
tensor->set_elem_type(_type[idx]);
vec_idx++;
}
VLOG(2) << "batch [" << bi << "] "
<< "int feed value prepared";
}
int64_t preprocess_end = timeline.TimeStampUS();
int64_t client_infer_start = timeline.TimeStampUS();
Response res;
int64_t client_infer_end = 0;
int64_t postprocess_start = 0;
int64_t postprocess_end = 0;
if (FLAGS_profile_client) {
if (FLAGS_profile_server) {
req.set_profile_server(true);
}
}
res.Clear();
if (_predictor->inference(&req, &res) != 0) {
LOG(ERROR) << "failed call predictor with req: " << req.ShortDebugString();
_api.thrd_clear();
return -1;
} else {
client_infer_end = timeline.TimeStampUS();
postprocess_start = client_infer_end;
VLOG(2) << "get model output num";
uint32_t model_num = res.outputs_size();
VLOG(2) << "model num: " << model_num;
for (uint32_t m_idx = 0; m_idx < model_num; ++m_idx) {
VLOG(2) << "process model output index: " << m_idx;
auto output = res.outputs(m_idx);
ModelRes model;
model.set_engine_name(output.engine_name());
int idx = 0;
for (auto &name : fetch_name) {
// int idx = _fetch_name_to_idx[name];
int shape_size = output.insts(0).tensor_array(idx).shape_size();
VLOG(2) << "fetch var " << name << " index " << idx << " shape size "
<< shape_size;
model._shape_map[name].resize(shape_size);
for (int i = 0; i < shape_size; ++i) {
model._shape_map[name][i] =
output.insts(0).tensor_array(idx).shape(i);
}
int lod_size = output.insts(0).tensor_array(idx).lod_size();
if (lod_size > 0) {
model._lod_map[name].resize(lod_size);
for (int i = 0; i < lod_size; ++i) {
model._lod_map[name][i] = output.insts(0).tensor_array(idx).lod(i);
}
}
idx += 1;
}
idx = 0;
for (auto &name : fetch_name) {
// int idx = _fetch_name_to_idx[name];
if (_fetch_name_to_type[name] == 0) {
VLOG(2) << "ferch var " << name << "type int64";
int size = output.insts(0).tensor_array(idx).int64_data_size();
model._int64_value_map[name] = std::vector<int64_t>(
output.insts(0).tensor_array(idx).int64_data().begin(),
output.insts(0).tensor_array(idx).int64_data().begin() + size);
} else if (_fetch_name_to_type[name] == 1) {
VLOG(2) << "fetch var " << name << "type float";
int size = output.insts(0).tensor_array(idx).float_data_size();
model._float_value_map[name] = std::vector<float>(
output.insts(0).tensor_array(idx).float_data().begin(),
output.insts(0).tensor_array(idx).float_data().begin() + size);
} else if (_fetch_name_to_type[name] == 2) {
VLOG(2) << "fetch var " << name << "type int32";
int size = output.insts(0).tensor_array(idx).int_data_size();
model._int32_value_map[name] = std::vector<int32_t>(
output.insts(0).tensor_array(idx).int_data().begin(),
output.insts(0).tensor_array(idx).int_data().begin() + size);
}
idx += 1;
}
predict_res_batch.add_model_res(std::move(model));
}
postprocess_end = timeline.TimeStampUS();
}
if (FLAGS_profile_client) {
std::ostringstream oss;
oss << "PROFILE\t"
<< "pid:" << pid << "\t"
<< "prepro_0:" << preprocess_start << " "
<< "prepro_1:" << preprocess_end << " "
<< "client_infer_0:" << client_infer_start << " "
<< "client_infer_1:" << client_infer_end << " ";
if (FLAGS_profile_server) {
int op_num = res.profile_time_size() / 2;
for (int i = 0; i < op_num; ++i) {
oss << "op" << i << "_0:" << res.profile_time(i * 2) << " ";
oss << "op" << i << "_1:" << res.profile_time(i * 2 + 1) << " ";
}
}
oss << "postpro_0:" << postprocess_start << " ";
oss << "postpro_1:" << postprocess_end;
fprintf(stderr, "%s\n", oss.str().c_str());
}
_api.thrd_clear();
return 0;
}
int PredictorClient::numpy_predict(
const std::vector<std::vector<py::array_t<float>>> &float_feed_batch,
const std::vector<std::string> &float_feed_name,
const std::vector<std::vector<int>> &float_shape,
const std::vector<std::vector<int>> &float_lod_slot_batch,
const std::vector<std::vector<py::array_t<int64_t>>> &int_feed_batch,
const std::vector<std::string> &int_feed_name,
const std::vector<std::vector<int>> &int_shape,
const std::vector<std::vector<int>> &int_lod_slot_batch,
const std::vector<std::string> &fetch_name,
PredictorRes &predict_res_batch,
const int &pid,
......@@ -412,6 +200,9 @@ int PredictorClient::numpy_predict(
for (uint32_t j = 0; j < float_shape[vec_idx].size(); ++j) {
tensor->add_shape(float_shape[vec_idx][j]);
}
for (uint32_t j = 0; j < float_lod_slot_batch[vec_idx].size(); ++j) {
tensor->add_lod(float_lod_slot_batch[vec_idx][j]);
}
tensor->set_elem_type(1);
const int float_shape_size = float_shape[vec_idx].size();
switch (float_shape_size) {
......@@ -470,6 +261,9 @@ int PredictorClient::numpy_predict(
for (uint32_t j = 0; j < int_shape[vec_idx].size(); ++j) {
tensor->add_shape(int_shape[vec_idx][j]);
}
for (uint32_t j = 0; j < int_lod_slot_batch[vec_idx].size(); ++j) {
tensor->add_lod(int_lod_slot_batch[vec_idx][j]);
}
tensor->set_elem_type(_type[idx]);
if (_type[idx] == 0) {
......
core/general-client/src/pybind_general_model.cpp
浏览文件 @ af65b28f
......@@ -95,42 +95,18 @@ PYBIND11_MODULE(serving_client, m) {
[](PredictorClient &self) { self.create_predictor(); })
.def("destroy_predictor",
[](PredictorClient &self) { self.destroy_predictor(); })
.def("batch_predict",
[](PredictorClient &self,
const std::vector<std::vector<std::vector<float>>>
&float_feed_batch,
const std::vector<std::string> &float_feed_name,
const std::vector<std::vector<int>> &float_shape,
const std::vector<std::vector<std::vector<int64_t>>>
&int_feed_batch,
const std::vector<std::string> &int_feed_name,
const std::vector<std::vector<int>> &int_shape,
const std::vector<std::string> &fetch_name,
PredictorRes &predict_res_batch,
const int &pid,
const uint64_t log_id) {
return self.batch_predict(float_feed_batch,
float_feed_name,
float_shape,
int_feed_batch,
int_feed_name,
int_shape,
fetch_name,
predict_res_batch,
pid,
log_id);
},
py::call_guard<py::gil_scoped_release>())
.def("numpy_predict",
[](PredictorClient &self,
const std::vector<std::vector<py::array_t<float>>>
&float_feed_batch,
const std::vector<std::string> &float_feed_name,
const std::vector<std::vector<int>> &float_shape,
const std::vector<std::vector<int>> &float_lod_slot_batch,
const std::vector<std::vector<py::array_t<int64_t>>>
&int_feed_batch,
const std::vector<std::string> &int_feed_name,
const std::vector<std::vector<int>> &int_shape,
const std::vector<std::vector<int>> &int_lod_slot_batch,
const std::vector<std::string> &fetch_name,
PredictorRes &predict_res_batch,
const int &pid,
......@@ -138,9 +114,11 @@ PYBIND11_MODULE(serving_client, m) {
return self.numpy_predict(float_feed_batch,
float_feed_name,
float_shape,
float_lod_slot_batch,
int_feed_batch,
int_feed_name,
int_shape,
int_lod_slot_batch,
fetch_name,
predict_res_batch,
pid,
......
core/general-server/CMakeLists.txt
浏览文件 @ af65b28f
......@@ -9,7 +9,7 @@ endif()
target_include_directories(serving PUBLIC
${CMAKE_CURRENT_BINARY_DIR}/../../core/predictor
)
include_directories(${CUDNN_ROOT}/include/)
if(WITH_GPU)
target_link_libraries(serving -Wl,--whole-archive fluid_gpu_engine
-Wl,--no-whole-archive)
......@@ -29,7 +29,11 @@ if(WITH_GPU)
endif()
if(WITH_MKL OR WITH_GPU)
if (WITH_TRT)
target_link_libraries(serving -liomp5 -lmklml_intel -lpthread -lcrypto -lm -lrt -lssl -ldl -lz -lbz2)
else()
target_link_libraries(serving -liomp5 -lmklml_intel -lmkldnn -lpthread -lcrypto -lm -lrt -lssl -ldl -lz -lbz2)
endif()
else()
target_link_libraries(serving openblas -lpthread -lcrypto -lm -lrt -lssl -ldl -lz -lbz2)
endif()
......
core/general-server/op/general_reader_op.cpp
浏览文件 @ af65b28f
......@@ -73,8 +73,6 @@ int GeneralReaderOp::inference() {
// reade request from client
const Request *req = dynamic_cast<const Request *>(get_request_message());
uint64_t log_id = req->log_id();
int batch_size = req->insts_size();
int input_var_num = 0;
std::vector<int64_t> elem_type;
std::vector<int64_t> elem_size;
......@@ -83,7 +81,6 @@ int GeneralReaderOp::inference() {
GeneralBlob *res = mutable_data<GeneralBlob>();
TensorVector *out = &res->tensor_vector;
res->SetBatchSize(batch_size);
res->SetLogId(log_id);
if (!res) {
......@@ -98,11 +95,11 @@ int GeneralReaderOp::inference() {
VLOG(2) << "(logid=" << log_id
<< ") start to call load general model_conf op";
baidu::paddle_serving::predictor::Resource &resource =
baidu::paddle_serving::predictor::Resource::instance();
VLOG(2) << "(logid=" << log_id << ") get resource pointer done.";
std::shared_ptr<PaddleGeneralModelConfig> model_config =
resource.get_general_model_config();
......@@ -122,13 +119,11 @@ int GeneralReaderOp::inference() {
elem_type.resize(var_num);
elem_size.resize(var_num);
capacity.resize(var_num);
// prepare basic information for input
for (int i = 0; i < var_num; ++i) {
paddle::PaddleTensor lod_tensor;
elem_type[i] = req->insts(0).tensor_array(i).elem_type();
VLOG(2) << "(logid=" << log_id << ") var[" << i
<< "] has elem type: " << elem_type[i];
VLOG(2) << "var[" << i << "] has elem type: " << elem_type[i];
if (elem_type[i] == 0) { // int64
elem_size[i] = sizeof(int64_t);
lod_tensor.dtype = paddle::PaddleDType::INT64;
......@@ -139,13 +134,24 @@ int GeneralReaderOp::inference() {
elem_size[i] = sizeof(int32_t);
lod_tensor.dtype = paddle::PaddleDType::INT32;
}
if (model_config->_is_lod_feed[i]) {
lod_tensor.lod.resize(1);
lod_tensor.lod[0].push_back(0);
// implement lod tensor here
if (req->insts(0).tensor_array(i).lod_size() > 0) {
VLOG(2) << "(logid=" << log_id << ") var[" << i << "] is lod_tensor";
lod_tensor.lod.resize(1);
for (int k = 0; k < req->insts(0).tensor_array(i).lod_size(); ++k) {
lod_tensor.lod[0].push_back(req->insts(0).tensor_array(i).lod(k));
}
capacity[i] = 1;
for (int k = 0; k < req->insts(0).tensor_array(i).shape_size(); ++k) {
int dim = req->insts(0).tensor_array(i).shape(k);
VLOG(2) << "(logid=" << log_id << ") shape for var[" << i
<< "]: " << dim;
capacity[i] *= dim;
lod_tensor.shape.push_back(dim);
}
VLOG(2) << "(logid=" << log_id << ") var[" << i
<< "] is tensor, capacity: " << capacity[i];
} else {
lod_tensor.shape.push_back(batch_size);
capacity[i] = 1;
for (int k = 0; k < req->insts(0).tensor_array(i).shape_size(); ++k) {
int dim = req->insts(0).tensor_array(i).shape(k);
......@@ -160,13 +166,11 @@ int GeneralReaderOp::inference() {
lod_tensor.name = model_config->_feed_name[i];
out->push_back(lod_tensor);
}
// specify the memory needed for output tensor_vector
for (int i = 0; i < var_num; ++i) {
if (out->at(i).lod.size() == 1) {
int tensor_size = 0;
for (int j = 0; j < batch_size; ++j) {
const Tensor &tensor = req->insts(j).tensor_array(i);
const Tensor &tensor = req->insts(0).tensor_array(i);
int data_len = 0;
if (tensor.int64_data_size() > 0) {
data_len = tensor.int64_data_size();
......@@ -188,23 +192,14 @@ int GeneralReaderOp::inference() {
} else {
sample_len = tensor.shape(0);
}
out->at(i).lod[0].push_back(cur_len + sample_len);
VLOG(2) << "(logid=" << log_id << ") new len: " << cur_len + sample_len;
}
out->at(i).data.Resize(tensor_size * elem_size[i]);
out->at(i).shape = {out->at(i).lod[0].back()};
for (int j = 1; j < req->insts(0).tensor_array(i).shape_size(); ++j) {
out->at(i).shape.push_back(req->insts(0).tensor_array(i).shape(j));
}
if (out->at(i).shape.size() == 1) {
out->at(i).shape.push_back(1);
}
VLOG(2) << "(logid=" << log_id << ") var[" << i
<< "] is lod_tensor and len=" << out->at(i).lod[0].back();
} else {
out->at(i).data.Resize(batch_size * capacity[i] * elem_size[i]);
out->at(i).data.Resize(capacity[i] * elem_size[i]);
VLOG(2) << "(logid=" << log_id << ") var[" << i
<< "] is tensor and capacity=" << batch_size * capacity[i];
<< "] is tensor and capacity=" << capacity[i];
}
}
......@@ -215,58 +210,36 @@ int GeneralReaderOp::inference() {
VLOG(2) << "(logid=" << log_id << ") first element data in var[" << i
<< "] is " << req->insts(0).tensor_array(i).int64_data(0);
int offset = 0;
for (int j = 0; j < batch_size; ++j) {
int elem_num = req->insts(j).tensor_array(i).int64_data_size();
int elem_num = req->insts(0).tensor_array(i).int64_data_size();
for (int k = 0; k < elem_num; ++k) {
dst_ptr[offset + k] = req->insts(j).tensor_array(i).int64_data(k);
}
if (out->at(i).lod.size() == 1) {
offset = out->at(i).lod[0][j + 1];
} else {
offset += capacity[i];
}
dst_ptr[offset + k] = req->insts(0).tensor_array(i).int64_data(k);
}
} else if (elem_type[i] == 1) {
float *dst_ptr = static_cast<float *>(out->at(i).data.data());
VLOG(2) << "(logid=" << log_id << ") first element data in var[" << i
<< "] is " << req->insts(0).tensor_array(i).float_data(0);
int offset = 0;
for (int j = 0; j < batch_size; ++j) {
int elem_num = req->insts(j).tensor_array(i).float_data_size();
int elem_num = req->insts(0).tensor_array(i).float_data_size();
for (int k = 0; k < elem_num; ++k) {
dst_ptr[offset + k] = req->insts(j).tensor_array(i).float_data(k);
}
if (out->at(i).lod.size() == 1) {
offset = out->at(i).lod[0][j + 1];
} else {
offset += capacity[i];
}
dst_ptr[offset + k] = req->insts(0).tensor_array(i).float_data(k);
}
} else if (elem_type[i] == 2) {
int32_t *dst_ptr = static_cast<int32_t *>(out->at(i).data.data());
VLOG(2) << "(logid=" << log_id << ") first element data in var[" << i
<< "] is " << req->insts(0).tensor_array(i).int_data(0);
int offset = 0;
for (int j = 0; j < batch_size; ++j) {
int elem_num = req->insts(j).tensor_array(i).int_data_size();
int elem_num = req->insts(0).tensor_array(i).int_data_size();
for (int k = 0; k < elem_num; ++k) {
dst_ptr[offset + k] = req->insts(j).tensor_array(i).int_data(k);
}
if (out->at(i).lod.size() == 1) {
offset = out->at(i).lod[0][j + 1];
} else {
offset += capacity[i];
}
dst_ptr[offset + k] = req->insts(0).tensor_array(i).int_data(k);
}
}
}
VLOG(2) << "(logid=" << log_id << ") output size: " << out->size();
timeline.Pause();
int64_t end = timeline.TimeStampUS();
res->p_size = 0;
res->_batch_size = batch_size;
res->_batch_size = 1;
AddBlobInfo(res, start);
AddBlobInfo(res, end);
......
core/general-server/op/general_response_op.cpp
浏览文件 @ af65b28f
......@@ -155,11 +155,13 @@ int GeneralResponseOp::inference() {
}
if (model_config->_is_lod_fetch[idx]) {
if (in->at(idx).lod.size() > 0) {
for (int j = 0; j < in->at(idx).lod[0].size(); ++j) {
fetch_p->mutable_tensor_array(var_idx)->add_lod(
in->at(idx).lod[0][j]);
}
}
}
VLOG(2) << "(logid=" << log_id << ") fetch var ["
<< model_config->_fetch_name[idx] << "] ready";
......
core/predictor/CMakeLists.txt
浏览文件 @ af65b28f
......@@ -13,7 +13,9 @@ set_source_files_properties(
PROPERTIES
COMPILE_FLAGS "-Wno-strict-aliasing -Wno-unused-variable -Wno-non-virtual-dtor -Wno-error=non-virtual-dtor -Wno-error=delete-non-virtual-dtor")
add_dependencies(pdserving protobuf boost brpc leveldb pdcodegen configure)
if (WITH_TRT)
add_definitions(-DWITH_TRT)
endif()
target_link_libraries(pdserving
brpc protobuf boost leveldb configure -lpthread -lcrypto -lm -lrt -lssl -ldl -lz)
......
core/predictor/framework/infer.h
浏览文件 @ af65b28f
......@@ -38,6 +38,7 @@ class InferEngineCreationParams {
_enable_ir_optimization = false;
_static_optimization = false;
_force_update_static_cache = false;
_use_trt = false;
}
void set_path(const std::string& path) { _path = path; }
......@@ -50,12 +51,16 @@ class InferEngineCreationParams {
_enable_ir_optimization = enable_ir_optimization;
}
void set_use_trt(bool use_trt) { _use_trt = use_trt; }
bool enable_memory_optimization() const {
return _enable_memory_optimization;
}
bool enable_ir_optimization() const { return _enable_ir_optimization; }
bool use_trt() const { return _use_trt; }
void set_static_optimization(bool static_optimization = false) {
_static_optimization = static_optimization;
}
......@@ -86,6 +91,7 @@ class InferEngineCreationParams {
bool _enable_ir_optimization;
bool _static_optimization;
bool _force_update_static_cache;
bool _use_trt;
};
class InferEngine {
......@@ -172,6 +178,10 @@ class ReloadableInferEngine : public InferEngine {
force_update_static_cache);
}
if (conf.has_use_trt()) {
_infer_engine_params.set_use_trt(conf.use_trt());
}
if (!check_need_reload() || load(_infer_engine_params) != 0) {
LOG(ERROR) << "Failed load model_data_path" << _model_data_path;
return -1;
......@@ -553,8 +563,12 @@ class CloneDBReloadableInferEngine
};
template <typename FluidFamilyCore>
#ifdef WITH_TRT
class FluidInferEngine : public DBReloadableInferEngine<FluidFamilyCore> {
#else
class FluidInferEngine : public CloneDBReloadableInferEngine<FluidFamilyCore> {
public:
#endif
public: // NOLINT
FluidInferEngine() {}
~FluidInferEngine() {}
......
core/sdk-cpp/include/abtest.h
浏览文件 @ af65b28f
......@@ -51,8 +51,8 @@ class WeightedRandomRender : public EndpointRouterBase {
new (std::nothrow) Factory<WeightedRandomRender, EndpointRouterBase>();
if (factory == NULL) {
RAW_LOG(ERROR,
"Failed regist factory: WeightedRandomRender->EndpointRouterBase \
in macro!");
"Failed regist factory: WeightedRandomRender->EndpointRouterBase "
"in macro!");
return -1;
}
......@@ -63,8 +63,8 @@ class WeightedRandomRender : public EndpointRouterBase {
if (FactoryPool<EndpointRouterBase>::instance().register_factory(
"WeightedRandomRender", factory) != 0) {
RAW_LOG(INFO,
"Factory has been registed: \
WeightedRandomRender->EndpointRouterBase.");
"Factory has been registed: "
"WeightedRandomRender->EndpointRouterBase.");
}
return 0;
......
doc/BERT_10_MINS.md
浏览文件 @ af65b28f
......@@ -56,21 +56,25 @@ the script of client side bert_client.py is as follow:
[//file]:#bert_client.py
``` python
import os
import sys
from paddle_serving_client import Client
from paddle_serving_client.utils import benchmark_args
from paddle_serving_app.reader import ChineseBertReader
import numpy as np
args = benchmark_args()
reader = ChineseBertReader()
reader = ChineseBertReader({"max_seq_len": 128})
fetch = ["pooled_output"]
endpoint_list = ["127.0.0.1:9292"]
endpoint_list = ['127.0.0.1:9292']
client = Client()
client.load_client_config("bert_seq20_client/serving_client_conf.prototxt")
client.load_client_config(args.model)
client.connect(endpoint_list)
for line in sys.stdin:
feed_dict = reader.process(line)
result = client.predict(feed=feed_dict, fetch=fetch)
for key in feed_dict.keys():
feed_dict[key] = np.array(feed_dict[key]).reshape((128, 1))
result = client.predict(feed=feed_dict, fetch=fetch, batch=False)
```
run
......
doc/BERT_10_MINS_CN.md
浏览文件 @ af65b28f
......@@ -52,18 +52,23 @@ pip install paddle_serving_app
``` python
import sys
from paddle_serving_client import Client
from paddle_serving_client.utils import benchmark_args
from paddle_serving_app.reader import ChineseBertReader
import numpy as np
args = benchmark_args()
reader = ChineseBertReader()
reader = ChineseBertReader({"max_seq_len": 128})
fetch = ["pooled_output"]
endpoint_list = ["127.0.0.1:9292"]
endpoint_list = ['127.0.0.1:9292']
client = Client()
client.load_client_config("bert_seq20_client/serving_client_conf.prototxt")
client.load_client_config(args.model)
client.connect(endpoint_list)
for line in sys.stdin:
feed_dict = reader.process(line)
result = client.predict(feed=feed_dict, fetch=fetch)
for key in feed_dict.keys():
feed_dict[key] = np.array(feed_dict[key]).reshape((128, 1))
result = client.predict(feed=feed_dict, fetch=fetch, batch=False)
```
执行
......
doc/COMPILE.md
浏览文件 @ af65b28f
......@@ -75,10 +75,12 @@ export PATH=$PATH:$GOPATH/bin
## Get go packages
```shell
go get -u github.com/grpc-ecosystem/grpc-gateway/protoc-gen-grpc-gateway
go get -u github.com/grpc-ecosystem/grpc-gateway/protoc-gen-swagger
go get -u github.com/golang/protobuf/protoc-gen-go
go get -u google.golang.org/grpc
go env -w GO111MODULE=on
go env -w GOPROXY=https://goproxy.cn,direct
go get -u github.com/grpc-ecosystem/grpc-gateway/protoc-gen-grpc-gateway@v1.15.2
go get -u github.com/grpc-ecosystem/grpc-gateway/protoc-gen-swagger@v1.15.2
go get -u github.com/golang/protobuf/protoc-gen-go@v1.4.3
go get -u google.golang.org/grpc@v1.33.0
```
......@@ -98,17 +100,47 @@ make -j10
you can execute `make install` to put targets under directory `./output`, you need to add`-DCMAKE_INSTALL_PREFIX=./output`to specify output path to cmake command shown above.
### Integrated GPU version paddle inference library
### CUDA_PATH is the cuda install path,use the command(whereis cuda) to check,it should be /usr/local/cuda.
### CUDNN_LIBRARY && CUDA_CUDART_LIBRARY is the lib path, it should be /usr/local/cuda/lib64/
``` shell
export CUDA_PATH='/usr/local'
export CUDNN_LIBRARY='/usr/local/cuda/lib64/'
export CUDA_CUDART_LIBRARY="/usr/local/cuda/lib64/"
mkdir server-build-gpu && cd server-build-gpu
cmake -DPYTHON_INCLUDE_DIR=$PYTHONROOT/include/python2.7/ \
-DPYTHON_LIBRARIES=$PYTHONROOT/lib/libpython2.7.so \
-DPYTHON_EXECUTABLE=$PYTHONROOT/bin/python \
-DCUDA_TOOLKIT_ROOT_DIR=${CUDA_PATH} \
-DCUDNN_LIBRARY=${CUDNN_LIBRARY} \
-DCUDA_CUDART_LIBRARY=${CUDA_CUDART_LIBRARY} \
-DSERVER=ON \
-DWITH_GPU=ON ..
make -j10
```
### Integrated TRT version paddle inference library
```
export CUDA_PATH='/usr/local'
export CUDNN_LIBRARY='/usr/local/cuda/lib64/'
export CUDA_CUDART_LIBRARY="/usr/local/cuda/lib64/"
mkdir server-build-trt && cd server-build-trt
cmake -DPYTHON_INCLUDE_DIR=$PYTHONROOT/include/python2.7/ \
-DPYTHON_LIBRARIES=$PYTHONROOT/lib/libpython2.7.so \
-DPYTHON_EXECUTABLE=$PYTHONROOT/bin/python \
-DTENSORRT_ROOT=${TENSORRT_LIBRARY_PATH} \
-DCUDA_TOOLKIT_ROOT_DIR=${CUDA_PATH} \
-DCUDNN_LIBRARY=${CUDNN_LIBRARY} \
-DCUDA_CUDART_LIBRARY=${CUDA_CUDART_LIBRARY} \
-DSERVER=ON \
-DWITH_GPU=ON \
-DWITH_TRT=ON ..
make -j10
```
execute `make install` to put targets under directory `./output`
**Attention：** After the compilation is successful, you need to set the path of `SERVING_BIN`. See [Note](https://github.com/PaddlePaddle/Serving/blob/develop/doc/COMPILE.md#Note) for details.
......@@ -134,7 +166,10 @@ execute `make install` to put targets under directory `./output`
```bash
mkdir app-build && cd app-build
cmake -DPYTHON_INCLUDE_DIR=$PYTHONROOT/include/python2.7/ -DPYTHON_LIBRARIES=$PYTHONROOT/lib/libpython2.7.so -DPYTHON_EXECUTABLE=$PYTHONROOT/bin/python -DAPP=ON ..
cmake -DPYTHON_INCLUDE_DIR=$PYTHONROOT/include/python2.7/ \
-DPYTHON_LIBRARIES=$PYTHONROOT/lib/libpython2.7.so \
-DPYTHON_EXECUTABLE=$PYTHONROOT/bin/python \
-DAPP=ON ..
make
```
......@@ -143,12 +178,14 @@ make
## Install wheel package
Regardless of the client, server or App part, after compiling, install the whl package in `python/dist/` in the temporary directory(`server-build-cpu`, `server-build-gpu`, `client-build`,`app-build`) of the compilation process.
for example：cd server-build-cpu/python/dist && pip install -U xxxxx.whl
## Note
When running the python server, it will check the `SERVING_BIN` environment variable. If you want to use your own compiled binary file, set the environment variable to the path of the corresponding binary file, usually`export SERVING_BIN=${BUILD_DIR}/core/general-server/serving`.
BUILD_DIR is the absolute path of server build CPU or server build GPU。
for example: cd server-build-cpu && export SERVING_BIN=${PWD}/core/general-server/serving
......@@ -165,7 +202,9 @@ Please use the example under `python/examples` to verify.
| WITH_AVX | Compile Paddle Serving with AVX intrinsics | OFF |
| WITH_MKL | Compile Paddle Serving with MKL support | OFF |
| WITH_GPU | Compile Paddle Serving with NVIDIA GPU | OFF |
| CUDNN_ROOT | Define CuDNN library and header path | |
| CUDNN_LIBRARY | Define CuDNN library and header path | |
| CUDA_TOOLKIT_ROOT_DIR | Define CUDA PATH | |
| TENSORRT_ROOT | Define TensorRT PATH | |
| CLIENT | Compile Paddle Serving Client | OFF |
| SERVER | Compile Paddle Serving Server | OFF |
| APP | Compile Paddle Serving App package | OFF |
......@@ -180,7 +219,8 @@ To compile the Paddle Serving GPU version on bare metal, you need to install the
- CUDA
- CuDNN
- NCCL2
To compile the TensorRT version, you need to install the TensorRT library.
Note here:
......@@ -190,21 +230,12 @@ Note here:
The following is the base library version matching relationship used by the PaddlePaddle release version for reference:
| | CUDA | CuDNN | NCCL2 |
| | CUDA | CuDNN | TensorRT |
| :----: | :-----: | :----------------------: | :----: |
| CUDA 8 | 8.0.61 | CuDNN 7.1.2 for CUDA 8.0 | 2.1.4 |
| CUDA 9 | 9.0.176 | CuDNN 7.3.1 for CUDA 9.0 | 2.2.12 |
| post9 | 9.0 | CuDNN 7.3.1 for CUDA 9.0 | |
| post10 | 10.0 | CuDNN 7.5.1 for CUDA 10.0| |
| trt | 10.1 | CuDNN 7.5.1 for CUDA 10.1| 6.0.1.5 |
### How to make the compiler detect the CuDNN library
Download the corresponding CUDNN version from NVIDIA developer official website and decompressing it, add `-DCUDNN_ROOT` to cmake command, to specify the path of CUDNN.
### How to make the compiler detect the nccl library
After downloading the corresponding version of the nccl2 library from the NVIDIA developer official website and decompressing it, add the following environment variables (take nccl2.1.4 as an example):
```shell
export C_INCLUDE_PATH=/path/to/nccl2/cuda8/nccl_2.1.4-1+cuda8.0_x86_64/include:$C_INCLUDE_PATH
export CPLUS_INCLUDE_PATH=/path/to/nccl2/cuda8/nccl_2.1.4-1+cuda8.0_x86_64/include:$CPLUS_INCLUDE_PATH
export LD_LIBRARY_PATH=/path/to/nccl2/cuda8/nccl_2.1.4-1+cuda8.0_x86_64/lib/:$LD_LIBRARY_PATH
```
doc/COMPILE_CN.md
浏览文件 @ af65b28f
......@@ -72,10 +72,12 @@ export PATH=$PATH:$GOPATH/bin
## 获取 Go packages
```shell
go get -u github.com/grpc-ecosystem/grpc-gateway/protoc-gen-grpc-gateway
go get -u github.com/grpc-ecosystem/grpc-gateway/protoc-gen-swagger
go get -u github.com/golang/protobuf/protoc-gen-go
go get -u google.golang.org/grpc
go env -w GO111MODULE=on
go env -w GOPROXY=https://goproxy.cn,direct
go get -u github.com/grpc-ecosystem/grpc-gateway/protoc-gen-grpc-gateway@v1.15.2
go get -u github.com/grpc-ecosystem/grpc-gateway/protoc-gen-swagger@v1.15.2
go get -u github.com/golang/protobuf/protoc-gen-go@v1.4.3
go get -u google.golang.org/grpc@v1.33.0
```
......@@ -85,31 +87,68 @@ go get -u google.golang.org/grpc
``` shell
mkdir server-build-cpu && cd server-build-cpu
cmake -DPYTHON_INCLUDE_DIR=$PYTHONROOT/include/python2.7/ -DPYTHON_LIBRARIES=$PYTHONROOT/lib/libpython2.7.so -DPYTHON_EXECUTABLE=$PYTHONROOT/bin/python -DSERVER=ON ..
cmake -DPYTHON_INCLUDE_DIR=$PYTHONROOT/include/python2.7/ \
-DPYTHON_LIBRARIES=$PYTHONROOT/lib/libpython2.7.so \
-DPYTHON_EXECUTABLE=$PYTHONROOT/bin/python \
-DSERVER=ON ..
make -j10
```
可以执行`make install`把目标产出放在`./output`目录下，cmake阶段需添加`-DCMAKE_INSTALL_PREFIX=./output`选项来指定存放路径。
### 集成GPU版本Paddle Inference Library
### CUDA_PATH是cuda的安装路径，可以使用命令行whereis cuda命令确认你的cuda安装路径，通常应该是/usr/local/cuda
### CUDNN_LIBRARY CUDA_CUDART_LIBRARY 是cuda库文件的路径，通常应该是/usr/local/cuda/lib64/
``` shell
export CUDA_PATH='/usr/local'
export CUDNN_LIBRARY='/usr/local/cuda/lib64/'
export CUDA_CUDART_LIBRARY="/usr/local/cuda/lib64/"
mkdir server-build-gpu && cd server-build-gpu
cmake -DPYTHON_INCLUDE_DIR=$PYTHONROOT/include/python2.7/ -DPYTHON_LIBRARIES=$PYTHONROOT/lib/libpython2.7.so -DPYTHON_EXECUTABLE=$PYTHONROOT/bin/python -DSERVER=ON -DWITH_GPU=ON ..
cmake -DPYTHON_INCLUDE_DIR=$PYTHONROOT/include/python2.7/ \
-DPYTHON_LIBRARIES=$PYTHONROOT/lib/libpython2.7.so \
-DPYTHON_EXECUTABLE=$PYTHONROOT/bin/python \
-DCUDA_TOOLKIT_ROOT_DIR=${CUDA_PATH} \
-DCUDNN_LIBRARY=${CUDNN_LIBRARY} \
-DCUDA_CUDART_LIBRARY=${CUDA_CUDART_LIBRARY} \
-DSERVER=ON \
-DWITH_GPU=ON ..
make -j10
```
执行`make install`可以把目标产出放在`./output`目录下。
### 集成TensorRT版本Paddle Inference Library
**注意：** 编译成功后，需要设置`SERVING_BIN`路径，详见后面的[注意事项](https://github.com/PaddlePaddle/Serving/blob/develop/doc/COMPILE_CN.md#注意事项)。
```
export CUDA_PATH='/usr/local'
export CUDNN_LIBRARY='/usr/local/cuda/lib64/'
export CUDA_CUDART_LIBRARY="/usr/local/cuda/lib64/"
mkdir server-build-trt && cd server-build-trt
cmake -DPYTHON_INCLUDE_DIR=$PYTHONROOT/include/python2.7/ \
-DPYTHON_LIBRARIES=$PYTHONROOT/lib/libpython2.7.so \
-DPYTHON_EXECUTABLE=$PYTHONROOT/bin/python \
-DTENSORRT_ROOT=${TENSORRT_LIBRARY_PATH} \
-DCUDA_TOOLKIT_ROOT_DIR=${CUDA_PATH} \
-DCUDNN_LIBRARY=${CUDNN_LIBRARY} \
-DCUDA_CUDART_LIBRARY=${CUDA_CUDART_LIBRARY} \
-DSERVER=ON \
-DWITH_GPU=ON \
-DWITH_TRT=ON ..
make -j10
```
执行`make install`可以把目标产出放在`./output`目录下。
**注意：** 编译成功后，需要设置`SERVING_BIN`路径，详见后面的[注意事项](https://github.com/PaddlePaddle/Serving/blob/develop/doc/COMPILE_CN.md#注意事项)。
## 编译Client部分
``` shell
mkdir client-build && cd client-build
cmake -DPYTHON_INCLUDE_DIR=$PYTHONROOT/include/python2.7/ -DPYTHON_LIBRARIES=$PYTHONROOT/lib/libpython2.7.so -DPYTHON_EXECUTABLE=$PYTHONROOT/bin/python -DCLIENT=ON ..
cmake -DPYTHON_INCLUDE_DIR=$PYTHONROOT/include/python2.7/ \
-DPYTHON_LIBRARIES=$PYTHONROOT/lib/libpython2.7.so \
-DPYTHON_EXECUTABLE=$PYTHONROOT/bin/python \
-DCLIENT=ON ..
make -j10
```
......@@ -121,7 +160,11 @@ make -j10
```bash
mkdir app-build && cd app-build
cmake -DPYTHON_INCLUDE_DIR=$PYTHONROOT/include/python2.7/ -DPYTHON_LIBRARIES=$PYTHONROOT/lib/libpython2.7.so -DPYTHON_EXECUTABLE=$PYTHONROOT/bin/python -DCMAKE_INSTALL_PREFIX=./output -DAPP=ON ..
cmake -DPYTHON_INCLUDE_DIR=$PYTHONROOT/include/python2.7/ \
-DPYTHON_LIBRARIES=$PYTHONROOT/lib/libpython2.7.so \
-DPYTHON_EXECUTABLE=$PYTHONROOT/bin/python \
-DCMAKE_INSTALL_PREFIX=./output \
-DAPP=ON ..
make
```
......@@ -130,12 +173,16 @@ make
## 安装wheel包
无论是Client端，Server端还是App部分，编译完成后，安装编译过程临时目录（`server-build-cpu`、`server-build-gpu`、`client-build`、`app-build`）下的`python/dist/` 中的whl包即可。
例如：cd server-build-cpu/python/dist && pip install -U xxxxx.whl
## 注意事项
运行python端Server时，会检查`SERVING_BIN`环境变量，如果想使用自己编译的二进制文件，请将设置该环境变量为对应二进制文件的路径，通常是`export SERVING_BIN=${BUILD_DIR}/core/general-server/serving`。
其中BUILD_DIR为server-build-cpu或server-build-gpu的绝对路径。
可以cd server-build-cpu路径下，执行export SERVING_BIN=${PWD}/core/general-server/serving
......@@ -152,7 +199,10 @@ make
| WITH_AVX | Compile Paddle Serving with AVX intrinsics | OFF |
| WITH_MKL | Compile Paddle Serving with MKL support | OFF |
| WITH_GPU | Compile Paddle Serving with NVIDIA GPU | OFF |
| CUDNN_ROOT | Define CuDNN library and header path | |
| WITH_TRT | Compile Paddle Serving with TensorRT | OFF |
| CUDNN_LIBRARY | Define CuDNN library and header path | |
| CUDA_TOOLKIT_ROOT_DIR | Define CUDA PATH | |
| TENSORRT_ROOT | Define TensorRT PATH | |
| CLIENT | Compile Paddle Serving Client | OFF |
| SERVER | Compile Paddle Serving Server | OFF |
| APP | Compile Paddle Serving App package | OFF |
......@@ -167,7 +217,8 @@ Paddle Serving通过PaddlePaddle预测库支持在GPU上做预测。WITH_GPU选
- CUDA
- CuDNN
- NCCL2
编译TensorRT版本，需要安装TensorRT库。
这里要注意的是：
......@@ -176,21 +227,12 @@ Paddle Serving通过PaddlePaddle预测库支持在GPU上做预测。WITH_GPU选
以下是PaddlePaddle发布版本所使用的基础库版本匹配关系，供参考：
| | CUDA | CuDNN | NCCL2 |
| | CUDA | CuDNN | TensorRT |
| :----: | :-----: | :----------------------: | :----: |
| CUDA 8 | 8.0.61 | CuDNN 7.1.2 for CUDA 8.0 | 2.1.4 |
| CUDA 9 | 9.0.176 | CuDNN 7.3.1 for CUDA 9.0 | 2.2.12 |
| post9 | 9.0 | CuDNN 7.3.1 for CUDA 9.0 | |
| post10 | 10.0 | CuDNN 7.5.1 for CUDA 10.0| |
| trt | 10.1 | CuDNN 7.5.1 for CUDA 10.1| 6.0.1.5 |
### 如何让Paddle Serving编译系统探测到CuDNN库
从NVIDIA developer官网下载对应版本CuDNN并在本地解压后，在cmake编译命令中增加`-DCUDNN_ROOT`参数，指定CuDNN库所在路径。
### 如何让Paddle Serving编译系统探测到nccl库
从NVIDIA developer官网下载对应版本nccl2库并解压后，增加如下环境变量 (以nccl2.1.4为例)：
```shell
export C_INCLUDE_PATH=/path/to/nccl2/cuda8/nccl_2.1.4-1+cuda8.0_x86_64/include:$C_INCLUDE_PATH
export CPLUS_INCLUDE_PATH=/path/to/nccl2/cuda8/nccl_2.1.4-1+cuda8.0_x86_64/include:$CPLUS_INCLUDE_PATH
export LD_LIBRARY_PATH=/path/to/nccl2/cuda8/nccl_2.1.4-1+cuda8.0_x86_64/lib/:$LD_LIBRARY_PATH
```
从NVIDIA developer官网下载对应版本CuDNN并在本地解压后，在cmake编译命令中增加`-DCUDNN_LIBRARY`参数，指定CuDNN库所在路径。
doc/DOCKER_IMAGES.md
浏览文件 @ af65b28f
......@@ -25,10 +25,10 @@ You can get images in two ways:
## Image description
Runtime images cannot be used for compilation.
If you want to customize your Serving based on source code, use the version with the suffix - devel.
| Description | OS | TAG | Dockerfile |
| :----------------------------------------------------------: | :-----: | :--------------------------: | :----------------------------------------------------------: |
......@@ -40,3 +40,13 @@ Runtime images cannot be used for compilation.
| GPU (cuda10.0-cudnn7) development | CentOS7 | latest-cuda10.0-cudnn7-devel | [Dockerfile.cuda10.0-cudnn7.devel](../tools/Dockerfile.cuda10.0-cudnn7.devel) |
| CPU development (Used to compile packages on Ubuntu) | CentOS6 | <None> | [Dockerfile.centos6.devel](../tools/Dockerfile.centos6.devel) |
| GPU (cuda9.0-cudnn7) development (Used to compile packages on Ubuntu) | CentOS6 | <None> | [Dockerfile.centos6.cuda9.0-cudnn7.devel](../tools/Dockerfile.centos6.cuda9.0-cudnn7.devel) |
## Requirements for running CUDA containers
Running a CUDA container requires a machine with at least one CUDA-capable GPU and a driver compatible with the CUDA toolkit version you are using.
The machine running the CUDA container **only requires the NVIDIA driver**, the CUDA toolkit doesn't have to be installed.
For the relationship between CUDA toolkit version, Driver version and GPU architecture, please refer to [nvidia-docker wiki](https://github.com/NVIDIA/nvidia-docker/wiki/CUDA).
doc/DOCKER_IMAGES_CN.md
浏览文件 @ af65b28f
......@@ -25,10 +25,10 @@
## 镜像说明
运行时镜像不能用于开发编译。
若需要基于源代码二次开发编译，请使用后缀为-devel的版本。
| 镜像说明 | 操作系统 | TAG | Dockerfile |
| -------------------------------------------------- | -------- | ---------------------------- | ------------------------------------------------------------ |
......@@ -40,3 +40,13 @@
| GPU (cuda10.0-cudnn7) 开发镜像 | CentOS7 | latest-cuda10.0-cudnn7-devel | [Dockerfile.cuda10.0-cudnn7.devel](../tools/Dockerfile.cuda10.0-cudnn7.devel) |
| CPU 开发镜像 (用于编译 Ubuntu 包) | CentOS6 | <无> | [Dockerfile.centos6.devel](../tools/Dockerfile.centos6.devel) |
| GPU (cuda9.0-cudnn7) 开发镜像 (用于编译 Ubuntu 包) | CentOS6 | <无> | [Dockerfile.centos6.cuda9.0-cudnn7.devel](../tools/Dockerfile.centos6.cuda9.0-cudnn7.devel) |
## 运行CUDA容器的要求
运行CUDA容器需要至少具有一个支持CUDA的GPU以及与您所使用的CUDA工具包版本兼容的驱动程序。
运行CUDA容器的机器**只需要相应的NVIDIA驱动程序**，而CUDA工具包不是必要的。
相关CUDA工具包版本、驱动版本和GPU架构的关系请参阅 [nvidia-docker wiki](https://github.com/NVIDIA/nvidia-docker/wiki/CUDA)。
doc/FAQ.md
浏览文件 @ af65b28f
# FAQ
- Q: 如何调整RPC服务的等待时间，避免超时？
A: 使用set_rpc_timeout_ms设置更长的等待时间，单位为毫秒，默认时间为20秒。
示例：
```
from paddle_serving_client import Client
## 基础知识
client = Client()
client.load_client_config(sys.argv[1])
client.set_rpc_timeout_ms(100000)
client.connect(["127.0.0.1:9393"])
```
#### Q: Paddle Serving 、Paddle Inference、PaddleHub Serving三者的区别及联系？
- Q: 如何使用自己编译的Paddle Serving进行预测？
**A:** paddle serving是远程服务，即发起预测的设备（手机、浏览器、客户端等）与实际预测的硬件不在一起。 paddle inference是一个library，适合嵌入到一个大系统中保证预测效率，paddle serving调用了paddle inference做远程服务。paddlehub serving可以认为是一个示例，都会使用paddle serving作为统一预测服务入口。如果在web端交互，一般是调用远程服务的形式，可以使用paddle serving的web service搭建。
A: 通过pip命令安装自己编译出的whl包，并设置SERVING_BIN环境变量为编译出的serving二进制文件路径。
#### Q: paddle-serving是否支持Int32支持
- Q: 执行GPU预测时遇到InvalidArgumentError: Device id must be less than GPU count, but received id is: 0. GPU count is: 0.
**A:** 在protobuf定feed_type和fetch_type编号与数据类型对应如下
A: 将显卡驱动对应的libcuda.so的目录添加到LD_LIBRARY_PATH环境变量中
​ 0-int64
- Q: 执行GPU预测时遇到ExternalError: Cudnn error, CUDNN_STATUS_BAD_PARAM at (/home/scmbuild/workspaces_cluster.dev/baidu.lib.paddlepaddle/baidu/lib/paddlepaddle/Paddle/paddle/fluid/operators/batch_norm_op.cu:198)
​ 1-float32
A: 将cudnn的lib64路径添加到LD_LIBRARY_PATH，安装自pypi的Paddle Serving中post9版使用的是cudnn 7.3,post10使用的是cudnn 7.5。如果是使用自己编译的Paddle Serving，可以在log/serving.INFO日志文件中查看对应的cudnn版本。
​ 2-int32
- Q: 执行GPU预测时遇到Error: Failed to find dynamic library: libcublas.so
#### Q: paddle-serving是否支持windows和Linux环境下的多线程调用
A: 将cuda的lib64路径添加到LD_LIBRARY_PATH, post9版本的Paddle Serving使用的是cuda 9.0，post10版本使用的cuda 10.0。
**A:** 客户端可以发起多线程访问调用服务端
- Q: 部署和预测中的日志信息在哪里查看？
#### Q: paddle-serving如何修改消息大小限制
- A: server端的日志分为两部分，一部分打印到标准输出，一部分打印到启动服务时的目录下的log/serving.INFO文件中。
**A:** 在server端和client但通过FLAGS_max_body_size来扩大数据量限制，单位为字节，默认为64MB
client端的日志直接打印到标准输出。
#### Q: paddle-serving客户端目前支持哪些语言
通过在部署服务之前 'export GLOG_v=3'可以输出更为详细的日志信息。
**A:** java c++ python
#### Q: paddle-serving目前支持哪些协议
**A:** http rpc
## 编译问题
#### Q: 如何使用自己编译的Paddle Serving进行预测？
**A:** 通过pip命令安装自己编译出的whl包，并设置SERVING_BIN环境变量为编译出的serving二进制文件路径。
#### Q: 使用Java客户端，mvn compile过程出现"No compiler is provided in this environment. Perhaps you are running on a JRE rather than a JDK?"错误
**A:** 没有安装JDK，或者JAVA_HOME路径配置错误（正确配置是JDK路径，常见错误配置成JRE路径，例如正确路径参考JAVA_HOME="/usr/lib/jvm/java-1.8.0-openjdk-1.8.0.262.b10-0.el7_8.x86_64/"）。Java JDK安装参考https://segmentfault.com/a/1190000015389941
## 部署问题
#### Q: GPU环境运行Serving报错，GPU count is: 0。
```
terminate called after throwing an instance of 'paddle::platform::EnforceNotMet'
what():
--------------------------------------------
C++ Call Stacks (More useful to developers):
--------------------------------------------
0 std::string paddle::platform::GetTraceBackString<std::string const&>(std::string const&, char const*, int)
1 paddle::platform::SetDeviceId(int)
2 paddle::AnalysisConfig::fraction_of_gpu_memory_for_pool() const
3 std::unique_ptr<paddle::PaddlePredictor, std::default_delete<paddle::PaddlePredictor> > paddle::CreatePaddlePredictor<paddle::AnalysisConfig, (paddle::PaddleEngineKind)2>(paddle::AnalysisConfig const&)
4 std::unique_ptr<paddle::PaddlePredictor, std::default_delete<paddle::PaddlePredictor> > paddle::CreatePaddlePredictor<paddle::AnalysisConfig>(paddle::AnalysisConfig const&)
----------------------
Error Message Summary:
----------------------
InvalidArgumentError: Device id must be less than GPU count, but received id is: 0. GPU count is: 0.
[Hint: Expected id < GetCUDADeviceCount(), but received id:0 >= GetCUDADeviceCount():0.] at (/home/scmbuild/workspaces_cluster.dev/baidu.lib.paddlepaddle/baidu/lib/paddlepaddle/Paddle/paddle/fluid/platform/gpu_info.cc:211)
```
**A:** libcuda.so没有链接成功。首先在机器上找到libcuda.so，ldd检查libnvidia版本与nvidia-smi中版本一致（libnvidia-fatbinaryloader.so.418.39，与NVIDIA-SMI 418.39 Driver Version: 418.39）,然后用export导出libcuda.so的路径即可（例如libcuda.so在/usr/lib64/，export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/lib64/）
#### Q: 遇到 GPU not found, please check your environment or use cpu version by "pip install paddle_serving_server"
**A:** 检查环境中是否有N卡：ls /dev/ | grep nvidia
#### Q: 目前Paddle Serving支持哪些镜像环境？
**A:** 目前（0.4.0）仅支持CentOS，具体列表查阅[这里](https://github.com/PaddlePaddle/Serving/blob/develop/doc/DOCKER_IMAGES.md)
#### Q: python编译的GCC版本与serving的版本不匹配
**A:**:1)使用[GPU docker](https://github.com/PaddlePaddle/Serving/blob/develop/doc/RUN_IN_DOCKER.md#gpunvidia-docker)解决环境问题
​ 2)修改anaconda的虚拟环境下安装的python的gcc版本[参考](https://www.jianshu.com/p/c498b3d86f77)
#### Q: paddle-serving是否支持本地离线安装
**A:** 支持离线部署，需要把一些相关的[依赖包](https://github.com/PaddlePaddle/Serving/blob/develop/doc/COMPILE.md)提前准备安装好
## 预测问题
#### Q: 使用GPU第一次预测时特别慢，如何调整RPC服务的等待时间避免超时？
**A:** GPU第一次预测需要初始化。使用set_rpc_timeout_ms设置更长的等待时间，单位为毫秒，默认时间为20秒。
示例：
```
from paddle_serving_client import Client
client = Client()
client.load_client_config(sys.argv[1])
client.set_rpc_timeout_ms(100000)
client.connect(["127.0.0.1:9393"])
```
#### Q: 执行GPU预测时遇到InvalidArgumentError: Device id must be less than GPU count, but received id is: 0. GPU count is: 0.
**A:** 将显卡驱动对应的libcuda.so的目录添加到LD_LIBRARY_PATH环境变量中
#### Q: 执行GPU预测时遇到ExternalError: Cudnn error, CUDNN_STATUS_BAD_PARAM at (../batch_norm_op.cu:198)
**A:** 将cudnn的lib64路径添加到LD_LIBRARY_PATH，安装自pypi的Paddle Serving中post9版使用的是cudnn 7.3,post10使用的是cudnn 7.5。如果是使用自己编译的Paddle Serving，可以在log/serving.INFO日志文件中查看对应的cudnn版本。
#### Q: 执行GPU预测时遇到Error: Failed to find dynamic library: libcublas.so
**A:** 将cuda的lib64路径添加到LD_LIBRARY_PATH, post9版本的Paddle Serving使用的是cuda 9.0，post10版本使用的cuda 10.0。
#### Q: Client端fetch的变量名如何设置
**A:** 可以查看配置文件serving_server_conf.prototxt，获取需要的变量名
#### Q: 如何使用多语言客户端
**A:** 多语言客户端要与多语言服务端配套使用。当前版本下（0.4.0），服务端需要将Server改为MultiLangServer（如果是以命令行启动的话只需要添加--use_multilang参数），Python客户端需要将Client改为MultiLangClient，同时去除load_client_config的过程。[Java客户端参考文档](https://github.com/PaddlePaddle/Serving/blob/develop/doc/JAVA_SDK_CN.md)
#### Q: 如何在Windows下使用Paddle Serving
**A:** 当前版本（0.4.0）在Windows上可以运行多语言RPC客户端，或使用HTTP方式访问。如果使用多语言RPC客户端，需要在Linux环境（比如本机容器，或远程Linux机器）中运行多语言服务端；如果使用HTTP方式，需要在Linux环境中运行普通服务端
#### Q: libnvinfer.so: cannot open shared object file: No such file or directory)
**A:** 参考该文档安装TensorRT: https://blog.csdn.net/hesongzefairy/article/details/105343525
## 日志排查
#### Q: 部署和预测中的日志信息在哪里查看？
**A:** server端的日志分为两部分，一部分打印到标准输出，一部分打印到启动服务时的目录下的log/serving.INFO文件中。
client端的日志直接打印到标准输出。
通过在部署服务之前 'export GLOG_v=3'可以输出更为详细的日志信息。
#### Q: paddle-serving启动成功后，相关的日志在哪里设置
**A:** 1)警告是glog组件打印的，告知glog初始化之前日志打印在STDERR
​ 2)一般采用GLOG_v方式启动服务同时设置日志级别。
例如：
```
GLOG_v=2 python -m paddle_serving_server.serve --model xxx_conf/ --port 9999
```
#### Q: （GLOG_v=2下）Server端日志一切正常，但Client端始终得不到正确的预测结果
**A:** 可能是配置文件有问题，检查下配置文件（is_load_tensor，fetch_type等有没有问题）
#### Q: 如何给Server传递Logid
**A:** Logid默认为0（后续应该有自动生成Logid的计划，当前版本0.4.0），Client端通过在predict函数中指定log_id参数传递
## 性能优化
doc/GRPC_IMPL_CN.md
浏览文件 @ af65b28f
# gRPC接口
# gRPC接口使用介绍
gRPC 接口实现形式类似 Web Service：
- [1.与bRPC接口对比](#1与brpc接口对比)
- [1.1 服务端对比](#11-服务端对比)
- [1.2 客服端对比](#12-客服端对比)
- [1.3 其他](#13-其他)
- [2.示例：线性回归预测服务](#2示例线性回归预测服务)
- [获取数据](#获取数据)
- [开启 gRPC 服务端](#开启-grpc-服务端)
- [客户端预测](#客户端预测)
- [同步预测](#同步预测)
- [异步预测](#异步预测)
- [Batch 预测](#batch-预测)
- [通用 pb 预测](#通用-pb-预测)
- [预测超时](#预测超时)
- [List 输入](#list-输入)
- [3.更多示例](#3更多示例)
![](grpc_impl.png)
使用gRPC接口，Client端可以在Win/Linux/MacOS平台上调用不同语言。gRPC 接口实现结构如下：
## 与bRPC接口对比
![](https://github.com/PaddlePaddle/Serving/blob/develop/doc/grpc_impl.png)
1. gRPC Server 端 `load_model_config` 函数添加 `client_config_path` 参数：
## 1.与bRPC接口对比
```python
#### 1.1 服务端对比
* gRPC Server 端 `load_model_config` 函数添加 `client_config_path` 参数：
```
def load_model_config(self, server_config_paths, client_config_path=None)
```
在一些例子中 bRPC Server 端与 bRPC Client 端的配置文件可能不同（如 在cube local 中，Client 端的数据先交给 cube，经过 cube 处理后再交给预测库），此时 gRPC Server 端需要手动设置 gRPC Client 端的配置`client_config_path`。
**`client_config_path` 默认为 `<server_config_path>/serving_server_conf.prototxt`。**
在一些例子中 bRPC Server 端与 bRPC Client 端的配置文件可能是不同的（如 cube local 例子中，Client 端的数据先交给 cube，经过 cube 处理后再交给预测库），所以 gRPC Server 端需要获取 gRPC Client 端的配置；同时为了取消 gRPC Client 端手动加载配置文件的过程，所以设计 gRPC Server 端同时加载两个配置文件。`client_config_path` 默认为 `<server_config_path>/serving_server_conf.prototxt`。
#### 1.2 客服端对比
2. gRPC Client 端取消 `load_client_config` 步骤：
* gRPC Client 端取消 `load_client_config` 步骤：
在 `connect` 步骤通过 RPC 获取相应的 prototxt（从任意一个 endpoint 获取即可）。
3. gRPC Client 需要通过 RPC 方式设置 timeout 时间（调用形式与 bRPC Client保持一致）
* gRPC Client 需要通过 RPC 方式设置 timeout 时间（调用形式与 bRPC Client保持一致）
因为 bRPC Client 在 `connect` 后无法更改 timeout 时间，所以当 gRPC Server 收到变更 timeout 的调用请求时会重新创建 bRPC Client 实例以变更 bRPC Client timeout时间，同时 gRPC Client 会设置 gRPC 的 deadline 时间。
**注意，设置 timeout 接口和 Inference 接口不能同时调用（非线程安全），出于性能考虑暂时不加锁。**
4. gRPC Client 端 `predict` 函数添加 `asyn` 和 `is_python` 参数：
* gRPC Client 端 `predict` 函数添加 `asyn` 和 `is_python` 参数：
```python
```
def predict(self, feed, fetch, need_variant_tag=False, asyn=False, is_python=True)
```
其中，`asyn` 为异步调用选项。当 `asyn=True` 时为异步调用，返回 `MultiLangPredictFuture` 对象，通过 `MultiLangPredictFuture.result()` 阻塞获取预测值；当 `asyn=Fasle` 为同步调用。
1. `asyn` 为异步调用选项。当 `asyn=True` 时为异步调用，返回 `MultiLangPredictFuture` 对象，通过 `MultiLangPredictFuture.result()` 阻塞获取预测值；当 `asyn=Fasle` 为同步调用。
2. `is_python` 为 proto 格式选项。当 `is_python=True` 时，基于 numpy bytes 格式进行数据传输，目前只适用于 Python；当 `is_python=False` 时，以普通数据格式传输，更加通用。使用 numpy bytes 格式传输耗时比普通数据格式小很多（详见 [#654](https://github.com/PaddlePaddle/Serving/pull/654)）。
#### 1.3 其他
* 异常处理：当 gRPC Server 端的 bRPC Client 预测失败（返回 `None`）时，gRPC Client 端同样返回None。其他 gRPC 异常会在 Client 内部捕获，并在返回的 fetch_map 中添加一个 "status_code" 字段来区分是否预测正常（参考 timeout 样例）。
* 由于 gRPC 只支持 pick_first 和 round_robin 负载均衡策略，ABTEST 特性还未打齐。
* 系统兼容性：
* [x] CentOS
* [x] macOS
* [x] Windows
* 已经支持的客户端语言：
- Python
- Java
- Go
## 2.示例：线性回归预测服务
以下是采用gRPC实现的关于线性回归预测的一个示例，具体代码详见此[链接](https://github.com/PaddlePaddle/Serving/tree/develop/python/examples/grpc_impl_example/fit_a_line)
#### 获取数据
```shell
sh get_data.sh
```
#### 开启 gRPC 服务端
``` shell
python test_server.py uci_housing_model/
```
也可以通过下面的一行代码开启默认 gRPC 服务：
```shell
python -m paddle_serving_server.serve --model uci_housing_model --thread 10 --port 9393 --use_multilang
```
注：--use_multilang参数用来启用多语言客户端
### 客户端预测
#### 同步预测
``` shell
python test_sync_client.py
```
#### 异步预测
``` shell
python test_asyn_client.py
```
#### Batch 预测
``` shell
python test_batch_client.py
```
`is_python` 为 proto 格式选项。当 `is_python=True` 时，基于 numpy bytes 格式进行数据传输，目前只适用于 Python；当 `is_python=False` 时，以普通数据格式传输，更加通用。使用 numpy bytes 格式传输耗时比普通数据格式小很多（详见 [#654](https://github.com/PaddlePaddle/Serving/pull/654)）。
#### 通用 pb 预测
5. 异常处理：当 gRPC Server 端的 bRPC Client 预测失败（返回 `None`）时，gRPC Client 端同样返回None。其他 gRPC 异常会在 Client 内部捕获，并在返回的 fetch_map 中添加一个 "status_code" 字段来区分是否预测正常（参考 timeout 样例）。
``` shell
python test_general_pb_client.py
```
6. 由于 gRPC 只支持 pick_first 和 round_robin 负载均衡策略，ABTEST 特性还未打齐。
#### 预测超时
7. 经测试，gRPC 版本可以在 Windows、macOS 平台使用。
``` shell
python test_timeout_client.py
```
8. 计划支持的客户端语言：
#### List 输入
- [x] Python
- [ ] Java
- [ ] Go
- [ ] JavaScript
``` shell
python test_list_input_client.py
```
## Python 端的一些例子
## 3.更多示例
详见 `python/examples/grpc_impl_example` 下的示例文件。
详见[`python/examples/grpc_impl_example`](https://github.com/PaddlePaddle/Serving/tree/develop/python/examples/grpc_impl_example)下的示例文件。
doc/INFERENCE_TO_SERVING.md
浏览文件 @ af65b28f
......@@ -24,13 +24,13 @@ inference_model_dir = "your_inference_model"
serving_client_dir = "serving_client_dir"
serving_server_dir = "serving_server_dir"
feed_var_names, fetch_var_names = inference_model_to_serving(
inference_model_dir, serving_client_dir, serving_server_dir)
inference_model_dir, serving_server_dir, serving_client_dir)
```
if your model file and params file are both standalone, please use the following api.
```
feed_var_names, fetch_var_names = inference_model_to_serving(
inference_model_dir, serving_client_dir, serving_server_dir,
inference_model_dir, serving_server_dir, serving_client_dir,
model_filename="model", params_filename="params")
```
doc/INFERENCE_TO_SERVING_CN.md
浏览文件 @ af65b28f
......@@ -23,11 +23,11 @@ inference_model_dir = "your_inference_model"
serving_client_dir = "serving_client_dir"
serving_server_dir = "serving_server_dir"
feed_var_names, fetch_var_names = inference_model_to_serving(
inference_model_dir, serving_client_dir, serving_server_dir)
inference_model_dir, serving_server_dir, serving_client_dir)
```
如果模型中有模型描述文件`model_filename` 和 模型参数文件`params_filename`，那么请用
```
feed_var_names, fetch_var_names = inference_model_to_serving(
inference_model_dir, serving_client_dir, serving_server_dir,
inference_model_dir, serving_server_dir, serving_client_dir,
model_filename="model", params_filename="params")
```
doc/LATEST_PACKAGES.md
浏览文件 @ af65b28f
......@@ -18,6 +18,8 @@ https://paddle-serving.bj.bcebos.com/whl/paddle_serving_server-0.0.0-py2-none-an
https://paddle-serving.bj.bcebos.com/whl/paddle_serving_server_gpu-0.0.0.post9-py3-none-any.whl
#cuda 10.0
https://paddle-serving.bj.bcebos.com/whl/paddle_serving_server_gpu-0.0.0.post10-py3-none-any.whl
#cuda10.1 with TensorRT 6
https://paddle-serving.bj.bcebos.com/whl/paddle_serving_server_gpu-0.0.0.trt-py3-none-any.whl
```
### Python 2
```
......@@ -25,6 +27,8 @@ https://paddle-serving.bj.bcebos.com/whl/paddle_serving_server_gpu-0.0.0.post10-
https://paddle-serving.bj.bcebos.com/whl/paddle_serving_server_gpu-0.0.0.post9-py2-none-any.whl
#cuda 10.0
https://paddle-serving.bj.bcebos.com/whl/paddle_serving_server_gpu-0.0.0.post10-py2-none-any.whl
##cuda10.1 with TensorRT 6
https://paddle-serving.bj.bcebos.com/whl/paddle_serving_server_gpu-0.0.0.trt-py2-none-any.whl
```
## Client
......@@ -36,6 +40,10 @@ https://paddle-serving.bj.bcebos.com/whl/paddle_serving_client-0.0.0-cp37-none-a
```
https://paddle-serving.bj.bcebos.com/whl/paddle_serving_client-0.0.0-cp36-none-any.whl
```
### Python 3.5
```
https://paddle-serving.bj.bcebos.com/whl/paddle_serving_client-0.0.0-cp35-none-any.whl
```
### Python 2.7
```
https://paddle-serving.bj.bcebos.com/whl/paddle_serving_client-0.0.0-cp27-none-any.whl
......
doc/PIPELINE_SERVING.md
浏览文件 @ af65b28f
......@@ -7,14 +7,46 @@ Paddle Serving is usually used for the deployment of single model, but the end-t
Paddle Serving provides a user-friendly programming framework for multi-model composite services, Pipeline Serving, which aims to reduce the threshold of programming, improve resource utilization (especially GPU), and improve the prediction efficiency.
## Architecture Design
## ★ Architecture Design
The Server side is built based on gRPC and graph execution engine. The relationship between them is shown in the following figure.
The Server side is built based on <b>RPC Service</b> and <b>graph execution engine</b>. The relationship between them is shown in the following figure.
<center>
<img src='pipeline_serving-image1.png' height = "250" align="middle"/>
</center>
### Graph Execution Engine
### 1. RPC Service
In order to meet the needs of different users, the RPC service starts one Web server and one RPC server at the same time, and can process 2 types of requests, RESTful API and gRPC.The gPRC gateway receives RESTful API requests and forwards requests to the gRPC server through the reverse proxy server; gRPC requests are received by the gRPC server, so the two types of requests are processed by the gRPC Service in a unified manner to ensure that the processing logic is consistent.
#### <b>1.1 Request and Respose of proto
gRPC service and gRPC gateway service are generated with service.proto.
```proto
message Request {
repeated string key = 1;
repeated string value = 2;
optional string name = 3;
optional string method = 4;
optional int64 logid = 5;
optional string clientip = 6;
};
message Response {
optional int32 err_no = 1;
optional string err_msg = 2;
repeated string key = 3;
repeated string value = 4;
};
```
The `key` and `value` in the Request are paired string arrays. The `name` and `method` correspond to the URL of the RESTful API://{ip}:{port}/{name}/{method}.The `logid` and `clientip` are convenient for users to connect service-level requests and customize strategies.
In Response, `err_no` and `err_msg` express the correctness and error information of the processing result, and `key` and `value` are the returned results.
### 2. Graph Execution Engine
The graph execution engine consists of OPs and Channels, and the connected OPs share one Channel.
......@@ -28,7 +60,7 @@ The graph execution engine consists of OPs and Channels, and the connected OPs s
</center>
### OP Design
#### <b>2.1 OP Design</b>
- The default function of a single OP is to access a single Paddle Serving Service based on the input Channel data and put the result into the output Channel.
- OP supports user customization, including preprocess, process, postprocess functions that can be inherited and implemented by the user.
......@@ -36,7 +68,7 @@ The graph execution engine consists of OPs and Channels, and the connected OPs s
- OP can obtain data from multiple different RPC requests for Auto-Batching.
- OP can be started by a thread or process.
### Channel Design
#### <b>2.2 Channel Design</b>
- Channel is the data structure for sharing data between OPs, responsible for sharing data or sharing data status information.
- Outputs from multiple OPs can be stored in the same Channel, and data from the same Channel can be used by multiple OPs.
......@@ -47,8 +79,17 @@ The graph execution engine consists of OPs and Channels, and the connected OPs s
</center>
#### <b>2.3 client type design</b>
### Extreme Case Consideration
- Prediction type (client_type) of Op has 3 types, brpc, grpc and local_predictor
- brpc: Using bRPC Client to interact with remote Serving by network, performance is better than grpc.
- grpc: Using gRPC Client to interact with remote Serving by network, cross-platform deployment supported.
- local_predictor: Load the model and predict in the local service without interacting with the network. Support multi-card deployment, and TensorRT prediction.
- Selection:
- Time cost(lower is better): local_predict < brpc <= grpc
- Microservice: Split the brpc or grpc model into independent services, simplify development and deployment complexity, and improve resource utilization
#### <b>2.4 Extreme Case Consideration</b>
- Request timeout
......@@ -65,9 +106,7 @@ The graph execution engine consists of OPs and Channels, and the connected OPs s
- For output buffer, you can use a similar process as input buffer, which adjusts the concurrency of OP3 and OP4 to control the buffer length of output buffer. (The length of the output buffer depends on the speed at which downstream OPs obtain data from the output buffer)
- The amount of data in the Channel will not exceed `worker_num` of gRPC, that is, it will not exceed the thread pool size.
## Detailed Design
### User Interface Design
## ★ Detailed Design
#### 1. General OP Definition
......@@ -79,11 +118,13 @@ def __init__(name=None,
server_endpoints=[],
fetch_list=[],
client_config=None,
client_type=None,
concurrency=1,
timeout=-1,
retry=1,
batch_size=1,
auto_batching_timeout=None)
auto_batching_timeout=None,
local_service_handler=None)
```
The meaning of each parameter is as follows:
......@@ -92,14 +133,16 @@ The meaning of each parameter is as follows:
| :-------------------: | :----------------------------------------------------------: |
| name | (str) String used to identify the OP type, which must be globally unique. |
| input_ops | (list) A list of all previous OPs of the current Op. |
| server_endpoints | (list) List of endpoints for remote Paddle Serving Service. If this parameter is not set, the OP will not access the remote Paddle Serving Service, that is, the process operation will not be performed. |
| server_endpoints | (list) List of endpoints for remote Paddle Serving Service. If this parameter is not set,it is considered as local_precditor mode, and the configuration is read from local_service_conf |
| fetch_list | (list) List of fetch variable names for remote Paddle Serving Service. |
| client_config | (str) The path of the client configuration file corresponding to the Paddle Serving Service. |
| client_type | （str)brpc, grpc or local_predictor. local_predictor does not start the Serving service, in-process prediction|
| concurrency | (int) The number of concurrent OPs. |
| timeout | (int) The timeout time of the process operation, in ms. If the value is less than zero, no timeout is considered. |
| retry | (int) Timeout number of retries. When the value is 1, no retries are made. |
| batch_size | (int) The expected batch_size of Auto-Batching, since building batches may time out, the actual batch_size may be less than the set value. |
| auto_batching_timeout | (float) Timeout for building batches of Auto-Batching (the unit is ms). |
| auto_batching_timeout | (float) Timeout for building batches of Auto-Batching (the unit is ms). When batch_size> 1, auto_batching_timeout should be set, otherwise the waiting will be blocked when the number of requests is insufficient for batch_size|
| local_service_handler | (object) local predictor handler，assigned by Op init() input parameters or created in Op init()|
#### 2. General OP Secondary Development Interface
......@@ -156,7 +199,7 @@ def init_op(self):
It should be **noted** that in the threaded version of OP, each OP will only call this function once, so the loaded resources must be thread safe.
#### 3. RequestOp Definition
#### 3. RequestOp Definition and Secondary Development Interface
RequestOp is used to process RPC data received by Pipeline Server, and the processed data will be added to the graph execution engine. Its constructor is as follows:
......@@ -164,7 +207,7 @@ RequestOp is used to process RPC data received by Pipeline Server, and the proce
def __init__(self)
```
#### 4. RequestOp Secondary Development Interface
When the default RequestOp cannot meet the parameter parsing requirements, you can customize the request parameter parsing method by rewriting the following two interfaces.
| Interface or Variable | Explain |
| :---------------------------------------: | :----------------------------------------------------------: |
......@@ -188,7 +231,7 @@ def unpack_request_package(self, request):
The return value is required to be a dictionary type.
#### 5. ResponseOp Definition
#### 4. ResponseOp Definition and Secondary Development Interface
ResponseOp is used to process the prediction results of the graph execution engine. The processed data will be used as the RPC return value of Pipeline Server. Its constructor is as follows:
......@@ -198,7 +241,7 @@ def __init__(self, input_ops)
`input_ops` is the last OP of graph execution engine. Users can construct different DAGs by setting different `input_ops` without modifying the topology of OPs.
#### 6. ResponseOp Secondary Development Interface
When the default ResponseOp cannot meet the requirements of the result return format, you can customize the return package packaging method by rewriting the following two interfaces.
| Interface or Variable | Explain |
| :------------------------------------------: | :----------------------------------------------------------: |
......@@ -237,7 +280,7 @@ def pack_response_package(self, channeldata):
return resp
```
#### 7. PipelineServer Definition
#### 5. PipelineServer Definition
The definition of PipelineServer is relatively simple, as follows:
......@@ -251,22 +294,137 @@ server.run_server()
Where `response_op` is the responseop mentioned above, PipelineServer will initialize Channels according to the topology relationship of each OP and build the calculation graph. `config_yml_path` is the configuration file of PipelineServer. The example file is as follows:
```yaml
rpc_port: 18080 # gRPC port
worker_num: 1 # gRPC thread pool size (the number of processes in the process version servicer). The default is 1
build_dag_each_worker: false # Whether to use process server or not. The default is false
http_port: 0 # HTTP service port. Do not start HTTP service when the value is less or equals 0. The default value is 0.
# gRPC port
rpc_port: 18080
# http port, do not start HTTP service when the value is less or equals 0. The default value is 0.
http_port: 18071
# gRPC thread pool size (the number of processes in the process version servicer). The default is 1
worker_num: 1
# Whether to use process server or not. The default is false
build_dag_each_worker: false
dag:
is_thread_op: true # Whether to use the thread version of OP. The default is true
client_type: brpc # Use brpc or grpc client. The default is brpc
retry: 1 # The number of times DAG executor retries after failure. The default value is 1, that is, no retrying
use_profile: false # Whether to print the log on the server side. The default is false
# Whether to use the thread version of OP. The default is true
is_thread_op: true
# The number of times DAG executor retries after failure. The default value is 1, that is, no retrying
retry: 1
# Whether to print the log on the server side. The default is false
use_profile: false
# Monitoring time interval of Tracer (in seconds). Do not start monitoring when the value is less than 1. The default value is -1
tracer:
interval_s: 600 # Monitoring time interval of Tracer (in seconds). Do not start monitoring when the value is less than 1. The default value is -1
interval_s: 600
op:
bow:
# Concurrency, when is_thread_op=True, it's thread concurrency; otherwise, it's process concurrency
concurrency: 1
# Client types, brpc, grpc and local_predictor
client_type: brpc
# Retry times, no retry by default
retry: 1
# Prediction timeout, ms
timeout: 3000
# Serving IPs
server_endpoints: ["127.0.0.1:9393"]
# Client config of bow model
client_config: "imdb_bow_client_conf/serving_client_conf.prototxt"
# Fetch list
fetch_list: ["prediction"]
# Batch size, default 1
batch_size: 1
# Batch query timeout
auto_batching_timeout: 2000
```
### 6. Special usages
#### 6.1 <b>Business custom error type</b>
Users can customize the error code according to the business, inherit ProductErrCode, and return it in the return list in Op's preprocess or postprocess. The next stage of processing will skip the post OP processing based on the custom error code.
```python
class ProductErrCode(enum.Enum):
"""
ProductErrCode is a base class for recording business error code.
product developers inherit this class and extend more error codes.
"""
pass
```
#### <b>6.2 Skip process stage</b>
The 2rd result of the result list returned by preprocess is `is_skip_process=True`, indicating whether to skip the process stage of the current OP and directly enter the postprocess processing
```python
def preprocess(self, input_dicts, data_id, log_id):
"""
In preprocess stage, assembling data for process stage. users can
override this function for model feed features.
Args:
input_dicts: input data to be preprocessed
data_id: inner unique id
log_id: global unique id for RTT
Return:
input_dict: data for process stage
is_skip_process: skip process stage or not, False default
prod_errcode: None default, otherwise, product errores occured.
It is handled in the same way as exception.
prod_errinfo: "" default
"""
# multiple previous Op
if len(input_dicts) != 1:
_LOGGER.critical(
self._log(
"Failed to run preprocess: this Op has multiple previous "
"inputs. Please override this func."))
os._exit(-1)
(_, input_dict), = input_dicts.items()
return input_dict, False, None, ""
```
#### <b>6.3 Custom proto Request and Response</b>
When the default proto structure does not meet the business requirements, at the same time, the Request and Response message structures of the proto in the following two files remain the same.
> pipeline/gateway/proto/gateway.proto
> pipeline/proto/pipeline_service.proto
Recompile Serving Server again.
#### <b>6.4 Custom URL</b>
## Example
The grpc gateway processes post requests. The default `method` is `prediction`, for example: 127.0.0.1:8080/ocr/prediction. Users can customize the name and method, and can seamlessly switch services with existing URLs.
```proto
service PipelineService {
rpc inference(Request) returns (Response) {
option (google.api.http) = {
post : "/{name=*}/{method=*}"
body : "*"
};
}
};
```
***
## ★ Classic examples
Here, we build a simple imdb model enable example to show how to use Pipeline Serving. The relevant code can be found in the `python/examples/pipeline/imdb_model_ensemble` folder. The Server-side structure in the example is shown in the following figure:
......@@ -277,7 +435,7 @@ Here, we build a simple imdb model enable example to show how to use Pipeline Se
</center>
### Get the model file and start the Paddle Serving Service
### 1. Get the model file and start the Paddle Serving Service
```shell
cd python/examples/pipeline/imdb_model_ensemble
......@@ -288,7 +446,83 @@ python -m paddle_serving_server.serve --model imdb_bow_model --port 9393 &> bow.
PipelineServing also supports local automatic startup of PaddleServingService. Please refer to the example `python/examples/pipeline/ocr`.
### Start PipelineServer
### 2. Create config.yaml
Because there is a lot of configuration information in config.yaml,, only part of the OP configuration is shown here. For full information, please refer to `python/examples/pipeline/imdb_model_ensemble/config.yaml`
```yaml
op:
bow:
# Concurrency, when is_thread_op=True, it's thread concurrency; otherwise, it's process concurrency
concurrency: 1
# Client types, brpc, grpc and local_predictor
client_type: brpc
# Retry times, no retry by default
retry: 1
# Predcition timeout, ms
timeout: 3000
# Serving IPs
server_endpoints: ["127.0.0.1:9393"]
# Client config of bow model
client_config: "imdb_bow_client_conf/serving_client_conf.prototxt"
# Fetch list
fetch_list: ["prediction"]
# Batch request size, default 1
batch_size: 1
# Batch query timeout
auto_batching_timeout: 2000
cnn:
# Concurrency
concurrency: 1
# Client types, brpc, grpc and local_predictor
client_type: brpc
# Retry times, no retry by default
retry: 1
# Predcition timeout, ms
timeout: 3000
# Serving IPs
server_endpoints: ["127.0.0.1:9292"]
# Client config of cnn model
client_config: "imdb_cnn_client_conf/serving_client_conf.prototxt"
# Fetch list
fetch_list: ["prediction"]
# Batch request size, default 1
batch_size: 1
# Batch query timeout
auto_batching_timeout: 2000
combine:
# Concurrency
concurrency: 1
#R etry times, no retry by default
retry: 1
# Predcition timeout, ms
timeout: 3000
# Batch request size, default 1
batch_size: 1
# Batch query timeout, ms
auto_batching_timeout: 2000
### 3. Start PipelineServer
Run the following code
......@@ -359,7 +593,7 @@ server.prepare_server('config.yml')
server.run_server()
```
### Perform prediction through PipelineClient
### 4. Perform prediction through PipelineClient
```python
from paddle_serving_client.pipeline import PipelineClient
......@@ -385,13 +619,16 @@ for f in futures:
exit(1)
```
***
## ★ Performance analysis
## How to optimize with the timeline tool
### 1. How to optimize with the timeline tool
In order to better optimize the performance, PipelineServing provides a timeline tool to monitor the time of each stage of the whole service.
### Output profile information on server side
### 2. Output profile information on server side
The server is controlled by the `use_profile` field in yaml:
......@@ -418,8 +655,29 @@ if __name__ == "__main__":
Specific operation: open Chrome browser, input in the address bar `chrome://tracing/` , jump to the tracing page, click the load button, open the saved `trace` file, and then visualize the time information of each stage of the prediction service.
### Output profile information on client side
### 3. Output profile information on client side
The profile function can be enabled by setting `profile=True` in the `predict` interface on the client side.
After the function is enabled, the client will print the log information corresponding to the prediction to the standard output during the prediction process, and the subsequent analysis and processing are the same as that of the server.
### 4. Analytical methods
```
cost of one single OP：
op_cost = process(pre + mid + post)
OP Concurrency:
op_concurrency = op_cost(s) * qps_expected
Service throughput：
service_throughput = 1 / slowest_op_cost * op_concurrency
Service average cost：
service_avg_cost = ∑op_concurrency in critical Path
Channel accumulations：
channel_acc_size = QPS(down - up) * time
Average cost of batch predictor：
avg_batch_cost = (N * pre + mid + post) / N
```
doc/PIPELINE_SERVING_CN.md
浏览文件 @ af65b28f
......@@ -7,15 +7,47 @@ Paddle Serving 通常用于单模型的一键部署，但端到端的深度学
Paddle Serving 提供了用户友好的多模型组合服务编程框架，Pipeline Serving，旨在降低编程门槛，提高资源使用率（尤其是GPU设备），提升整体的预估效率。
## 整体架构设计
## ★ 整体架构设计
Server端基于 gRPC 和图执行引擎构建，两者的关系如下图所示。
Server端基于<b>RPC服务层</b>和<b>图执行引擎</b>构建，两者的关系如下图所示。
<center>
<img src='pipeline_serving-image1.png' height = "250" align="middle"/>
</center>
### 图执行引擎
</n>
### 1. RPC服务层
为满足用户不同的使用需求，RPC服务层同时启动1个Web服务器和1个RPC服务器，可同时处理RESTful API、gRPC 2种类型请求。gPRC gateway接收RESTful API请求通过反向代理服务器将请求转发给gRPC Service；gRPC请求由gRPC service接收，所以，2种类型的请求统一由gRPC Service处理，确保处理逻辑一致。
#### <b>1.1 proto的输入输出结构</b>
gRPC服务和gRPC gateway服务统一用service.proto生成。
```proto
message Request {
repeated string key = 1;
repeated string value = 2;
optional string name = 3;
optional string method = 4;
optional int64 logid = 5;
optional string clientip = 6;
};
message Response {
optional int32 err_no = 1;
optional string err_msg = 2;
repeated string key = 3;
repeated string value = 4;
};
```
Request中`key`与`value`是配对的string数组。 `name`与`method`对应RESTful API的URL://{ip}:{port}/{name}/{method}。`logid`和`clientip`便于用户串联服务级请求和自定义策略。
Response中`err_no`和`err_msg`表达处理结果的正确性和错误信息，`key`和`value`为返回结果。
### 2. 图执行引擎
图执行引擎由 OP 和 Channel 构成，相连接的 OP 之间会共享一个 Channel。
......@@ -29,7 +61,7 @@ Server端基于 gRPC 和图执行引擎构建，两者的关系如下图所示
</center>
### OP的设计
#### <b>2.1 OP的设计</b>
- 单个 OP 默认的功能是根据输入的 Channel 数据，访问一个 Paddle Serving 的单模型服务，并将结果存在输出的 Channel
- 单个 OP 可以支持用户自定义，包括 preprocess，process，postprocess 三个函数都可以由用户继承和实现
......@@ -37,7 +69,7 @@ Server端基于 gRPC 和图执行引擎构建，两者的关系如下图所示
- 单个 OP 可以获取多个不同 RPC 请求的数据，以实现 Auto-Batching
- OP 可以由线程或进程启动
### Channel的设计
#### <b>2.2 Channel的设计</b>
- Channel 是 OP 之间共享数据的数据结构，负责共享数据或者共享数据状态信息
- Channel 可以支持多个OP的输出存储在同一个 Channel，同一个 Channel 中的数据可以被多个 OP 使用
......@@ -47,8 +79,18 @@ Server端基于 gRPC 和图执行引擎构建，两者的关系如下图所示
<img src='pipeline_serving-image3.png' height = "500" align="middle"/>
</center>
#### <b>2.3 预测类型的设计</b>
- OP的预测类型(client_type)有3种类型，brpc、grpc和local_predictor
- brpc: 使用bRPC Client与远端的Serving服务网络交互，性能优于grpc
- grpc: 使用gRPC Client与远端的Serving服务网络交互，支持跨平台部署
- local_predictor: 本地服务内加载模型并完成预测，不需要与网络交互。支持多卡部署，和TensorRT高性能预测。
- 选型:
- 延时(越少越好): local_predict < brpc <= grpc
- 微服务: brpc或grpc模型分拆成独立服务，简化开发和部署复杂度，提升资源利用率
### 极端情况的考虑
#### <b>2.4 极端情况的考虑</b>
- 请求超时的处理
......@@ -65,9 +107,11 @@ Server端基于 gRPC 和图执行引擎构建，两者的关系如下图所示
- 对于 output buffer，可以采用和 input buffer 类似的处理方法，即调整 OP3 和 OP4 的并发数，使得 output buffer 的缓冲长度得到控制（output buffer 的长度取决于下游 OP 从 output buffer 获取数据的速度）
- 同时 Channel 中数据量不会超过 gRPC 的 `worker_num`，即线程池大小
### 用户接口设计
***
## ★ 详细设计
#### 1. 普通 OP 定义
### 1. 普通 OP 定义
普通 OP 作为图执行引擎中的基本单元，其构造函数如下：
......@@ -77,11 +121,13 @@ def __init__(name=None,
server_endpoints=[],
fetch_list=[],
client_config=None,
client_type=None,
concurrency=1,
timeout=-1,
retry=1,
batch_size=1,
auto_batching_timeout=None)
auto_batching_timeout=None,
local_service_handler=None)
```
各参数含义如下
......@@ -90,17 +136,21 @@ def __init__(name=None,
| :-------------------: | :----------------------------------------------------------: |
| name | （str）用于标识 OP 类型的字符串，该字段必须全局唯一。 |
| input_ops | （list）当前 OP 的所有前继 OP 的列表。 |
| server_endpoints | （list）远程 Paddle Serving Service 的 endpoints 列表。如果不设置该参数，则不访问远程 Paddle Serving Service，即 不会执行 process 操作。 |
| server_endpoints | （list）远程 Paddle Serving Service 的 endpoints 列表。如果不设置该参数，认为是local_precditor模式，从local_service_conf中读取配置。 |
| fetch_list | （list）远程 Paddle Serving Service 的 fetch 列表。 |
| client_config | （str）Paddle Serving Service 对应的 Client 端配置文件路径。 |
| client_type | (str) 可选择brpc、grpc或local_predictor。local_predictor不启动Serving服务，进程内预测。 |
| concurrency | （int）OP 的并发数。 |
| timeout | （int）process 操作的超时时间，单位为毫秒。若该值小于零，则视作不超时。 |
| retry | （int）超时重试次数。当该值为 1 时，不进行重试。 |
| batch_size | （int）进行 Auto-Batching 的期望 batch_size 大小，由于构建 batch 可能超时，实际 batch_size 可能小于设定值。 |
| auto_batching_timeout | （float）进行 Auto-Batching 构建 batch 的超时时间，单位为毫秒。 |
| batch_size | （int）进行 Auto-Batching 的期望 batch_size 大小，由于构建 batch 可能超时，实际 batch_size 可能小于设定值，默认为 1。 |
| auto_batching_timeout | （float）进行 Auto-Batching 构建 batch 的超时时间，单位为毫秒。batch_size > 1时，要设置auto_batching_timeout，否则请求数量不足batch_size时会阻塞等待。 |
| local_service_handler | (object) local predictor handler，Op init()入参赋值 或 在Op init()中创建|
#### 2. 普通 OP二次开发接口
### 2. 普通 OP二次开发接口
OP 二次开发的目的是满足业务开发人员控制OP处理策略。
| 变量或接口 | 说明 |
| :----------------------------------------------: | :----------------------------------------------------------: |
......@@ -154,7 +204,7 @@ def init_op(self):
需要**注意**的是，在线程版 OP 中，每个 OP 只会调用一次该函数，故加载的资源必须要求是线程安全的。
#### 3. RequestOp 定义
### 3. RequestOp 定义 与 二次开发接口
RequestOp 用于处理 Pipeline Server 接收到的 RPC 数据，处理后的数据将会被加入到图执行引擎中。其构造函数如下：
......@@ -162,7 +212,7 @@ RequestOp 用于处理 Pipeline Server 接收到的 RPC 数据，处理后的数
def __init__(self)
```
#### 4. RequestOp 二次开发接口
当默认的RequestOp无法满足参数解析需求时，可通过重写下面2个接口自定义请求参数解析方法。
| 变量或接口 | 说明 |
| :---------------------------------------: | :----------------------------------------: |
......@@ -186,7 +236,7 @@ def unpack_request_package(self, request):
要求返回值是一个字典类型。
#### 5. ResponseOp 定义
#### 4. ResponseOp 定义 与 二次开发接口
ResponseOp 用于处理图执行引擎的预测结果，处理后的数据将会作为 Pipeline Server 的RPC 返回值，其构造函数如下：
......@@ -196,7 +246,7 @@ def __init__(self, input_ops)
其中，`input_ops` 是图执行引擎的最后一个 OP，用户可以通过设置不同的 `input_ops` 以在不修改 OP 的拓扑关系下构造不同的 DAG。
#### 6. ResponseOp 二次开发接口
当默认的 ResponseOp 无法满足结果返回格式要求时，可通过重写下面2个接口自定义返回包打包方法。
| 变量或接口 | 说明 |
| :------------------------------------------: | :-----------------------------------------: |
......@@ -235,7 +285,7 @@ def pack_response_package(self, channeldata):
return resp
```
#### 7. PipelineServer定义
#### 5. PipelineServer定义
PipelineServer 的定义比较简单，如下所示：
......@@ -249,22 +299,134 @@ server.run_server()
其中，`response_op` 为上面提到的 ResponseOp，PipelineServer 将会根据各个 OP 的拓扑关系初始化 Channel 并构建计算图。`config_yml_path` 为 PipelineServer 的配置文件，示例文件如下：
```yaml
rpc_port: 18080 # gRPC端口号
worker_num: 1 # gRPC线程池大小（进程版 Servicer 中为进程数），默认为 1
build_dag_each_worker: false # 是否使用进程版 Servicer，默认为 false
http_port: 0 # HTTP 服务的端口号，若该值小于或等于 0 则不开启 HTTP 服务，默认为 0
# gRPC端口号
rpc_port: 18080
# http端口号，若该值小于或等于 0 则不开启 HTTP 服务，默认为 0
http_port: 18071
# #worker_num, 最大并发数。当build_dag_each_worker=True时, 框架会创建worker_num个进程，每个进程内构建grpcSever和DAG
worker_num: 1
# 是否使用进程版 Servicer，默认为 false
build_dag_each_worker: false
dag:
is_thread_op: true # 是否使用线程版Op，默认为 true
client_type: brpc # 使用 brpc 或 grpc client，默认为 brpc
retry: 1 # DAG Executor 在失败后重试次数，默认为 1，即不重试
use_profile: false # 是否在 Server 端打印日志，默认为 false
# op资源类型, True, 为线程模型；False，为进程模型，默认为 True
is_thread_op: true
# DAG Executor 在失败后重试次数，默认为 1，即不重试
retry: 1
# 是否在 Server 端打印日志，默认为 false
use_profile: false
# 跟踪框架吞吐，每个OP和channel的工作情况。无tracer时不生成数据
tracer:
interval_s: 600 # Tracer 监控的时间间隔，单位为秒。当该值小于 1 时不启动监控，默认为 -1
interval_s: 600 # 监控的时间间隔，单位为秒。当该值小于 1 时不启动监控，默认为 -1
op:
bow:
# 并发数，is_thread_op=True时，为线程并发；否则为进程并发
concurrency: 1
# client连接类型，brpc
client_type: brpc
# Serving交互重试次数，默认不重试
retry: 1
# Serving交互超时时间, 单位ms
timeout: 3000
# Serving IPs
server_endpoints: ["127.0.0.1:9393"]
# bow模型client端配置
client_config: "imdb_bow_client_conf/serving_client_conf.prototxt"
# Fetch结果列表，以client_config中fetch_var的alias_name为准
fetch_list: ["prediction"]
# 批量查询Serving的数量, 默认1。batch_size>1要设置 auto_batching_timeout，否则不足batch_size时会阻塞
batch_size: 1
# 批量查询超时，与batch_size配合使用
auto_batching_timeout: 2000
```
### 6. 特殊用法
#### 6.1 <b>业务自定义错误类型</b>
用户可根据业务场景自定义错误码，继承ProductErrCode，在Op的preprocess或postprocess中返回列表中返回，下一阶段处理会根据自定义错误码跳过后置OP处理。
```python
class ProductErrCode(enum.Enum):
"""
ProductErrCode is a base class for recording business error code.
product developers inherit this class and extend more error codes.
"""
pass
```
#### <b>6.2 跳过OP process阶段</b>
preprocess返回结果列表的第二个结果是`is_skip_process=True`表示是否跳过当前OP的process阶段，直接进入postprocess处理
```python
def preprocess(self, input_dicts, data_id, log_id):
"""
In preprocess stage, assembling data for process stage. users can
override this function for model feed features.
Args:
input_dicts: input data to be preprocessed
data_id: inner unique id
log_id: global unique id for RTT
Return:
input_dict: data for process stage
is_skip_process: skip process stage or not, False default
prod_errcode: None default, otherwise, product errores occured.
It is handled in the same way as exception.
prod_errinfo: "" default
"""
# multiple previous Op
if len(input_dicts) != 1:
_LOGGER.critical(
self._log(
"Failed to run preprocess: this Op has multiple previous "
"inputs. Please override this func."))
os._exit(-1)
(_, input_dict), = input_dicts.items()
return input_dict, False, None, ""
```
#### <b>6.3 自定义proto Request 和 Response结构</b>
## 例子
当默认proto结构不满足业务需求时，同时下面2个文件的proto的Request和Response message结构，保持一致。
> pipeline/gateway/proto/gateway.proto
> pipeline/proto/pipeline_service.proto
再重新编译Serving Server。
#### <b>6.4 自定义URL</b>
grpc gateway处理post请求，默认`method`是`prediction`，例如:127.0.0.1:8080/ocr/prediction。用户可自定义name和method，对于已有url的服务可无缝切换
```proto
service PipelineService {
rpc inference(Request) returns (Response) {
option (google.api.http) = {
post : "/{name=*}/{method=*}"
body : "*"
};
}
};
```
***
## ★ 典型示例
这里通过搭建简单的 imdb model ensemble 例子来展示如何使用 Pipeline Serving，相关代码在 `python/examples/pipeline/imdb_model_ensemble` 文件夹下可以找到，例子中的 Server 端结构如下图所示：
......@@ -275,7 +437,7 @@ dag:
</center>
### 获取模型文件并启动 Paddle Serving Service
### 1. 获取模型文件并启动 Paddle Serving Service
```shell
cd python/examples/pipeline/imdb_model_ensemble
......@@ -286,9 +448,84 @@ python -m paddle_serving_server.serve --model imdb_bow_model --port 9393 &> bow.
PipelineServing 也支持本地自动启动 PaddleServingService，请参考 `python/examples/pipeline/ocr` 下的例子。
### 启动 PipelineServer
### 2. 创建config.yaml
由于config.yaml配置信息量很多，这里仅展示OP部分配置，全量信息参考`python/examples/pipeline/imdb_model_ensemble/config.yaml`
```yaml
op:
bow:
# 并发数，is_thread_op=True时，为线程并发；否则为进程并发
concurrency: 1
# client连接类型，brpc
client_type: brpc
# Serving交互重试次数，默认不重试
retry: 1
# Serving交互超时时间, 单位ms
timeout: 3000
# Serving IPs
server_endpoints: ["127.0.0.1:9393"]
# bow模型client端配置
client_config: "imdb_bow_client_conf/serving_client_conf.prototxt"
# Fetch结果列表，以client_config中fetch_var的alias_name为准
fetch_list: ["prediction"]
# 批量查询Serving的数量, 默认1。batch_size>1要设置auto_batching_timeout，否则不足batch_size时会阻塞
batch_size: 1
# 批量查询超时，与batch_size配合使用
auto_batching_timeout: 2000
cnn:
# 并发数，is_thread_op=True时，为线程并发；否则为进程并发
concurrency: 1
# client连接类型，brpc
client_type: brpc
# Serving交互重试次数，默认不重试
retry: 1
# 预测超时时间, 单位ms
timeout: 3000
运行下面代码
# Serving IPs
server_endpoints: ["127.0.0.1:9292"]
# cnn模型client端配置
client_config: "imdb_cnn_client_conf/serving_client_conf.prototxt"
# Fetch结果列表，以client_config中fetch_var的alias_name为准
fetch_list: ["prediction"]
# 批量查询Serving的数量, 默认1。
batch_size: 1
# 批量查询超时，与batch_size配合使用
auto_batching_timeout: 2000
combine:
# 并发数，is_thread_op=True时，为线程并发；否则为进程并发
concurrency: 1
# Serving交互重试次数，默认不重试
retry: 1
# 预测超时时间, 单位ms
timeout: 3000
# 批量查询Serving的数量, 默认1。
batch_size: 1
# 批量查询超时，与batch_size配合使用
auto_batching_timeout: 2000
```
### 3. 启动 PipelineServer
代码示例中，重点留意3个自定义Op的proprocess、postprocess处理，以及Combin Op初始化列表input_ops=[bow_op, cnn_op]，设置Combin Op的前置OP列表。
```python
from paddle_serving_server.pipeline import Op, RequestOp, ResponseOp
......@@ -356,7 +593,7 @@ server.prepare_server('config.yml')
server.run_server()
```
### 通过 PipelineClient 执行预测
### 4. 通过 PipelineClient 执行预测
```python
from paddle_serving_client.pipeline import PipelineClient
......@@ -382,13 +619,16 @@ for f in futures:
exit(1)
```
***
## ★ 性能分析
## 如何通过 Timeline 工具进行优化
### 1. 如何通过 Timeline 工具进行优化
为了更好地对性能进行优化，PipelineServing 提供了 Timeline 工具，对整个服务的各个阶段时间进行打点。
### 在 Server 端输出 Profile 信息
### 2. 在 Server 端输出 Profile 信息
Server 端用 yaml 中的 `use_profile` 字段进行控制：
......@@ -415,8 +655,29 @@ if __name__ == "__main__":
具体操作：打开 chrome 浏览器，在地址栏输入 `chrome://tracing/` ，跳转至 tracing 页面，点击 load 按钮，打开保存的 `trace` 文件，即可将预测服务的各阶段时间信息可视化。
### 在 Client 端输出 Profile 信息
### 3. 在 Client 端输出 Profile 信息
Client 端在 `predict` 接口设置 `profile=True`，即可开启 Profile 功能。
开启该功能后，Client 端在预测的过程中会将该次预测对应的日志信息打印到标准输出，后续分析处理同 Server。
### 4. 分析方法
```
单OP耗时：
op_cost = process(pre + mid + post)
OP期望并发数：
op_concurrency = 单OP耗时(s) * 期望QPS
服务吞吐量：
service_throughput = 1 / 最慢OP的耗时 * 并发数
服务平响：
service_avg_cost = ∑op_concurrency 【关键路径】
Channel堆积：
channel_acc_size = QPS(down - up) * time
批量预测平均耗时：
avg_batch_cost = (N * pre + mid + post) / N
```
doc/RUN_IN_DOCKER.md
浏览文件 @ af65b28f
......@@ -2,6 +2,8 @@
([简体中文](RUN_IN_DOCKER_CN.md)|English)
One of the biggest benefits of Docker is portability, which can be deployed on multiple operating systems and mainstream cloud computing platforms. The Paddle Serving Docker image can be deployed on Linux, Mac and Windows platforms.
## Requirements
Docker (GPU version requires nvidia-docker to be installed on the GPU machine)
......@@ -30,63 +32,9 @@ The `-p` option is to map the `9292` port of the container to the `9292` port of
### Install PaddleServing
In order to make the image smaller, the PaddleServing package is not installed in the image. You can run the following command to install it:
```bash
pip install paddle-serving-server
```
You may need to use a domestic mirror source (in China, you can use the Tsinghua mirror source of the following example) to speed up the download:
```shell
pip install paddle-serving-server -i https://pypi.tuna.tsinghua.edu.cn/simple
```
### Test example
Get the trained Boston house price prediction model by the following command:
```bash
wget --no-check-certificate https://paddle-serving.bj.bcebos.com/uci_housing.tar.gz
tar -xzf uci_housing.tar.gz
```
- Test HTTP service
Running on the Server side (inside the container):
```bash
python -m paddle_serving_server.serve --model uci_housing_model --thread 10 --port 9292 --name uci >std.log 2>err.log &
```
Running on the Client side (inside or outside the container):
```bash
curl -H "Content-Type:application/json" -X POST -d '{"feed":{"x": [0.0137, -0.1136, 0.2553, -0.0692, 0.0582, -0.0727, -0.1583, -0.0584, 0.6283, 0.4919, 0.1856, 0.0795, -0.0332]}, "fetch":["price"]}' http://127.0.0.1:9292/uci/prediction
```
- Test RPC service
Running on the Server side (inside the container):
```bash
python -m paddle_serving_server.serve --model uci_housing_model --thread 10 --port 9292 >std.log 2>err.log &
```
Running following Python code on the Client side (inside or outside the container, The `paddle-serving-client` package needs to be installed):
```bash
from paddle_serving_client import Client
client = Client()
client.load_client_config("uci_housing_client/serving_client_conf.prototxt")
client.connect(["127.0.0.1:9292"])
data = [0.0137, -0.1136, 0.2553, -0.0692, 0.0582, -0.0727,
-0.1583, -0.0584, 0.6283, 0.4919, 0.1856, 0.0795, -0.0332]
fetch_map = client.predict(feed={"x": data}, fetch=["price"])
print(fetch_map)
```
The mirror comes with `paddle_serving_server`, `paddle_serving_client`, and `paddle_serving_app` corresponding to the mirror tag version. If users don’t need to change the version, they can use it directly, which is suitable for environments without extranet services.
If you need to change the version, please refer to the instructions on the homepage to download the pip package of the corresponding version.
## GPU
......@@ -98,7 +46,7 @@ The GPU version is basically the same as the CPU version, with only some differe
Refer to [this document](DOCKER_IMAGES.md) for a docker image, the following is an example of an `cuda9.0-cudnn7` image:
```shell
nvidia-docker pull hub.baidubce.com/paddlepaddle/serving:latest-cuda9.0-cudnn7
docker pull hub.baidubce.com/paddlepaddle/serving:latest-cuda9.0-cudnn7
```
### Create container
......@@ -108,77 +56,21 @@ nvidia-docker run -p 9292:9292 --name test -dit hub.baidubce.com/paddlepaddle/se
nvidia-docker exec -it test bash
```
The `-p` option is to map the `9292` port of the container to the `9292` port of the host.
### Install PaddleServing
In order to make the image smaller, the PaddleServing package is not installed in the image. You can run the following command to install it:
```bash
pip install paddle-serving-server-gpu
```
You may need to use a domestic mirror source (in China, you can use the Tsinghua mirror source of the following example) to speed up the download:
```shell
pip install paddle-serving-server-gpu -i https://pypi.tuna.tsinghua.edu.cn/simple
```
### Test example
When running the GPU Server, you need to set the GPUs used by the prediction service through the `--gpu_ids` option, and the CPU is used by default. An error will be reported when the value of `--gpu_ids` exceeds the environment variable `CUDA_VISIBLE_DEVICES`. The following example specifies to use a GPU with index 0:
```shell
export CUDA_VISIBLE_DEVICES=0,1
python -m paddle_serving_server_gpu.serve --model uci_housing_model --port 9292 --gpu_ids 0
```
Get the trained Boston house price prediction model by the following command:
or
```bash
wget --no-check-certificate https://paddle-serving.bj.bcebos.com/uci_housing.tar.gz
tar -xzf uci_housing.tar.gz
docker run --gpus all -p 9292:9292 --name test -dit hub.baidubce.com/paddlepaddle/serving:latest-cuda9.0-cudnn7
docker exec -it test bash
```
- Test HTTP service
Running on the Server side (inside the container):
```bash
python -m paddle_serving_server_gpu.serve --model uci_housing_model --thread 10 --port 9292 --name uci --gpu_ids 0
```
Running on the Client side (inside or outside the container):
```bash
curl -H "Content-Type:application/json" -X POST -d '{"feed":{"x": [0.0137, -0.1136, 0.2553, -0.0692, 0.0582, -0.0727, -0.1583, -0.0584, 0.6283, 0.4919, 0.1856, 0.0795, -0.0332]}, "fetch":["price"]}' http://127.0.0.1:9292/uci/prediction
```
- Test RPC service
Running on the Server side (inside the container):
```bash
python -m paddle_serving_server_gpu.serve --model uci_housing_model --thread 10 --port 9292 --gpu_ids 0
```
Running following Python code on the Client side (inside or outside the container, The `paddle-serving-client` package needs to be installed):
```bash
from paddle_serving_client import Client
client = Client()
client.load_client_config("uci_housing_client/serving_client_conf.prototxt")
client.connect(["127.0.0.1:9292"])
data = [0.0137, -0.1136, 0.2553, -0.0692, 0.0582, -0.0727,
-0.1583, -0.0584, 0.6283, 0.4919, 0.1856, 0.0795, -0.0332]
fetch_map = client.predict(feed={"x": data}, fetch=["price"])
print(fetch_map)
```
The `-p` option is to map the `9292` port of the container to the `9292` port of the host.
### Install PaddleServing
The mirror comes with `paddle_serving_server_gpu`, `paddle_serving_client`, and `paddle_serving_app` corresponding to the mirror tag version. If users don’t need to change the version, they can use it directly, which is suitable for environments without extranet services.
If you need to change the version, please refer to the instructions on the homepage to download the pip package of the corresponding version.
## Attention
## Precautious
Runtime images cannot be used for compilation. If you want to compile from source, refer to [COMPILE](COMPILE.md).
doc/RUN_IN_DOCKER_CN.md
浏览文件 @ af65b28f
......@@ -2,6 +2,8 @@
(简体中文|[English](RUN_IN_DOCKER.md))
Docker最大的好处之一就是可移植性，可在多种操作系统和主流的云计算平台部署。使用Paddle Serving Docker镜像可在Linux、Mac和Windows平台部署。
## 环境要求
Docker（GPU版本需要在GPU机器上安装nvidia-docker）
......@@ -18,7 +20,6 @@ Docker（GPU版本需要在GPU机器上安装nvidia-docker）
docker pull hub.baidubce.com/paddlepaddle/serving:latest
```
### 创建容器并进入
```bash
......@@ -30,74 +31,11 @@ docker exec -it test bash
### 安装PaddleServing
为了减小镜像的体积，镜像中没有安装Serving包，要执行下面命令进行安装。
```bash
pip install paddle-serving-server
```
您可能需要使用国内镜像源（例如清华源）来加速下载。
```shell
pip install paddle-serving-server -i https://pypi.tuna.tsinghua.edu.cn/simple
```
### 测试example
通过下面命令获取训练好的Boston房价预估模型：
```bash
wget --no-check-certificate https://paddle-serving.bj.bcebos.com/uci_housing.tar.gz
tar -xzf uci_housing.tar.gz
```
- 测试HTTP服务
镜像里自带对应镜像tag版本的`paddle_serving_server`，`paddle_serving_client`，`paddle_serving_app`，如果用户不需要更改版本，可以直接使用，适用于没有外网服务的环境。
在Server端（容器内）运行：
如果需要更换版本，请参照首页的指导，下载对应版本的pip包。
```bash
python -m paddle_serving_server.serve --model uci_housing_model --thread 10 --port 9292 --name uci >std.log 2>err.log &
```
在Client端（容器内或容器外）运行：
```bash
curl -H "Content-Type:application/json" -X POST -d '{"feed":{"x": [0.0137, -0.1136, 0.2553, -0.0692, 0.0582, -0.0727, -0.1583, -0.0584, 0.6283, 0.4919, 0.1856, 0.0795, -0.0332]}, "fetch":["price"]}' http://127.0.0.1:9292/uci/prediction
```
- 测试RPC服务
在Server端（容器内）运行：
```bash
python -m paddle_serving_server.serve --model uci_housing_model --thread 10 --port 9292 >std.log 2>err.log &
```
在Client端（容器内或容器外，需要安装`paddle-serving-client`包）运行下面Python代码：
```python
from paddle_serving_client import Client
client = Client()
client.load_client_config("uci_housing_client/serving_client_conf.prototxt")
client.connect(["127.0.0.1:9292"])
data = [0.0137, -0.1136, 0.2553, -0.0692, 0.0582, -0.0727,
-0.1583, -0.0584, 0.6283, 0.4919, 0.1856, 0.0795, -0.0332]
fetch_map = client.predict(feed={"x": data}, fetch=["price"])
print(fetch_map)
```
## GPU版本
GPU版本与CPU版本基本一致，只有部分接口命名的差别（GPU版本需要在GPU机器上安装nvidia-docker）。
### 获取镜像
参考[该文档](DOCKER_IMAGES_CN.md)获取镜像，这里以 `cuda9.0-cudnn7` 的镜像为例：
```shell
nvidia-docker pull hub.baidubce.com/paddlepaddle/serving:latest-cuda9.0-cudnn7
```
## GPU 版本
### 创建容器并进入
......@@ -105,74 +43,19 @@ nvidia-docker pull hub.baidubce.com/paddlepaddle/serving:latest-cuda9.0-cudnn7
nvidia-docker run -p 9292:9292 --name test -dit hub.baidubce.com/paddlepaddle/serving:latest-cuda9.0-cudnn7
nvidia-docker exec -it test bash
```
`-p`选项是为了将容器的`9292`端口映射到宿主机的`9292`端口。
### 安装PaddleServing
为了减小镜像的体积，镜像中没有安装Serving包，要执行下面命令进行安装。
```bash
pip install paddle-serving-server-gpu
```
您可能需要使用国内镜像源（例如清华源）来加速下载。
```shell
pip install paddle-serving-server-gpu -i https://pypi.tuna.tsinghua.edu.cn/simple
```
### 测试example
在运行GPU版Server时需要通过`--gpu_ids`选项设置预测服务使用的GPU，缺省状态默认使用CPU。当设置的`--gpu_ids`超出环境变量`CUDA_VISIBLE_DEVICES`时会报错。下面的示例为指定使用索引为0的GPU：
```shell
export CUDA_VISIBLE_DEVICES=0,1
python -m paddle_serving_server_gpu.serve --model uci_housing_model --port 9292 --gpu_ids 0
```
通过下面命令获取训练好的Boston房价预估模型：
或者
```bash
wget --no-check-certificate https://paddle-serving.bj.bcebos.com/uci_housing.tar.gz
tar -xzf uci_housing.tar.gz
docker run --gpus all -p 9292:9292 --name test -dit hub.baidubce.com/paddlepaddle/serving:latest-cuda9.0-cudnn7
docker exec -it test bash
```
- 测试HTTP服务
在Server端（容器内）运行：
```bash
python -m paddle_serving_server_gpu.serve --model uci_housing_model --thread 10 --port 9292 --name uci --gpu_ids 0
```
在Client端（容器内或容器外）运行：
```bash
curl -H "Content-Type:application/json" -X POST -d '{"feed":{"x": [0.0137, -0.1136, 0.2553, -0.0692, 0.0582, -0.0727, -0.1583, -0.0584, 0.6283, 0.4919, 0.1856, 0.0795, -0.0332]}, "fetch":["price"]}' http://127.0.0.1:9292/uci/prediction
```
- 测试RPC服务
在Server端（容器内）运行：
```bash
python -m paddle_serving_server_gpu.serve --model uci_housing_model --thread 10 --port 9292 --gpu_ids 0
```
`-p`选项是为了将容器的`9292`端口映射到宿主机的`9292`端口。
在Client端（容器内或容器外，需要安装`paddle-serving-client`包）运行下面Python代码：
### 安装PaddleServing
```bash
from paddle_serving_client import Client
镜像里自带对应镜像tag版本的`paddle_serving_server_gpu`，`paddle_serving_client`，`paddle_serving_app`，如果用户不需要更改版本，可以直接使用，适用于没有外网服务的环境。
client = Client()
client.load_client_config("uci_housing_client/serving_client_conf.prototxt")
client.connect(["127.0.0.1:9292"])
data = [0.0137, -0.1136, 0.2553, -0.0692, 0.0582, -0.0727,
-0.1583, -0.0584, 0.6283, 0.4919, 0.1856, 0.0795, -0.0332]
fetch_map = client.predict(feed={"x": data}, fetch=["price"])
print(fetch_map)
```
如果需要更换版本，请参照首页的指导，下载对应版本的pip包。
## 注意事项
......
doc/WINDOWS_TUTORIAL.md 0 → 100644
浏览文件 @ af65b28f
## Paddle Serving for Windows Users
(English|[简体中文](./WINDOWS_TUTORIAL_CN.md))
### Summary
This document guides users how to build Paddle Serving service on the Windows platform. Due to the limited support of third-party libraries, the Windows platform currently only supports the use of web services to build local predictor prediction services. If you want to experience all the services, you need to use Docker for Windows to simulate the operating environment of Linux.
### Running Paddle Serving on Native Windows System
**Configure Python environment variables to PATH**: **We only support Python 3.5+ on Native Windows System.**. First, you need to add the directory where the Python executable program is located to the PATH. Usually in **System Properties/My Computer Properties**-**Advanced**-**Environment Variables**, click Path and add the path at the beginning. For example, `C:\Users\$USER\AppData\Local\Programs\Python\Python36`, and finally click **OK** continuously. If you enter python on Powershell, you can enter the python interactive interface, indicating that the environment variable configuration is successful.
**Install wget**: Because all the downloads in the tutorial and the built-in model download function in `paddle_serving_app` all use the wget tool, download the binary package at the [link](http://gnuwin32.sourceforge.net/packages/wget.htm), unzip and copy it to `C:\Windows\System32`, if there is a security prompt, you need to pass it.
**Install Git**: For details, see [Git official website](https://git-scm.com/downloads)
**Install the necessary C++ library (optional)**: Some users may encounter the problem that the dll cannot be linked during the `import paddle` stage. It is recommended to [Install Visual Studio Community Edition](https://visualstudio.microsoft.com/), and install the relevant components of C++.
**Install Paddle and Serving**: In Powershell, execute
```
python -m pip install -U paddle_serving_server paddle_serving_client paddle_serving_app paddlepaddle`
```
for GPU users,
```
python -m pip install -U paddle_serving_server_gpu paddle_serving_client paddle_serving_app paddlepaddle-gpu
```
**Git clone Serving Project:**
```
git clone https://github.com/paddlepaddle/Serving
pip install -r python/requirements_win.txt
```
**Run OCR example**:
```
cd Serving/python/example/ocr
python -m paddle_serving_app.package --get_model ocr_rec
tar -xzvf ocr_rec.tar.gz
python -m paddle_serving_app.package --get_model ocr_det
tar -xzvf ocr_det.tar.gz
python ocr_debugger_server.py cpu &
python ocr_web_client.py
```
### Create a new Paddle Serving Web Service on Windows
Currently Windows supports the Local Predictor of the Web Service framework. The server code framework is as follows
```
# filename:your_webservice.py
from paddle_serving_server.web_service import WebService
# If it is the GPU version, please use from paddle_serving_server_gpu.web_service import WebService
class YourWebService(WebService):
def preprocess(self, feed=[], fetch=[]):
#Implement pre-processing here
#feed_dict is key: var names, value: numpy array input
#fetch_names is a list of fetch variable names
The meaning of #is_batch is whether the numpy array in the value of feed_dict contains the batch dimension
return feed_dict, fetch_names, is_batch
def postprocess(self, feed={}, fetch=[], fetch_map=None):
#fetch map is the returned dictionary after prediction, the key is the fetch names given when the process returns, and the value is the var specific value corresponding to the fetch names
#After processing here, the result needs to be converted into a dictionary again, and the type of values should be a list, so that it can be serialized in JSON to facilitate web return
return response
your_service = YourService(name="XXX")
your_service.load_model_config("your_model_path")
your_service.prepare_server(workdir="workdir", port=9292)
# If you are a GPU user, you can refer to the python example under python/examples/ocr
your_service.run_debugger_service()
# Windows platform cannot use run_rpc_service() interface
your_service.run_web_service()
```
Client code example
```
# filename:your_client.py
import requests
import json
import base64
import os, sys
import time
import cv2 # If you need to upload pictures
# Used for image reading, the principle is to use base64 encoding file content
def cv2_to_base64(image):
return base64.b64encode(image).decode(
'utf8') #data.tostring()).decode('utf8')
headers = {"Content-type": "application/json"}
url = "http://127.0.0.1:9292/XXX/prediction" # XXX depends on the initial name parameter of the server YourService
r = requests.post(url=url, headers=headers, data=json.dumps(data))
print(r.json())
```
The user only needs to follow the above instructions and implement the relevant content in the corresponding function. For more information, please refer to [How to develop a new Web Service? ](./NEW_WEB_SERVICE.md)
Execute after development
```
python your_webservice.py &
python your_client.py
```
Because the port needs to be occupied, there may be a security prompt during the startup process. Please click through and an IP address will be generated. It should be noted that when the Windows platform starts the service, the local IP address may not be 127.0.0.1. You need to confirm the IP address and then see how the Client should set the access IP.
### Docker for Windows User Guide
The above content is used for native Windows. If users want to experience complete functions, they need to use Docker tools to model Linux systems.
Please refer to [Docker Desktop](https://www.docker.com/products/docker-desktop) to install Docker
After installation, start the docker linux engine and download the relevant image. In the Serving directory
```
docker pull hub.baidubce.com/paddlepaddle/serving:latest-devel
# There is no expose port here, users can set -p to perform port mapping as needed
docker run --rm -dit --name serving_devel -v $PWD:/Serving hub.baidubce.com/paddlepaddle/serving:latest-devel
docker exec -it serving_devel bash
cd /Serving
```
The rest of the operations are exactly the same as the Linux version.
doc/WINDOWS_TUTORIAL_CN.md 0 → 100644
浏览文件 @ af65b28f
## Windows平台使用Paddle Serving指导
([English](./WINDOWS_TUTORIAL.md)|简体中文）
### 综述
本文档指导用户如何在Windows平台手把手搭建Paddle Serving服务。由于受限第三方库的支持，Windows平台目前只支持用web service的方式搭建local predictor预测服务。如果想要体验全部的服务，需要使用Docker for Windows，来模拟Linux的运行环境。
### 原生Windows系统运行Paddle Serving
**配置Python环境变量到PATH**：**目前原生Windows仅支持Python 3.5或更高版本**。首先需要将Python的可执行程序所在目录加入到PATH当中。通常在**系统属性/我的电脑属性**-**高级**-**环境变量** ，点选Path，并在开头加上路径。例如`C:\Users\$USER\AppData\Local\Programs\Python\Python36`，最后连续点击**确定** 。在Powershell上如果输入python可以进入python交互界面，说明环境变量配置成功。
**安装wget工具**：由于教程当中所有的下载，以及`paddle_serving_app`当中内嵌的模型下载功能，都是用到wget工具，在链接[下载wget](http://gnuwin32.sourceforge.net/packages/wget.htm)，解压后复制到`C:\Windows\System32`下，如有安全提示需要通过。
**安装Git工具**： 详情参见[Git官网](https://git-scm.com/downloads)
**安装必要的C++库（可选）**：部分用户可能会在`import paddle`阶段遇见dll无法链接的问题，建议可以[安装Visual Studio社区版本](`https://visualstudio.microsoft.com/`) ，并且安装C++的相关组件。
**安装Paddle和Serving**：在Powershell，执行
```
python -m pip install -U paddle_serving_server paddle_serving_client paddle_serving_app paddlepaddle`
```
如果是GPU用户
```
python -m pip install -U paddle_serving_server_gpu paddle_serving_client paddle_serving_app paddlepaddle-gpu
```
**下载Serving库**：
```
git clone https://github.com/paddlepaddle/Serving
pip install -r python/requirements_win.txt
```
**运行OCR示例**：
```
cd Serving/python/example/ocr
python -m paddle_serving_app.package --get_model ocr_rec
tar -xzvf ocr_rec.tar.gz
python -m paddle_serving_app.package --get_model ocr_det
tar -xzvf ocr_det.tar.gz
python ocr_debugger_server.py cpu &
python ocr_web_client.py
```
### 创建新的Windows支持的Paddle Serving服务
目前Windows支持Web Service框架的Local Predictor。服务端代码框架如下
```
# filename:your_webservice.py
from paddle_serving_server.web_service import WebService
# 如果是GPU版本，请使用 from paddle_serving_server_gpu.web_service import WebService
class YourWebService(WebService):
def preprocess(self, feed=[], fetch=[]):
#在这里实现前处理
#feed_dict是 key: var names, value: numpy array input
#fetch_names 是fetch变量名列表
#is_batch的含义是feed_dict的value里的numpy array是否包含了batch维度
return feed_dict, fetch_names, is_batch
def postprocess(self, feed={}, fetch=[], fetch_map=None):
#fetch map是经过预测之后的返回字典，key是process返回时给定的fetch names，value是对应fetch names的var具体值
#在这里做处理之后，结果需重新转换成字典，并且values的类型应是列表list，这样可以JSON序列化方便web返回
return response
your_service = YourService(name="XXX")
your_service.load_model_config("your_model_path")
your_service.prepare_server(workdir="workdir", port=9292)
# 如果是GPU用户，可以参照python/examples/ocr下的python示例
your_service.run_debugger_service()
# Windows平台不可以使用 run_rpc_service()接口
your_service.run_web_service()
```
客户端代码示例
```
# filename：your_client.py
import requests
import json
import base64
import os, sys
import time
import cv2 # 如果需要上传图片
# 用于图片读取，原理是采用base64编码文件内容
def cv2_to_base64(image):
return base64.b64encode(image).decode(
'utf8') #data.tostring()).decode('utf8')
headers = {"Content-type": "application/json"}
url = "http://127.0.0.1:9292/XXX/prediction" # XXX取决于服务端YourService的初始化name参数
r = requests.post(url=url, headers=headers, data=json.dumps(data))
print(r.json())
```
用户只需要按照如上指示，在对应函数中实现相关内容即可。更多信息请参见[如何开发一个新的Web Service？](./NEW_WEB_SERVICE_CN.md)
开发完成后执行
```
python your_webservice.py &
python your_client.py
```
因为需要占用端口，因此启动过程可能会有安全提示，请点选通过，就会有IP地址生成。需要注意的是，Windows平台启动服务时，本地IP地址可能不是127.0.0.1，需要确认好IP地址再看Client应该如何设定访问IP。
### Docker for Windows 使用指南
以上内容用于原生的Windows，如果用户想要体验完整的功能，需要使用Docker工具，来模拟Linux系统。
安装Docker请参考[Docker Desktop](https://www.docker.com/products/docker-desktop)
安装之后启动docker的linux engine，下载相关镜像。在Serving目录下
```
docker pull hub.baidubce.com/paddlepaddle/serving:latest-devel
# 此处没有expose端口，用户可根据需要设置-p来进行端口映射
docker run --rm -dit --name serving_devel -v $PWD:/Serving hub.baidubce.com/paddlepaddle/serving:latest-devel
docker exec -it serving_devel bash
cd /Serving
```
其余操作与Linux版本完全一致。
java/README.md 0 → 100644
浏览文件 @ af65b28f
## Tutorial of Java Client for Paddle Serving
(English|[简体中文](./README_CN.md))
### Development Environment
In order to facilitate users to use java for development, we provide the compiled Serving project to be placed in the java mirror. The way to get the mirror and enter the development environment is
```
docker pull hub.baidubce.com/paddlepaddle/serving:0.4.0-java
docker run --rm -dit --name java_serving hub.baidubce.com/paddlepaddle/serving:0.4.0-java
docker exec -it java_serving bash
cd Serving/java
```
The Serving folder is at the develop branch when the docker image is generated. You need to git pull to the latest version or git checkout to the desired branch.
### Install client dependencies
Due to the large number of dependent libraries, the image has been compiled once at the time of generation, and the user can perform the following operations
```
mvn compile
mvn install
cd examples
mvn compile
mvn install
```
### Start the server
Take the fit_a_line model as an example, the server starts
```
cd ../../python/examples/fit_a_line
sh get_data.sh
python -m paddle_serving_server.serve --model uci_housing_model --thread 10 --port 9393 --use_multilang &
```
Client prediction
```
cd ../../../java/examples/target
java -cp paddle-serving-sdk-java-examples-0.0.1-jar-with-dependencies.jar PaddleServingClientExample fit_a_line
```
Take yolov4 as an example, the server starts
```
python -m paddle_serving_app.package --get_model yolov4
tar -xzvf yolov4.tar.gz
python -m paddle_serving_server_gpu.serve --model yolov4_model --port 9393 --gpu_ids 0 --use_multilang & #It needs to be executed in GPU Docker, otherwise the execution method of CPU must be used.
```
Client prediction
```
# in /Serving/java/examples/target
java -cp paddle-serving-sdk-java-examples-0.0.1-jar-with-dependencies.jar PaddleServingClientExample yolov4 ../../../python/examples/yolov4/000000570688.jpg
# The case of yolov4 needs to specify a picture as input
```
### Customization guidance
The above example is running in CPU mode. If GPU mode is required, there are two options.
The first is that GPU Serving and Java Client are in the same image. After starting the corresponding image, the user needs to move /Serving/java in the java image to the corresponding image.
The second is to deploy GPU Serving and Java Client separately. If they are on the same host, you can learn the IP address of the corresponding container through ifconfig, and then when you connect to client.connect in `examples/src/main/java/PaddleServingClientExample.java` Make changes to the endpoint, and then compile it again. Or select `--net=host` to bind the network device of docker and host when docker starts, so that it can run directly without customizing java code.
**It should be noted that in the example, all models need to use `--use_multilang` to start GRPC multi-programming language support, and the port number is 9393. If you need another port, you need to modify it in the java file**
**Currently Serving has launched the Pipeline mode (see [Pipeline Serving](../doc/PIPELINE_SERVING.md) for details). The next version (0.4.1) of the Pipeline Serving Client for Java will be released. **
java/README_CN.md 0 → 100644
浏览文件 @ af65b28f
## 用于Paddle Serving的Java客户端
([English](./README.md)|简体中文)
### 开发环境
为了方便用户使用java进行开发，我们提供了编译好的Serving工程放置在java镜像当中，获取镜像并进入开发环境的方式是
```
docker pull hub.baidubce.com/paddlepaddle/serving:0.4.0-java
docker run --rm -dit --name java_serving hub.baidubce.com/paddlepaddle/serving:0.4.0-java
docker exec -it java_serving bash
cd Serving/java
```
Serving文件夹是镜像生成时的develop分支工程目录，需要git pull 到最新版本，或者git checkout 到想要的分支。
### 安装客户端依赖
由于依赖库数量庞大，因此镜像已经在生成时编译过一次，用户执行以下操作即可
```
mvn compile
mvn install
cd examples
mvn compile
mvn install
```
### 启动服务端
以fit_a_line模型为例，服务端启动
```
cd ../../python/examples/fit_a_line
sh get_data.sh
python -m paddle_serving_server.serve --model uci_housing_model --thread 10 --port 9393 --use_multilang &
```
客户端预测
```
cd ../../../java/examples/target
java -cp paddle-serving-sdk-java-examples-0.0.1-jar-with-dependencies.jar PaddleServingClientExample fit_a_line
```
以yolov4为例子，服务端启动
```
python -m paddle_serving_app.package --get_model yolov4
tar -xzvf yolov4.tar.gz
python -m paddle_serving_server_gpu.serve --model yolov4_model --port 9393 --gpu_ids 0 --use_multilang & #需要在GPU Docker当中执行，否则要使用CPU的执行方式。
```
客户端预测
```
# in /Serving/java/examples/target
java -cp paddle-serving-sdk-java-examples-0.0.1-jar-with-dependencies.jar PaddleServingClientExample yolov4 ../../../python/examples/yolov4/000000570688.jpg
# yolov4的案例需要指定一个图片作为输入
```
### 二次开发指导
上述示例是在CPU模式下运行，如果需要GPU模式，可以有两种选择。
第一种是GPU Serving和Java Client在同一个镜像，需要用户在启动对应的镜像后，把java镜像当中的/Serving/java移动到对应的镜像中。
第二种是GPU Serving和Java Client分开部署，如果在同一台宿主机，可以通过ifconfig了解对应容器的IP地址，然后在`examples/src/main/java/PaddleServingClientExample.java`当中对client.connect时的endpoint做修改，然后再编译一次。 或者在docker启动时选择 `--net=host`来绑定docker和宿主机的网络设备，这样不需要定制java代码可以直接运行。
**需要注意的是，在示例中，所有模型都需要使用`--use_multilang`来启动GRPC多编程语言支持，以及端口号都是9393，如果需要别的端口，需要在java文件里修改**
**目前Serving已推出Pipeline模式（详见[Pipeline Serving](../doc/PIPELINE_SERVING_CN.md)），下个版本（0.4.1）面向Java的Pipeline Serving Client将会发布，敬请期待。**
java/examples/pom.xml
浏览文件 @ af65b28f
......@@ -75,7 +75,7 @@
<dependency>
<groupId>junit</groupId>
<artifactId>junit</artifactId>
<version>4.11</version>
<version>4.13.1</version>
<scope>test</scope>
</dependency>
<dependency>
......
paddle_inference/inferencer-fluid-cpu/include/fluid_cpu_engine.h
浏览文件 @ af65b28f
......@@ -23,7 +23,6 @@
#include "core/configure/inferencer_configure.pb.h"
#include "core/predictor/framework/infer.h"
#include "paddle_inference_api.h" // NOLINT
//#include "predictor/framework/infer.h"
namespace baidu {
namespace paddle_serving {
......
paddle_inference/inferencer-fluid-gpu/CMakeLists.txt
浏览文件 @ af65b28f
......@@ -2,6 +2,7 @@ FILE(GLOB fluid_gpu_engine_srcs ${CMAKE_CURRENT_LIST_DIR}/src/*.cpp)
add_library(fluid_gpu_engine ${fluid_gpu_engine_srcs})
target_include_directories(fluid_gpu_engine PUBLIC
${CMAKE_BINARY_DIR}/Paddle/fluid_install_dir/)
add_dependencies(fluid_gpu_engine pdserving extern_paddle configure)
target_link_libraries(fluid_gpu_engine pdserving paddle_fluid iomp5 mklml_intel -lpthread -lcrypto -lm -lrt -lssl -ldl -lz)
......
paddle_inference/inferencer-fluid-gpu/include/fluid_gpu_engine.h
浏览文件 @ af65b28f
......@@ -190,7 +190,7 @@ class FluidGpuAnalysisDirCore : public FluidFamilyCore {
paddle::AnalysisConfig analysis_config;
analysis_config.SetModel(data_path);
analysis_config.EnableUseGpu(100, FLAGS_gpuid);
analysis_config.EnableUseGpu(1500, FLAGS_gpuid);
analysis_config.SwitchSpecifyInputNames(true);
analysis_config.SetCpuMathLibraryNumThreads(1);
......@@ -198,12 +198,68 @@ class FluidGpuAnalysisDirCore : public FluidFamilyCore {
analysis_config.EnableMemoryOptim();
}
#if 0 // todo: support flexible shape
int min_seq_len = 1;
int max_seq_len = 512;
int opt_seq_len = 128;
int head_number = 12;
int batch = 50;
std::vector<int> min_in_shape = {batch, min_seq_len, 1};
std::vector<int> max_in_shape = {batch, max_seq_len, 1};
std::vector<int> opt_in_shape = {batch, opt_seq_len, 1};
std::string input1_name = "src_text_a_ids";
std::string input2_name = "pos_text_a_ids";
std::string input3_name = "sent_text_a_ids";
std::string input4_name = "stack_0.tmp_0";
std::map<std::string, std::vector<int>> min_input_shape = {
{input1_name, min_in_shape},
{input2_name, min_in_shape},
{input3_name, min_in_shape},
{input4_name, {batch, head_number, min_seq_len, min_seq_len}},
};
std::map<std::string, std::vector<int>> max_input_shape = {
{input1_name, max_in_shape},
{input2_name, max_in_shape},
{input3_name, max_in_shape},
{input4_name, {batch, head_number, max_seq_len, max_seq_len}},
};
std::map<std::string, std::vector<int>> opt_input_shape = {
{input1_name, opt_in_shape},
{input2_name, opt_in_shape},
{input3_name, opt_in_shape},
{input4_name, {batch, head_number, opt_seq_len, opt_seq_len}},
};
analysis_config.SetTRTDynamicShapeInfo(
min_input_shape, max_input_shape, opt_input_shape);
#endif
int max_batch = 32;
int min_subgraph_size = 3;
if (params.use_trt()) {
analysis_config.EnableTensorRtEngine(
1 << 20,
max_batch,
min_subgraph_size,
paddle::AnalysisConfig::Precision::kFloat32,
false,
false);
LOG(INFO) << "create TensorRT predictor";
} else {
if (params.enable_memory_optimization()) {
analysis_config.EnableMemoryOptim();
}
if (params.enable_ir_optimization()) {
analysis_config.SwitchIrOptim(true);
} else {
analysis_config.SwitchIrOptim(false);
}
}
AutoLock lock(GlobalPaddleCreateMutex::instance());
_core =
paddle::CreatePaddlePredictor<paddle::AnalysisConfig>(analysis_config);
......
python/CMakeLists.txt
浏览文件 @ af65b28f
......@@ -80,6 +80,16 @@ if (SERVER)
COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py bdist_wheel
DEPENDS ${SERVING_SERVER_CORE} server_config_py_proto ${PY_FILES})
add_custom_target(paddle_python ALL DEPENDS ${PADDLE_SERVING_BINARY_DIR}/.timestamp)
elseif(WITH_TRT)
add_custom_command(
OUTPUT ${PADDLE_SERVING_BINARY_DIR}/.timestamp
COMMAND cp -r
${CMAKE_CURRENT_SOURCE_DIR}/paddle_serving_server_gpu/ ${PADDLE_SERVING_BINARY_DIR}/python/
COMMAND env ${py_env} ${PYTHON_EXECUTABLE} gen_version.py
"server_gpu" trt
COMMAND env ${py_env} ${PYTHON_EXECUTABLE} setup.py bdist_wheel
DEPENDS ${SERVING_SERVER_CORE} server_config_py_proto ${PY_FILES})
add_custom_target(paddle_python ALL DEPENDS ${PADDLE_SERVING_BINARY_DIR}/.timestamp)
else()
add_custom_command(
OUTPUT ${PADDLE_SERVING_BINARY_DIR}/.timestamp
......
python/examples/bert/bert_client.py
浏览文件 @ af65b28f
......@@ -18,16 +18,20 @@ import sys
from paddle_serving_client import Client
from paddle_serving_client.utils import benchmark_args
from paddle_serving_app.reader import ChineseBertReader
import numpy as np
args = benchmark_args()
reader = ChineseBertReader({"max_seq_len": 128})
fetch = ["pooled_output"]
endpoint_list = ["127.0.0.1:9292"]
endpoint_list = ['127.0.0.1:9292']
client = Client()
client.load_client_config(args.model)
client.connect(endpoint_list)
for line in sys.stdin:
feed_dict = reader.process(line)
result = client.predict(feed=feed_dict, fetch=fetch)
for key in feed_dict.keys():
feed_dict[key] = np.array(feed_dict[key]).reshape((128, 1))
#print(feed_dict)
result = client.predict(feed=feed_dict, fetch=fetch, batch=False)
print(result)
python/examples/bert/bert_web_service.py
浏览文件 @ af65b28f
......@@ -13,10 +13,11 @@
# See the License for the specific language governing permissions and
# limitations under the License.
# pylint: disable=doc-string-missing
from paddle_serving_server_gpu.web_service import WebService
from paddle_serving_server.web_service import WebService
from paddle_serving_app.reader import ChineseBertReader
import sys
import os
import numpy as np
class BertService(WebService):
......@@ -27,18 +28,21 @@ class BertService(WebService):
})
def preprocess(self, feed=[], fetch=[]):
feed_res = [
self.reader.process(ins["words"].encode("utf-8")) for ins in feed
]
return feed_res, fetch
feed_res = []
is_batch = False
for ins in feed:
feed_dict = self.reader.process(ins["words"].encode("utf-8"))
for key in feed_dict.keys():
feed_dict[key] = np.array(feed_dict[key]).reshape(
(len(feed_dict[key]), 1))
feed_res.append(feed_dict)
return feed_res, fetch, is_batch
bert_service = BertService(name="bert")
bert_service.load()
bert_service.load_model_config(sys.argv[1])
gpu_ids = os.environ["CUDA_VISIBLE_DEVICES"]
bert_service.set_gpus(gpu_ids)
bert_service.prepare_server(
workdir="workdir", port=int(sys.argv[2]), device="gpu")
workdir="workdir", port=int(sys.argv[2]), device="cpu")
bert_service.run_rpc_service()
bert_service.run_web_service()
python/examples/bert/test_multi_fetch_client.py
浏览文件 @ af65b28f
......@@ -15,6 +15,7 @@
from paddle_serving_client import Client
from paddle_serving_app.reader import ChineseBertReader
import sys
import numpy as np
client = Client()
client.load_client_config("./bert_seq32_client/serving_client_conf.prototxt")
......@@ -28,12 +29,21 @@ expected_shape = {
"pooled_output": (4, 768)
}
batch_size = 4
feed_batch = []
feed_batch = {}
batch_len = 0
for line in sys.stdin:
feed = reader.process(line)
if batch_len == 0:
for key in feed.keys():
val_len = len(feed[key])
feed_batch[key] = np.array(feed[key]).reshape((1, val_len, 1))
continue
if len(feed_batch) < batch_size:
feed_batch.append(feed)
for key in feed.keys():
np.concatenate([
feed_batch[key], np.array(feed[key]).reshape((1, val_len, 1))
])
else:
fetch_map = client.predict(feed=feed_batch, fetch=fetch)
feed_batch = []
......
python/examples/criteo_ctr_with_cube/test_server_gpu.py → python/examples/cascade_rcnn/test_client.py 100755 → 100644
浏览文件 @ af65b28f
......@@ -11,31 +11,30 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# pylint: disable=doc-string-missing
import os
import sys
from paddle_serving_server_gpu import OpMaker
from paddle_serving_server_gpu import OpSeqMaker
from paddle_serving_server_gpu import Server
from paddle_serving_client import Client
from paddle_serving_app.reader import *
import numpy as np
op_maker = OpMaker()
read_op = op_maker.create('general_reader')
general_dist_kv_infer_op = op_maker.create('general_dist_kv_infer')
response_op = op_maker.create('general_response')
op_seq_maker = OpSeqMaker()
op_seq_maker.add_op(read_op)
op_seq_maker.add_op(general_dist_kv_infer_op)
op_seq_maker.add_op(response_op)
server = Server()
server.set_op_sequence(op_seq_maker.get_op_sequence())
server.set_num_threads(4)
server.load_model_config(sys.argv[1])
server.prepare_server(
workdir="work_dir1",
port=9292,
device="cpu",
cube_conf="./cube/conf/cube.conf")
server.run_server()
preprocess = Sequential([
File2Image(), BGR2RGB(), Div(255.0),
Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225], False),
Resize(800, 1333), Transpose((2, 0, 1)), PadStride(32)
])
postprocess = RCNNPostprocess("label_list.txt", "output")
client = Client()
client.load_client_config("serving_client/serving_client_conf.prototxt")
client.connect(['127.0.0.1:9292'])
im = preprocess('000000570688.jpg')
fetch_map = client.predict(
feed={
"image": im,
"im_info": np.array(list(im.shape[1:]) + [1.0]),
"im_shape": np.array(list(im.shape[1:]) + [1.0])
},
fetch=["multiclass_nms_0.tmp_0"],
batch=False)
fetch_map["image"] = '000000570688.jpg'
print(fetch_map)
postprocess(fetch_map)
print(fetch_map)
python/examples/criteo_ctr/test_client.py
浏览文件 @ af65b28f
......@@ -20,7 +20,7 @@ import os
import time
import criteo_reader as criteo
from paddle_serving_client.metric import auc
import numpy as np
import sys
py_version = sys.version_info[0]
......@@ -49,7 +49,8 @@ for ei in range(1000):
data = reader().__next__()
feed_dict = {}
for i in range(1, 27):
feed_dict["sparse_{}".format(i - 1)] = data[0][i]
feed_dict["sparse_{}".format(i - 1)] = np.array(data[0][i]).reshape(-1)
feed_dict["sparse_{}.lod".format(i - 1)] = [0, len(data[0][i])]
fetch_map = client.predict(feed=feed_dict, fetch=["prob"])
end = time.time()
print(end - start)
python/examples/criteo_ctr_with_cube/README.md 已删除 100755 → 0
浏览文件 @ bcc0c749
## Criteo CTR with Sparse Parameter Indexing Service
([简体中文](./README_CN.md)|English)
### Get Sample Dataset
go to directory `python/examples/criteo_ctr_with_cube`
```
sh get_data.sh
```
### Download Model and Sparse Parameter Sequence Files
```
wget https://paddle-serving.bj.bcebos.com/unittest/ctr_cube_unittest.tar.gz
tar xf ctr_cube_unittest.tar.gz
mv models/ctr_client_conf ./
mv models/ctr_serving_model_kv ./
mv models/data ./cube/
```
the model will be in ./ctr_server_model_kv and ./ctr_client_config.
### Start Sparse Parameter Indexing Service
```
wget https://paddle-serving.bj.bcebos.com/others/cube_app.tar.gz
tar xf cube_app.tar.gz
mv cube_app/cube* ./cube/
sh cube_prepare.sh &
```
Here, the sparse parameter is loaded by cube sparse parameter indexing service Cube.
### Start RPC Predictor, the number of serving thread is 4（configurable in test_server.py）
```
python test_server.py ctr_serving_model_kv
```
### Run Prediction
```
python test_client.py ctr_client_conf/serving_client_conf.prototxt ./raw_data
```
### Benchmark
CPU ：Intel(R) Xeon(R) CPU 6148 @ 2.40GHz
Model ：[Criteo CTR](https://github.com/PaddlePaddle/Serving/blob/develop/python/examples/criteo_ctr_with_cube/network_conf.py)
server core/thread num ： 4/8
Run
```
bash benchmark.sh
```
1000 batches will be sent by every client
| client thread num | prepro | client infer | op0 | op1 | op2 | postpro | avg_latency | qps |
| ------------------ | ------ | ------------ | ------ | ----- | ------ | ------- | ----- | ----- |
| 1 | 0.035 | 1.596 | 0.021 | 0.518 | 0.0024 | 0.0025 | 6.774 | 147.7 |
| 2 | 0.034 | 1.780 | 0.027 | 0.463 | 0.0020 | 0.0023 | 6.931 | 288.3 |
| 4 | 0.038 | 2.954 | 0.025 | 0.455 | 0.0019 | 0.0027 | 8.378 | 477.5 |
| 8 | 0.044 | 8.230 | 0.028 | 0.464 | 0.0023 | 0.0034 | 14.191 | 563.8 |
| 16 | 0.048 | 21.037 | 0.028 | 0.455 | 0.0025 | 0.0041 | 27.236 | 587.5 |
the average latency of threads
![avg cost](../../../doc/criteo-cube-benchmark-avgcost.png)
The QPS is
![qps](../../../doc/criteo-cube-benchmark-qps.png)
python/examples/criteo_ctr_with_cube/README_CN.md 已删除 100644 → 0
浏览文件 @ bcc0c749
## 带稀疏参数索引服务的CTR预测服务
(简体中文|[English](./README.md))
### 获取样例数据
进入目录 `python/examples/criteo_ctr_with_cube`
```
sh get_data.sh
```
### 下载模型和稀疏参数序列文件
```
wget https://paddle-serving.bj.bcebos.com/unittest/ctr_cube_unittest.tar.gz
tar xf ctr_cube_unittest.tar.gz
mv models/ctr_client_conf ./
mv models/ctr_serving_model_kv ./
mv models/data ./cube/
```
执行脚本后会在当前目录有ctr_server_model_kv和ctr_client_config文件夹。
### 启动稀疏参数索引服务
```
wget https://paddle-serving.bj.bcebos.com/others/cube_app.tar.gz
tar xf cube_app.tar.gz
mv cube_app/cube* ./cube/
sh cube_prepare.sh &
```
此处，模型当中的稀疏参数会被存放在稀疏参数索引服务Cube当中。
### 启动RPC预测服务，服务端线程数为4（可在test_server.py配置）
```
python test_server.py ctr_serving_model_kv
```
### 执行预测
```
python test_client.py ctr_client_conf/serving_client_conf.prototxt ./raw_data
```
### Benchmark
设备 ：Intel(R) Xeon(R) CPU 6148 @ 2.40GHz
模型 ：[Criteo CTR](https://github.com/PaddlePaddle/Serving/blob/develop/python/examples/criteo_ctr_with_cube/network_conf.py)
server core/thread num ： 4/8
执行
```
bash benchmark.sh
```
客户端每个线程会发送1000个batch
| client thread num | prepro | client infer | op0 | op1 | op2 | postpro | avg_latency | qps |
| ------------------ | ------ | ------------ | ------ | ----- | ------ | ------- | ----- | ----- |
| 1 | 0.035 | 1.596 | 0.021 | 0.518 | 0.0024 | 0.0025 | 6.774 | 147.7 |
| 2 | 0.034 | 1.780 | 0.027 | 0.463 | 0.0020 | 0.0023 | 6.931 | 288.3 |
| 4 | 0.038 | 2.954 | 0.025 | 0.455 | 0.0019 | 0.0027 | 8.378 | 477.5 |
| 8 | 0.044 | 8.230 | 0.028 | 0.464 | 0.0023 | 0.0034 | 14.191 | 563.8 |
| 16 | 0.048 | 21.037 | 0.028 | 0.455 | 0.0025 | 0.0041 | 27.236 | 587.5 |
平均每个线程耗时图如下
![avg cost](../../../doc/criteo-cube-benchmark-avgcost.png)
每个线程QPS耗时如下
![qps](../../../doc/criteo-cube-benchmark-qps.png)
python/examples/criteo_ctr_with_cube/args.py 已删除 100755 → 0
浏览文件 @ bcc0c749
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# pylint: disable=doc-string-missing
import argparse
def parse_args():
parser = argparse.ArgumentParser(description="PaddlePaddle CTR example")
parser.add_argument(
'--train_data_path',
type=str,
default='./data/raw/train.txt',
help="The path of training dataset")
parser.add_argument(
'--sparse_only',
type=bool,
default=False,
help="Whether we use sparse features only")
parser.add_argument(
'--test_data_path',
type=str,
default='./data/raw/valid.txt',
help="The path of testing dataset")
parser.add_argument(
'--batch_size',
type=int,
default=1000,
help="The size of mini-batch (default:1000)")
parser.add_argument(
'--embedding_size',
type=int,
default=10,
help="The size for embedding layer (default:10)")
parser.add_argument(
'--num_passes',
type=int,
default=10,
help="The number of passes to train (default: 10)")
parser.add_argument(
'--model_output_dir',
type=str,
default='models',
help='The path for model to store (default: models)')
parser.add_argument(
'--sparse_feature_dim',
type=int,
default=1000001,
help='sparse feature hashing space for index processing')
parser.add_argument(
'--is_local',
type=int,
default=1,
help='Local train or distributed train (default: 1)')
parser.add_argument(
'--cloud_train',
type=int,
default=0,
help='Local train or distributed train on paddlecloud (default: 0)')
parser.add_argument(
'--async_mode',
action='store_true',
default=False,
help='Whether start pserver in async mode to support ASGD')
parser.add_argument(
'--no_split_var',
action='store_true',
default=False,
help='Whether split variables into blocks when update_method is pserver')
parser.add_argument(
'--role',
type=str,
default='pserver', # trainer or pserver
help='The path for model to store (default: models)')
parser.add_argument(
'--endpoints',
type=str,
default='127.0.0.1:6000',
help='The pserver endpoints, like: 127.0.0.1:6000,127.0.0.1:6001')
parser.add_argument(
'--current_endpoint',
type=str,
default='127.0.0.1:6000',
help='The path for model to store (default: 127.0.0.1:6000)')
parser.add_argument(
'--trainer_id',
type=int,
default=0,
help='The path for model to store (default: models)')
parser.add_argument(
'--trainers',
type=int,
default=1,
help='The num of trianers, (default: 1)')
return parser.parse_args()
python/examples/criteo_ctr_with_cube/benchmark.sh 已删除 100755 → 0
浏览文件 @ bcc0c749
rm profile_log
export FLAGS_profile_client=1
export FLAGS_profile_server=1
wget https://paddle-serving.bj.bcebos.com/unittest/ctr_cube_unittest.tar.gz --no-check-certificate
tar xf ctr_cube_unittest.tar.gz
mv models/ctr_client_conf ./
mv models/ctr_serving_model_kv ./
mv models/data ./cube/
wget https://paddle-serving.bj.bcebos.com/others/cube_app.tar.gz --no-check-certificate
tar xf cube_app.tar.gz
mv cube_app/cube* ./cube/
sh cube_prepare.sh &
python test_server.py ctr_serving_model_kv > serving_log 2>&1 &
for thread_num in 1 4 16
do
for batch_size in 1 4 16 64
do
$PYTHONROOT/bin/python benchmark.py --thread $thread_num --batch_size $batch_size --model serving_client_conf/serving_client_conf.prototxt --request rpc > profile 2>&1
echo "batch size : $batch_size"
echo "thread num : $thread_num"
echo "========================================"
echo "batch size : $batch_size" >> profile_log
$PYTHONROOT/bin/python ../util/show_profile.py profile $thread_num >> profile_log
tail -n 3 profile >> profile_log
done
done
ps -ef|grep 'serving'|grep -v grep|cut -c 9-15 | xargs kill -9
python/examples/criteo_ctr_with_cube/benchmark_cube.sh 已删除 100755 → 0
浏览文件 @ bcc0c749
rm profile_log
#wget https://paddle-serving.bj.bcebos.com/unittest/ctr_cube_unittest.tar.gz --no-check-certificate
#tar xf ctr_cube_unittest.tar.gz
mv models/ctr_client_conf ./
mv models/ctr_serving_model_kv ./
mv models/data ./cube/
#wget https://paddle-serving.bj.bcebos.com/others/cube_app.tar.gz --no-check-certificate
#tar xf cube_app.tar.gz
mv cube_app/cube* ./cube/
sh cube_prepare.sh &
cp ../../../build_server/core/cube/cube-api/cube-cli .
python gen_key.py
for thread_num in 1 4 16 32
do
for batch_size in 1000
do
./cube-cli -config_file ./cube/conf/cube.conf -keys key -dict test_dict -thread_num $thread_num --batch $batch_size > profile 2>&1
echo "batch size : $batch_size"
echo "thread num : $thread_num"
echo "========================================"
echo "batch size : $batch_size" >> profile_log
echo "thread num : $thread_num" >> profile_log
tail -n 8 profile >> profile_log
done
done
ps -ef|grep 'cube'|grep -v grep|cut -c 9-15 | xargs kill -9
python/examples/criteo_ctr_with_cube/clean.sh 已删除 100755 → 0
浏览文件 @ bcc0c749
ps -ef | grep cube | awk {'print $2'} | xargs kill -9
rm -rf cube/cube_data cube/data cube/log* cube/nohup* cube/output/ cube/donefile cube/input cube/monitor cube/cube-builder.INFO
ps -ef | grep test | awk {'print $2'} | xargs kill -9
ps -ef | grep serving | awk {'print $2'} | xargs kill -9
python/examples/criteo_ctr_with_cube/criteo.py 已删除 100755 → 0
浏览文件 @ bcc0c749
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import sys
class CriteoDataset(object):
def setup(self, sparse_feature_dim):
self.cont_min_ = [0, -3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
self.cont_max_ = [
20, 600, 100, 50, 64000, 500, 100, 50, 500, 10, 10, 10, 50
]
self.cont_diff_ = [
20, 603, 100, 50, 64000, 500, 100, 50, 500, 10, 10, 10, 50
]
self.hash_dim_ = sparse_feature_dim
# here, training data are lines with line_index < train_idx_
self.train_idx_ = 41256555
self.continuous_range_ = range(1, 14)
self.categorical_range_ = range(14, 40)
def _process_line(self, line):
features = line.rstrip('\n').split('\t')
dense_feature = []
sparse_feature = []
for idx in self.continuous_range_:
if features[idx] == '':
dense_feature.append(0.0)
else:
dense_feature.append((float(features[idx]) - self.cont_min_[idx - 1]) / \
self.cont_diff_[idx - 1])
for idx in self.categorical_range_:
sparse_feature.append(
[hash(str(idx) + features[idx]) % self.hash_dim_])
return dense_feature, sparse_feature, [int(features[0])]
def infer_reader(self, filelist, batch, buf_size):
def local_iter():
for fname in filelist:
with open(fname.strip(), "r") as fin:
for line in fin:
dense_feature, sparse_feature, label = self._process_line(
line)
#yield dense_feature, sparse_feature, label
yield [dense_feature] + sparse_feature + [label]
import paddle
batch_iter = paddle.batch(
paddle.reader.shuffle(
local_iter, buf_size=buf_size),
batch_size=batch)
return batch_iter
def generate_sample(self, line):
def data_iter():
dense_feature, sparse_feature, label = self._process_line(line)
feature_name = ["dense_input"]
for idx in self.categorical_range_:
feature_name.append("C" + str(idx - 13))
feature_name.append("label")
yield zip(feature_name, [dense_feature] + sparse_feature + [label])
return data_iter
if __name__ == "__main__":
criteo_dataset = CriteoDataset()
criteo_dataset.setup(int(sys.argv[1]))
criteo_dataset.run_from_stdin()
python/examples/criteo_ctr_with_cube/criteo_reader.py 已删除 100755 → 0
浏览文件 @ bcc0c749
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# pylint: disable=doc-string-missing
import sys
import paddle.fluid.incubate.data_generator as dg
class CriteoDataset(dg.MultiSlotDataGenerator):
def setup(self, sparse_feature_dim):
self.cont_min_ = [0, -3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
self.cont_max_ = [
20, 600, 100, 50, 64000, 500, 100, 50, 500, 10, 10, 10, 50
]
self.cont_diff_ = [
20, 603, 100, 50, 64000, 500, 100, 50, 500, 10, 10, 10, 50
]
self.hash_dim_ = sparse_feature_dim
# here, training data are lines with line_index < train_idx_
self.train_idx_ = 41256555
self.continuous_range_ = range(1, 14)
self.categorical_range_ = range(14, 40)
def _process_line(self, line):
features = line.rstrip('\n').split('\t')
dense_feature = []
sparse_feature = []
for idx in self.continuous_range_:
if features[idx] == '':
dense_feature.append(0.0)
else:
dense_feature.append((float(features[idx]) - self.cont_min_[idx - 1]) / \
self.cont_diff_[idx - 1])
for idx in self.categorical_range_:
sparse_feature.append(
[hash(str(idx) + features[idx]) % self.hash_dim_])
return dense_feature, sparse_feature, [int(features[0])]
def infer_reader(self, filelist, batch, buf_size):
def local_iter():
for fname in filelist:
with open(fname.strip(), "r") as fin:
for line in fin:
dense_feature, sparse_feature, label = self._process_line(
line)
#yield dense_feature, sparse_feature, label
yield [dense_feature] + sparse_feature + [label]
import paddle
batch_iter = paddle.batch(
paddle.reader.shuffle(
local_iter, buf_size=buf_size),
batch_size=batch)
return batch_iter
def generate_sample(self, line):
def data_iter():
dense_feature, sparse_feature, label = self._process_line(line)
feature_name = ["dense_input"]
for idx in self.categorical_range_:
feature_name.append("C" + str(idx - 13))
feature_name.append("label")
yield zip(feature_name, [dense_feature] + sparse_feature + [label])
return data_iter
if __name__ == "__main__":
criteo_dataset = CriteoDataset()
criteo_dataset.setup(int(sys.argv[1]))
criteo_dataset.run_from_stdin()
python/examples/criteo_ctr_with_cube/cube/conf/cube.conf 已删除 100755 → 0
浏览文件 @ bcc0c749
[{
"dict_name": "test_dict",
"shard": 1,
"dup": 1,
"timeout": 200,
"retry": 3,
"backup_request": 100,
"type": "ipport_list",
"load_balancer": "rr",
"nodes": [{
"ipport_list": "list://127.0.0.1:8027"
}]
}]
python/examples/criteo_ctr_with_cube/cube/conf/gflags.conf 已删除 100755 → 0
浏览文件 @ bcc0c749
--port=8027
--dict_split=1
--in_mem=true
--log_dir=./log/
python/examples/criteo_ctr_with_cube/cube_prepare.sh 已删除 100755 → 0
浏览文件 @ bcc0c749
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# pylint: disable=doc-string-missing
#! /bin/bash
mkdir -p cube_model
mkdir -p cube/data
./cube/cube-builder -dict_name=test_dict -job_mode=base -last_version=0 -cur_version=0 -depend_version=0 -input_path=./cube_model -output_path=${PWD}/cube/data -shard_num=1 -only_build=false
cd cube && ./cube
python/examples/criteo_ctr_with_cube/cube_quant_prepare.sh 已删除 100755 → 0
浏览文件 @ bcc0c749
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# pylint: disable=doc-string-missing
#! /bin/bash
mkdir -p cube_model
mkdir -p cube/data
./seq_generator ctr_serving_model/SparseFeatFactors ./cube_model/feature 8
./cube/cube-builder -dict_name=test_dict -job_mode=base -last_version=0 -cur_version=0 -depend_version=0 -input_path=./cube_model -output_path=${PWD}/cube/data -shard_num=1 -only_build=false
mv ./cube/data/0_0/test_dict_part0/* ./cube/data/
cd cube && ./cube
python/examples/criteo_ctr_with_cube/get_data.sh 已删除 100755 → 0
浏览文件 @ bcc0c749
wget --no-check-certificate https://paddle-serving.bj.bcebos.com/data/ctr_prediction/ctr_data.tar.gz
tar -zxvf ctr_data.tar.gz
python/examples/criteo_ctr_with_cube/local_train.py 已删除 100755 → 0
浏览文件 @ bcc0c749
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# pylint: disable=doc-string-missing
from __future__ import print_function
from args import parse_args
import os
import paddle.fluid as fluid
import sys
from network_conf import dnn_model
dense_feature_dim = 13
def train():
args = parse_args()
sparse_only = args.sparse_only
if not os.path.isdir(args.model_output_dir):
os.mkdir(args.model_output_dir)
dense_input = fluid.layers.data(
name="dense_input", shape=[dense_feature_dim], dtype='float32')
sparse_input_ids = [
fluid.layers.data(
name="C" + str(i), shape=[1], lod_level=1, dtype="int64")
for i in range(1, 27)
]
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
#nn_input = None if sparse_only else dense_input
nn_input = dense_input
predict_y, loss, auc_var, batch_auc_var, infer_vars = dnn_model(
nn_input, sparse_input_ids, label, args.embedding_size,
args.sparse_feature_dim)
optimizer = fluid.optimizer.SGD(learning_rate=1e-4)
optimizer.minimize(loss)
exe = fluid.Executor(fluid.CPUPlace())
exe.run(fluid.default_startup_program())
dataset = fluid.DatasetFactory().create_dataset("InMemoryDataset")
dataset.set_use_var([dense_input] + sparse_input_ids + [label])
python_executable = "python"
pipe_command = "{} criteo_reader.py {}".format(python_executable,
args.sparse_feature_dim)
dataset.set_pipe_command(pipe_command)
dataset.set_batch_size(128)
thread_num = 10
dataset.set_thread(thread_num)
whole_filelist = [
"raw_data/part-%d" % x for x in range(len(os.listdir("raw_data")))
]
print(whole_filelist)
dataset.set_filelist(whole_filelist[:100])
dataset.load_into_memory()
fluid.layers.Print(auc_var)
epochs = 1
for i in range(epochs):
exe.train_from_dataset(
program=fluid.default_main_program(), dataset=dataset, debug=True)
print("epoch {} finished".format(i))
import paddle_serving_client.io as server_io
feed_var_dict = {}
feed_var_dict['dense_input'] = dense_input
for i, sparse in enumerate(sparse_input_ids):
feed_var_dict["embedding_{}.tmp_0".format(i)] = sparse
fetch_var_dict = {"prob": predict_y}
feed_kv_dict = {}
feed_kv_dict['dense_input'] = dense_input
for i, emb in enumerate(infer_vars):
feed_kv_dict["embedding_{}.tmp_0".format(i)] = emb
fetch_var_dict = {"prob": predict_y}
server_io.save_model("ctr_serving_model", "ctr_client_conf", feed_var_dict,
fetch_var_dict, fluid.default_main_program())
server_io.save_model("ctr_serving_model_kv", "ctr_client_conf_kv",
feed_kv_dict, fetch_var_dict,
fluid.default_main_program())
if __name__ == '__main__':
train()
python/examples/criteo_ctr_with_cube/network_conf.py 已删除 100755 → 0
浏览文件 @ bcc0c749
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# pylint: disable=doc-string-missing
import paddle.fluid as fluid
import math
def dnn_model(dense_input, sparse_inputs, label, embedding_size,
sparse_feature_dim):
def embedding_layer(input):
emb = fluid.layers.embedding(
input=input,
is_sparse=True,
is_distributed=False,
size=[sparse_feature_dim, embedding_size],
param_attr=fluid.ParamAttr(
name="SparseFeatFactors",
initializer=fluid.initializer.Uniform()))
x = fluid.layers.sequence_pool(input=emb, pool_type='sum')
return emb, x
def mlp_input_tensor(emb_sums, dense_tensor):
#if isinstance(dense_tensor, fluid.Variable):
# return fluid.layers.concat(emb_sums, axis=1)
#else:
return fluid.layers.concat(emb_sums + [dense_tensor], axis=1)
def mlp(mlp_input):
fc1 = fluid.layers.fc(input=mlp_input,
size=400,
act='relu',
param_attr=fluid.ParamAttr(
initializer=fluid.initializer.Normal(
scale=1 / math.sqrt(mlp_input.shape[1]))))
fc2 = fluid.layers.fc(input=fc1,
size=400,
act='relu',
param_attr=fluid.ParamAttr(
initializer=fluid.initializer.Normal(
scale=1 / math.sqrt(fc1.shape[1]))))
fc3 = fluid.layers.fc(input=fc2,
size=400,
act='relu',
param_attr=fluid.ParamAttr(
initializer=fluid.initializer.Normal(
scale=1 / math.sqrt(fc2.shape[1]))))
pre = fluid.layers.fc(input=fc3,
size=2,
act='softmax',
param_attr=fluid.ParamAttr(
initializer=fluid.initializer.Normal(
scale=1 / math.sqrt(fc3.shape[1]))))
return pre
emb_pair_sums = list(map(embedding_layer, sparse_inputs))
emb_sums = [x[1] for x in emb_pair_sums]
infer_vars = [x[0] for x in emb_pair_sums]
mlp_in = mlp_input_tensor(emb_sums, dense_input)
predict = mlp(mlp_in)
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.reduce_sum(cost)
accuracy = fluid.layers.accuracy(input=predict, label=label)
auc_var, batch_auc_var, auc_states = \
fluid.layers.auc(input=predict, label=label, num_thresholds=2 ** 12, slide_steps=20)
return predict, avg_cost, auc_var, batch_auc_var, infer_vars
python/examples/criteo_ctr_with_cube/test_client.py 已删除 100755 → 0
浏览文件 @ bcc0c749
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# pylint: disable=doc-string-missing
from paddle_serving_client import Client
import sys
import os
import criteo as criteo
import time
from paddle_serving_client.metric import auc
py_version = sys.version_info[0]
client = Client()
client.load_client_config(sys.argv[1])
client.connect(["127.0.0.1:9292"])
batch = 1
buf_size = 100
dataset = criteo.CriteoDataset()
dataset.setup(1000001)
test_filelists = ["{}/part-0".format(sys.argv[2])]
reader = dataset.infer_reader(test_filelists, batch, buf_size)
label_list = []
prob_list = []
start = time.time()
for ei in range(10000):
if py_version == 2:
data = reader().next()
else:
data = reader().__next__()
feed_dict = {}
feed_dict['dense_input'] = data[0][0]
for i in range(1, 27):
feed_dict["embedding_{}.tmp_0".format(i - 1)] = data[0][i]
fetch_map = client.predict(feed=feed_dict, fetch=["prob"])
prob_list.append(fetch_map['prob'][0][1])
label_list.append(data[0][-1][0])
print(auc(label_list, prob_list))
end = time.time()
print(end - start)
python/examples/criteo_ctr_with_cube/test_server.py 已删除 100755 → 0
浏览文件 @ bcc0c749
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# pylint: disable=doc-string-missing
import os
import sys
from paddle_serving_server import OpMaker
from paddle_serving_server import OpSeqMaker
from paddle_serving_server import Server
op_maker = OpMaker()
read_op = op_maker.create('general_reader')
general_dist_kv_infer_op = op_maker.create('general_dist_kv_infer')
response_op = op_maker.create('general_response')
op_seq_maker = OpSeqMaker()
op_seq_maker.add_op(read_op)
op_seq_maker.add_op(general_dist_kv_infer_op)
op_seq_maker.add_op(response_op)
server = Server()
server.set_op_sequence(op_seq_maker.get_op_sequence())
server.set_num_threads(4)
server.load_model_config(sys.argv[1])
server.prepare_server(
workdir="work_dir1",
port=9292,
device="cpu",
cube_conf="./cube/conf/cube.conf")
server.run_server()
python/examples/criteo_ctr_with_cube/test_server_quant.py 已删除 100755 → 0
浏览文件 @ bcc0c749
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# pylint: disable=doc-string-missing
import os
import sys
from paddle_serving_server import OpMaker
from paddle_serving_server import OpSeqMaker
from paddle_serving_server import Server
op_maker = OpMaker()
read_op = op_maker.create('general_reader')
general_dist_kv_infer_op = op_maker.create('general_dist_kv_quant_infer')
response_op = op_maker.create('general_response')
op_seq_maker = OpSeqMaker()
op_seq_maker.add_op(read_op)
op_seq_maker.add_op(general_dist_kv_infer_op)
op_seq_maker.add_op(response_op)
server = Server()
server.set_op_sequence(op_seq_maker.get_op_sequence())
server.set_num_threads(4)
server.load_model_config(sys.argv[1])
server.prepare_server(
workdir="work_dir1",
port=9292,
device="cpu",
cube_conf="./cube/conf/cube.conf")
server.run_server()
python/examples/criteo_ctr_with_cube/benchmark.py → python/examples/faster_rcnn/benchmark.py
浏览文件 @ af65b28f
......@@ -15,77 +15,109 @@
# limitations under the License.
# pylint: disable=doc-string-missing
from paddle_serving_client import Client
import sys
from __future__ import unicode_literals, absolute_import
import os
import criteo as criteo
import sys
import time
import json
import requests
from paddle_serving_client import Client
from paddle_serving_client.utils import MultiThreadRunner
from paddle_serving_client.utils import benchmark_args
from paddle_serving_client.metric import auc
from paddle_serving_client.utils import benchmark_args, show_latency
from paddle_serving_app.reader import ChineseBertReader
from paddle_serving_app.reader import *
import numpy as np
py_version = sys.version_info[0]
args = benchmark_args()
def single_func(idx, resource):
client = Client()
print([resource["endpoint"][idx % len(resource["endpoint"])]])
client.load_client_config('ctr_client_conf/serving_client_conf.prototxt')
client.connect(['127.0.0.1:9292'])
batch = 1
buf_size = 100
dataset = criteo.CriteoDataset()
dataset.setup(1000001)
test_filelists = [
"./raw_data/part-%d" % x for x in range(len(os.listdir("./raw_data")))
]
reader = dataset.infer_reader(test_filelists[len(test_filelists) - 40:],
batch, buf_size)
img = "./000000570688.jpg"
profile_flags = False
latency_flags = False
if os.getenv("FLAGS_profile_client"):
profile_flags = True
if os.getenv("FLAGS_serving_latency"):
latency_flags = True
latency_list = []
if args.request == "rpc":
fetch = ["prob"]
preprocess = Sequential([
File2Image(), BGR2RGB(), Div(255.0),
Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225], False),
Resize(640, 640), Transpose((2, 0, 1))
])
postprocess = RCNNPostprocess("label_list.txt", "output")
client = Client()
client.load_client_config(args.model)
client.connect([resource["endpoint"][idx % len(resource["endpoint"])]])
start = time.time()
itr = 1000
for ei in range(itr):
if args.batch_size > 0:
for i in range(turns):
if args.batch_size >= 1:
l_start = time.time()
feed_batch = []
b_start = time.time()
im = preprocess(img)
for bi in range(args.batch_size):
if py_version == 2:
data = reader().next()
else:
data = reader().__next__()
feed_dict = {}
feed_dict['dense_input'] = data[0][0]
for i in range(1, 27):
feed_dict["embedding_{}.tmp_0".format(i - 1)] = data[0][
i]
feed_batch.append(feed_dict)
result = client.predict(feed=feed_batch, fetch=fetch)
print("1111batch")
print(bi)
feed_batch.append({
"image": im,
"im_info": np.array(list(im.shape[1:]) + [1.0]),
"im_shape": np.array(list(im.shape[1:]) + [1.0])
})
# im = preprocess(img)
b_end = time.time()
if profile_flags:
sys.stderr.write(
"PROFILE\tpid:{}\tbert_pre_0:{} bert_pre_1:{}\n".format(
os.getpid(),
int(round(b_start * 1000000)),
int(round(b_end * 1000000))))
#result = client.predict(feed=feed_batch, fetch=fetch)
fetch_map = client.predict(
feed=feed_batch, fetch=["multiclass_nms"])
fetch_map["image"] = img
postprocess(fetch_map)
l_end = time.time()
if latency_flags:
latency_list.append(l_end * 1000 - l_start * 1000)
else:
print("unsupport batch size {}".format(args.batch_size))
elif args.request == "http":
raise ("Not support http service.")
else:
raise ValueError("not implemented {} request".format(args.request))
end = time.time()
qps = itr * args.batch_size / (end - start)
return [[end - start, qps]]
if latency_flags:
return [[end - start], latency_list]
else:
return [[end - start]]
if __name__ == '__main__':
multi_thread_runner = MultiThreadRunner()
endpoint_list = ["127.0.0.1:9292"]
#result = single_func(0, {"endpoint": endpoint_list})
endpoint_list = ["127.0.0.1:7777"]
turns = 10
start = time.time()
result = multi_thread_runner.run(single_func, args.thread,
{"endpoint": endpoint_list})
result = multi_thread_runner.run(
single_func, args.thread, {"endpoint": endpoint_list,
"turns": turns})
end = time.time()
total_cost = end - start
avg_cost = 0
qps = 0
for i in range(args.thread):
avg_cost += result[0][i * 2 + 0]
qps += result[0][i * 2 + 1]
avg_cost += result[0][i]
avg_cost = avg_cost / args.thread
print("total cost: {}".format(total_cost))
print("average total cost {} s.".format(avg_cost))
print("qps {} ins/s".format(qps))
print("total cost: {}s".format(total_cost))
print("each thread cost: {}s. ".format(avg_cost))
print("qps: {}samples/s".format(args.batch_size * args.thread * turns /
total_cost))
if os.getenv("FLAGS_serving_latency"):
show_latency(result[1])
python/examples/faster_rcnn/benchmark.sh 0 → 100755
浏览文件 @ af65b28f
rm profile_log*
export CUDA_VISIBLE_DEVICES=0
export FLAGS_profile_server=1
export FLAGS_profile_client=1
export FLAGS_serving_latency=1
gpu_id=0
#save cpu and gpu utilization log
if [ -d utilization ];then
rm -rf utilization
else
mkdir utilization
fi
#start server
$PYTHONROOT/bin/python3 -m paddle_serving_server_gpu.serve --model $1 --port 7777 --thread 4 --gpu_ids 0 --ir_optim > elog 2>&1 &
sleep 5
#warm up
$PYTHONROOT/bin/python3 benchmark.py --thread 4 --batch_size 1 --model $2/serving_client_conf.prototxt --request rpc > profile 2>&1
echo -e "import psutil\ncpu_utilization=psutil.cpu_percent(1,False)\nprint('CPU_UTILIZATION:', cpu_utilization)\n" > cpu_utilization.py
for thread_num in 1 4 8 16
do
for batch_size in 1
do
job_bt=`date '+%Y%m%d%H%M%S'`
nvidia-smi --id=0 --query-compute-apps=used_memory --format=csv -lms 100 > gpu_use.log 2>&1 &
nvidia-smi --id=0 --query-gpu=utilization.gpu --format=csv -lms 100 > gpu_utilization.log 2>&1 &
gpu_memory_pid=$!
$PYTHONROOT/bin/python3 benchmark.py --thread $thread_num --batch_size $batch_size --model $2/serving_client_conf.prototxt --request rpc > profile 2>&1
kill ${gpu_memory_pid}
kill `ps -ef|grep used_memory|awk '{print $2}'`
echo "model_name:" $1
echo "thread_num:" $thread_num
echo "batch_size:" $batch_size
echo "=================Done===================="
echo "model_name:$1" >> profile_log_$1
echo "batch_size:$batch_size" >> profile_log_$1
$PYTHONROOT/bin/python3 cpu_utilization.py >> profile_log_$1
job_et=`date '+%Y%m%d%H%M%S'`
awk 'BEGIN {max = 0} {if(NR>1){if ($1 > max) max=$1}} END {print "MAX_GPU_MEMORY:", max}' gpu_use.log >> profile_log_$1
awk 'BEGIN {max = 0} {if(NR>1){if ($1 > max) max=$1}} END {print "GPU_UTILIZATION:", max}' gpu_utilization.log >> profile_log_$1
rm -rf gpu_use.log gpu_utilization.log
$PYTHONROOT/bin/python3 ../util/show_profile.py profile $thread_num >> profile_log_$1
tail -n 8 profile >> profile_log_$1
echo "" >> profile_log_$1
done
done
#Divided log
awk 'BEGIN{RS="\n\n"}{i++}{print > "bert_log_"i}' profile_log_$1
mkdir bert_log && mv bert_log_* bert_log
ps -ef|grep 'serving'|grep -v grep|cut -c 9-15 | xargs kill -9
python/examples/faster_rcnn/label_list.txt 0 → 100644
浏览文件 @ af65b28f
background
person
bicycle
car
motorcycle
airplane
bus
train
truck
boat
traffic light
fire hydrant
stop sign
parking meter
bench
bird
cat
dog
horse
sheep
cow
elephant
bear
zebra
giraffe
backpack
umbrella
handbag
tie
suitcase
frisbee
skis
snowboard
sports ball
kite
baseball bat
baseball glove
skateboard
surfboard
tennis racket
bottle
wine glass
cup
fork
knife
spoon
bowl
banana
apple
sandwich
orange
broccoli
carrot
hot dog
pizza
donut
cake
chair
couch
potted plant
bed
dining table
toilet
tv
laptop
mouse
remote
keyboard
cell phone
microwave
oven
toaster
sink
refrigerator
book
clock
vase
scissors
teddy bear
hair drier
toothbrush
python/examples/faster_rcnn_model/test_client.py → python/examples/faster_rcnn/test_client.py
浏览文件 @ af65b28f
......@@ -36,6 +36,7 @@ fetch_map = client.predict(
"im_info": np.array(list(im.shape[1:]) + [1.0]),
"im_shape": np.array(list(im.shape[1:]) + [1.0])
},
fetch=["multiclass_nms"])
fetch=["multiclass_nms"],
batch=False)
fetch_map["image"] = sys.argv[3]
postprocess(fetch_map)
python/examples/fit_a_line/README.md
浏览文件 @ af65b28f
......@@ -14,12 +14,6 @@ sh get_data.sh
### Start server
``` shell
python test_server.py uci_housing_model/
```
You can also start the default RPC service with the following line of code:
```shell
python -m paddle_serving_server.serve --model uci_housing_model --thread 10 --port 9393
```
......@@ -40,7 +34,7 @@ python test_client.py uci_housing_client/serving_client_conf.prototxt
Start a web service with default web service hosting modules:
``` shell
python -m paddle_serving_server.serve --model uci_housing_model --thread 10 --port 9393 --name uci
python test_server.py
```
### Client prediction
......
python/examples/fit_a_line/README_CN.md
浏览文件 @ af65b28f
......@@ -41,7 +41,7 @@ python test_client.py uci_housing_client/serving_client_conf.prototxt
通过下面的一行代码开启默认web服务：
``` shell
python -m paddle_serving_server.serve --model uci_housing_model --thread 10 --port 9393 --name uci
python test_server.py
```
### 客户端预测
......
python/examples/fit_a_line/test_client.py
浏览文件 @ af65b28f
......@@ -15,6 +15,7 @@
from paddle_serving_client import Client
import sys
import numpy as np
client = Client()
client.load_client_config(sys.argv[1])
......@@ -27,5 +28,9 @@ test_reader = paddle.batch(
batch_size=1)
for data in test_reader():
fetch_map = client.predict(feed={"x": data[0][0]}, fetch=["price"])
new_data = np.zeros((1, 1, 13)).astype("float32")
new_data[0] = data[0][0]
fetch_map = client.predict(
feed={"x": new_data}, fetch=["price"], batch=True)
print("{} {}".format(fetch_map["price"][0], data[0][1][0]))
print(fetch_map)
python/examples/fit_a_line/test_multi_process_client.py
浏览文件 @ af65b28f
......@@ -15,6 +15,7 @@
from paddle_serving_client import Client
from paddle_serving_client.utils import MultiThreadRunner
import paddle
import numpy as np
def single_func(idx, resource):
......@@ -26,6 +27,7 @@ def single_func(idx, resource):
0.0137, -0.1136, 0.2553, -0.0692, 0.0582, -0.0727, -0.1583, -0.0584,
0.6283, 0.4919, 0.1856, 0.0795, -0.0332
]
x = np.array(x)
for i in range(1000):
fetch_map = client.predict(feed={"x": x}, fetch=["price"])
if fetch_map is None:
......
python/examples/imagenet/resnet50_rpc_client.py
浏览文件 @ af65b28f
......@@ -38,7 +38,8 @@ start = time.time()
image_file = "https://paddle-serving.bj.bcebos.com/imagenet-example/daisy.jpg"
for i in range(10):
img = seq(image_file)
fetch_map = client.predict(feed={"image": img}, fetch=["score"])
fetch_map = client.predict(
feed={"image": img}, fetch=["score"], batch=False)
prob = max(fetch_map["score"][0])
label = label_dict[fetch_map["score"][0].tolist().index(prob)].strip(
).replace(",", "")
......
python/examples/imdb/benchmark.py
浏览文件 @ af65b28f
此差异已折叠。
python/examples/imdb/test_client.py
浏览文件 @ af65b28f
此差异已折叠。
python/examples/lac/lac_client.py
浏览文件 @ af65b28f
此差异已折叠。
python/examples/ocr/README.md
浏览文件 @ af65b28f
此差异已折叠。
python/examples/ocr/README_CN.md
浏览文件 @ af65b28f
......@@ -34,8 +34,8 @@ python ocr_web_server.py gpu
python ocr_web_client.py
```
如果用户需要更快的执行速度，请尝试Debugger版Web服务
## 启动Debugger版Web服务
如果用户需要更快的执行速度，请尝试LocalPredictor版Web服务
## 启动LocalPredictor版Web服务
```
#根据CPU/GPU设备选择一种启动方式
#for cpu user
......@@ -60,7 +60,7 @@ GPU: Nvidia Tesla V100单卡
| engine | 客户端读图(ms) | 客户端发送请求到服务端(ms) | 服务端读图（ms） | 检测预处理耗时(ms) | 检测模型耗时(ms) | 检测后处理耗时(ms) | 识别预处理耗时(ms) | 识别模型耗时(ms) | 识别后处理耗时(ms) | 服务端回传客户端时间(ms) | 服务端整体耗时(ms) | 空跑耗时(ms) | 整体耗时（ms) |
|------------------------------|----------------|----------------------------|------------------|--------------------|------------------|--------------------|--------------------|------------------|--------------------|--------------------------|--------------------|--------------|---------------|
| Serving web service | 8.69 | 13.41 | 109.97 | 2.82 | 87.76 | 4.29 | 3.98 | 78.51 | 3.66 | 4.12 | 181.02 | 136.49 | 317.51 |
| Serving Debugger web service | 8.73 | 16.42 | 115.27 | 2.93 | 20.63 | 3.97 | 4.48 | 13.84 | 3.60 | 6.91 | 49.45 | 147.33 | 196.78 |
| Serving LocalPredictor web service | 8.73 | 16.42 | 115.27 | 2.93 | 20.63 | 3.97 | 4.48 | 13.84 | 3.60 | 6.91 | 49.45 | 147.33 | 196.78 |
## 附录： 检测/识别单服务启动
......
python/examples/ocr_detection/7.jpg 已删除 100644 → 0
浏览文件 @ bcc0c749
此差异已折叠。
python/examples/pipeline/imagenet/README.md 0 → 100644
浏览文件 @ af65b28f
此差异已折叠。
python/examples/pipeline/imagenet/README_CN.md 0 → 100644
浏览文件 @ af65b28f
此差异已折叠。
python/examples/pipeline/imagenet/config.yml 0 → 100644
浏览文件 @ af65b28f
此差异已折叠。
python/examples/pipeline/imagenet/daisy.jpg 0 → 100644
浏览文件 @ af65b28f
此差异已折叠。
python/examples/pipeline/imagenet/get_model.sh 0 → 100644
浏览文件 @ af65b28f
此差异已折叠。
python/examples/pipeline/imagenet/imagenet.label 0 → 100644
浏览文件 @ af65b28f
此差异已折叠。
python/examples/pipeline/imdb_model_ensemble/README.md 0 → 100644
浏览文件 @ af65b28f
此差异已折叠。
python/paddle_serving_app/README.md
浏览文件 @ af65b28f
此差异已折叠。
python/pipeline/__init__.py
浏览文件 @ af65b28f
此差异已折叠。
python/pipeline/analyse.py
浏览文件 @ af65b28f
此差异已折叠。
python/pipeline/channel.py
浏览文件 @ af65b28f
此差异已折叠。
python/pipeline/dag.py
浏览文件 @ af65b28f
此差异已折叠。
python/pipeline/operator.py
浏览文件 @ af65b28f
此差异已折叠。
python/pipeline/pipeline_client.py
浏览文件 @ af65b28f
此差异已折叠。
python/pipeline/pipeline_server.py
浏览文件 @ af65b28f
此差异已折叠。
python/requirements_mac.txt 0 → 100644
浏览文件 @ af65b28f
此差异已折叠。
python/setup.py.app.in
浏览文件 @ af65b28f
此差异已折叠。
python/setup.py.client.in
浏览文件 @ af65b28f
此差异已折叠。
python/setup.py.server.in
浏览文件 @ af65b28f
此差异已折叠。
python/setup.py.server_gpu.in
浏览文件 @ af65b28f
此差异已折叠。
python/util.py
浏览文件 @ af65b28f
此差异已折叠。
tools/Dockerfile.centos6.devel
浏览文件 @ af65b28f
此差异已折叠。
tools/Dockerfile.ci
浏览文件 @ af65b28f
此差异已折叠。
tools/Dockerfile.cuda10.1-cudnn7-trt6.devel 0 → 100644
浏览文件 @ af65b28f
此差异已折叠。
tools/serving_build.sh
浏览文件 @ af65b28f
此差异已折叠。
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或