def __init__(self):
self.uv_client = LocalPredictor()
self.uv_client.load_model_config(
"user_vector_model/serving_server_dir")
# milvus_host = '127.0.0.1'
# milvus_port = '19530'
self.milvus_client = RecallByMilvus()
self.collection_name = 'demo_films'
def get_client(self, concurrency_idx):
"""
Function get_client is only used for local predictor case, creates one
LocalPredictor object, and initializes the paddle predictor by function
load_model_config.The concurrency_idx is used to select running devices.
Args:
concurrency_idx: process/thread index
Returns:
_local_predictor_client
"""
#checking the legality of concurrency_idx.
device_num = len(self._devices)
if device_num <= 0:
_LOGGER.error(
"device_num must be not greater than 0. devices({})".format(
self._devices))
raise ValueError("The number of self._devices error")
if concurrency_idx < 0:
_LOGGER.error(
"concurrency_idx({}) must be one positive number".format(
concurrency_idx))
concurrency_idx = 0
elif concurrency_idx >= device_num:
concurrency_idx = concurrency_idx % device_num
_LOGGER.info("GET_CLIENT : concurrency_idx={}, device_num={}".format(
concurrency_idx, device_num))
from paddle_serving_app.local_predict import LocalPredictor
if self._local_predictor_client is None:
self._local_predictor_client = LocalPredictor()
# load model config and init predictor
self._local_predictor_client.load_model_config(
model_path=self._model_config,
use_gpu=self._use_gpu,
gpu_id=self._devices[concurrency_idx],
use_profile=self._use_profile,
thread_num=self._thread_num,
mem_optim=self._mem_optim,
ir_optim=self._ir_optim,
use_trt=self._use_trt,
use_lite=self._use_lite,
use_xpu=self._use_xpu,
precision=self._precision,
use_mkldnn=self._use_mkldnn,
mkldnn_cache_capacity=self._mkldnn_cache_capacity,
mkldnn_op_list=self._mkldnn_op_list,
mkldnn_bf16_op_list=self._mkldnn_bf16_op_list,
use_ascend_cl=self._use_ascend_cl,
min_subgraph_size=self.min_subgraph_size,
dynamic_shape_info=self.dynamic_shape_info,
use_calib=self._use_calib)
return self._local_predictor_client
def init_det_debugger(self, det_model_config):
self.det_preprocess = Sequential([
ResizeByFactor(32, 960), Div(255),
Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]), Transpose(
(2, 0, 1))
])
self.det_client = LocalPredictor()
if sys.argv[1] == 'gpu':
self.det_client.load_model_config(
det_model_config, use_gpu=True, gpu_id=0)
elif sys.argv[1] == 'cpu':
self.det_client.load_model_config(det_model_config)
self.ocr_reader = OCRReader(
char_dict_path="../../../ppocr/utils/ppocr_keys_v1.txt")
def _launch_local_predictor(self, gpu):
# actually, LocalPredictor is like a server, but it is WebService Request initiator
# for WebService it is a Client.
# local_predictor only support single-Model DirPath - Type:str
# so the input must be self.server_config_dir_paths[0]
from paddle_serving_app.local_predict import LocalPredictor
self.client = LocalPredictor()
if gpu:
# if user forget to call function `set_gpus` to set self.gpus.
# default self.gpus = [0].
if len(self.gpus) == 0 or self.gpus == ["-1"]:
self.gpus = ["0"]
# right now, local Predictor only support 1 card.
# no matter how many gpu_id is in gpus, we only use the first one.
gpu_id = (self.gpus[0].split(","))[0]
self.client.load_model_config(self.server_config_dir_paths[0],
use_gpu=True,
gpu_id=gpu_id)
else:
self.client.load_model_config(self.server_config_dir_paths[0],
use_gpu=False)
# 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.
from paddle_serving_app.reader import Sequential, File2Image, Resize, CenterCrop
from paddle_serving_app.reader import RGB2BGR, Transpose, Div, Normalize
from paddle_serving_app.local_predict import LocalPredictor
import sys
debugger = LocalPredictor()
debugger.load_model_config(sys.argv[1], gpu=True)
seq = Sequential([
File2Image(),
Resize(256),
CenterCrop(224),
RGB2BGR(),
Transpose((2, 0, 1)),
Div(255),
Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225], True)
])
image_file = "daisy.jpg"
img = seq(image_file)
fetch_map = debugger.predict(feed={"image": img}, fetch=["feature_map"])
class OCRService(WebService):
def init_det_debugger(self, det_model_config):
self.det_preprocess = Sequential([
ResizeByFactor(32, 960), Div(255),
Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]), Transpose(
(2, 0, 1))
])
self.det_client = LocalPredictor()
if sys.argv[1] == 'gpu':
self.det_client.load_model_config(
det_model_config, use_gpu=True, gpu_id=0)
elif sys.argv[1] == 'cpu':
self.det_client.load_model_config(det_model_config)
self.ocr_reader = OCRReader(
char_dict_path="../../../ppocr/utils/ppocr_keys_v1.txt")
def preprocess(self, feed=[], fetch=[]):
data = base64.b64decode(feed[0]["image"].encode('utf8'))
data = np.fromstring(data, np.uint8)
im = cv2.imdecode(data, cv2.IMREAD_COLOR)
ori_h, ori_w, _ = im.shape
det_img = self.det_preprocess(im)
_, new_h, new_w = det_img.shape
det_img = det_img[np.newaxis, :]
det_img = det_img.copy()
det_out = self.det_client.predict(
feed={"x": det_img}, fetch=["save_infer_model/scale_0.tmp_1"], batch=True)
filter_func = FilterBoxes(10, 10)
post_func = DBPostProcess({
"thresh": 0.3,
"box_thresh": 0.5,
"max_candidates": 1000,
"unclip_ratio": 1.5,
"min_size": 3
})
sorted_boxes = SortedBoxes()
ratio_list = [float(new_h) / ori_h, float(new_w) / ori_w]
dt_boxes_list = post_func(det_out["save_infer_model/scale_0.tmp_1"], [ratio_list])
dt_boxes = filter_func(dt_boxes_list[0], [ori_h, ori_w])
dt_boxes = sorted_boxes(dt_boxes)
get_rotate_crop_image = GetRotateCropImage()
img_list = []
max_wh_ratio = 0
for i, dtbox in enumerate(dt_boxes):
boximg = get_rotate_crop_image(im, dt_boxes[i])
img_list.append(boximg)
h, w = boximg.shape[0:2]
wh_ratio = w * 1.0 / h
max_wh_ratio = max(max_wh_ratio, wh_ratio)
if len(img_list) == 0:
return [], []
_, w, h = self.ocr_reader.resize_norm_img(img_list[0],
max_wh_ratio).shape
imgs = np.zeros((len(img_list), 3, w, h)).astype('float32')
for id, img in enumerate(img_list):
norm_img = self.ocr_reader.resize_norm_img(img, max_wh_ratio)
imgs[id] = norm_img
feed = {"x": imgs.copy()}
fetch = ["save_infer_model/scale_0.tmp_1"]
return feed, fetch, True
def postprocess(self, feed={}, fetch=[], fetch_map=None):
rec_res = self.ocr_reader.postprocess(fetch_map, with_score=True)
res_lst = []
for res in rec_res:
res_lst.append(res[0])
res = {"res": res_lst}
return res
# 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.
from paddle_serving_app.reader import Sequential, File2Image, Resize, CenterCrop
from paddle_serving_app.reader import RGB2BGR, Transpose, Div, Normalize
from paddle_serving_app.local_predict import LocalPredictor
import sys
predictor = LocalPredictor()
predictor.load_model_config(sys.argv[1],
use_lite=True,
use_xpu=True,
ir_optim=True)
seq = Sequential([
File2Image(),
Resize(256),
CenterCrop(224),
RGB2BGR(),
Transpose((2, 0, 1)),
Div(255),
Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225], True)
])
class LocalServiceHandler(object):
"""
LocalServiceHandler is the processor of the local service, contains
three client types, brpc, grpc and local_predictor.If you use the
brpc or grpc, serveing startup ability is provided.If you use
local_predictor, local predict ability is provided by paddle_serving_app.
"""
def __init__(self,
model_config,
client_type='local_predictor',
workdir="",
thread_num=2,
device_type=-1,
devices="",
fetch_names=None,
mem_optim=True,
ir_optim=False,
available_port_generator=None,
use_profile=False,
precision="fp32",
use_mkldnn=False,
mkldnn_cache_capacity=0,
mkldnn_op_list=None,
mkldnn_bf16_op_list=None,
min_subgraph_size=3,
dynamic_shape_info={},
use_calib=False):
"""
Initialization of localservicehandler
Args:
model_config: model config path
client_type: brpc, grpc and local_predictor[default]
workdir: work directory
thread_num: number of threads, concurrent quantity.
device_type: support multiple devices. -1=Not set, determined by
`devices`. 0=cpu, 1=gpu, 2=tensorRT, 3=arm cpu, 4=kunlun xpu
devices: gpu id list[gpu], "" default[cpu]
fetch_names: get fetch names out of LocalServiceHandler in
local_predictor mode. fetch_names_ is compatible for Client().
mem_optim: use memory/graphics memory optimization, True default.
ir_optim: use calculation chart optimization, False default.
available_port_generator: generate available ports
use_profile: use profiling, False default.
precision: inference precesion, e.g. "fp32", "fp16", "int8"
use_mkldnn: use mkldnn, default False.
mkldnn_cache_capacity: cache capacity of mkldnn, 0 means no limit.
mkldnn_op_list: OP list optimized by mkldnn, None default.
mkldnn_bf16_op_list: OP list optimized by mkldnn bf16, None default.
use_calib: set inference use_calib_mode param, False default.
Returns:
None
"""
if available_port_generator is None:
available_port_generator = util.GetAvailablePortGenerator()
self._model_config = model_config
self._port_list = []
self._device_name = "cpu"
self._use_gpu = False
self._use_trt = False
self._use_lite = False
self._use_xpu = False
self._use_ascend_cl = False
self._use_mkldnn = False
self._mkldnn_cache_capacity = 0
self._mkldnn_op_list = None
self._mkldnn_bf16_op_list = None
self.min_subgraph_size = 3
self.dynamic_shape_info = {}
self._use_calib = False
if device_type == -1:
# device_type is not set, determined by `devices`,
if devices == "":
# CPU
self._device_name = "cpu"
devices = [-1]
else:
# GPU
self._device_name = "gpu"
self._use_gpu = True
devices = [int(x) for x in devices.split(",")]
elif device_type == 0:
# CPU
self._device_name = "cpu"
devices = [-1]
elif device_type == 1:
# GPU
self._device_name = "gpu"
self._use_gpu = True
devices = [int(x) for x in devices.split(",")]
elif device_type == 2:
# Nvidia Tensor RT
self._device_name = "gpu"
self._use_gpu = True
devices = [int(x) for x in devices.split(",")]
self._use_trt = True
self.min_subgraph_size = min_subgraph_size
self.dynamic_shape_info = dynamic_shape_info
elif device_type == 3:
# ARM CPU
self._device_name = "arm"
devices = [-1]
self._use_lite = True
elif device_type == 4:
# Kunlun XPU
self._device_name = "arm"
devices = [int(x) for x in devices.split(",")]
self._use_lite = True
self._use_xpu = True
elif device_type == 5:
# Ascend 310 ARM CPU
self._device_name = "arm"
devices = [int(x) for x in devices.split(",")]
self._use_lite = True
self._use_ascend_cl = True
elif device_type == 6:
# Ascend 910 ARM CPU
self._device_name = "arm"
devices = [int(x) for x in devices.split(",")]
self._use_ascend_cl = True
else:
_LOGGER.error(
"LocalServiceHandler initialization fail. device_type={}".
format(device_type))
if client_type == "brpc" or client_type == "grpc":
for _ in devices:
self._port_list.append(available_port_generator.next())
_LOGGER.info("Create ports for devices:{}. Port:{}".format(
devices, self._port_list))
self._client_type = client_type
self._workdir = workdir
self._devices = devices
self._thread_num = thread_num
self._mem_optim = mem_optim
self._ir_optim = ir_optim
self._local_predictor_client = None
self._rpc_service_list = []
self._server_pros = []
self._use_profile = use_profile
self._fetch_names = fetch_names
self._precision = precision
self._use_mkldnn = use_mkldnn
self._mkldnn_cache_capacity = mkldnn_cache_capacity
self._mkldnn_op_list = mkldnn_op_list
self._mkldnn_bf16_op_list = mkldnn_bf16_op_list
self._use_calib = use_calib
_LOGGER.info(
"Models({}) will be launched by device {}. use_gpu:{}, "
"use_trt:{}, use_lite:{}, use_xpu:{}, device_type:{}, devices:{}, "
"mem_optim:{}, ir_optim:{}, use_profile:{}, thread_num:{}, "
"client_type:{}, fetch_names:{}, precision:{}, use_calib:{}, "
"use_mkldnn:{}, mkldnn_cache_capacity:{}, mkldnn_op_list:{}, "
"mkldnn_bf16_op_list:{}, use_ascend_cl:{}, min_subgraph_size:{},"
"is_set_dynamic_shape_info:{}".format(
model_config, self._device_name, self._use_gpu, self._use_trt,
self._use_lite, self._use_xpu, device_type, self._devices,
self._mem_optim, self._ir_optim, self._use_profile,
self._thread_num, self._client_type, self._fetch_names,
self._precision, self._use_calib, self._use_mkldnn,
self._mkldnn_cache_capacity, self._mkldnn_op_list,
self._mkldnn_bf16_op_list, self._use_ascend_cl,
self.min_subgraph_size, bool(len(self.dynamic_shape_info))))
def get_fetch_list(self):
return self._fetch_names
def get_port_list(self):
return self._port_list
def get_client(self, concurrency_idx):
"""
Function get_client is only used for local predictor case, creates one
LocalPredictor object, and initializes the paddle predictor by function
load_model_config.The concurrency_idx is used to select running devices.
Args:
concurrency_idx: process/thread index
Returns:
_local_predictor_client
"""
#checking the legality of concurrency_idx.
device_num = len(self._devices)
if device_num <= 0:
_LOGGER.error(
"device_num must be not greater than 0. devices({})".format(
self._devices))
raise ValueError("The number of self._devices error")
if concurrency_idx < 0:
_LOGGER.error(
"concurrency_idx({}) must be one positive number".format(
concurrency_idx))
concurrency_idx = 0
elif concurrency_idx >= device_num:
concurrency_idx = concurrency_idx % device_num
_LOGGER.info("GET_CLIENT : concurrency_idx={}, device_num={}".format(
concurrency_idx, device_num))
from paddle_serving_app.local_predict import LocalPredictor
if self._local_predictor_client is None:
self._local_predictor_client = LocalPredictor()
# load model config and init predictor
self._local_predictor_client.load_model_config(
model_path=self._model_config,
use_gpu=self._use_gpu,
gpu_id=self._devices[concurrency_idx],
use_profile=self._use_profile,
thread_num=self._thread_num,
mem_optim=self._mem_optim,
ir_optim=self._ir_optim,
use_trt=self._use_trt,
use_lite=self._use_lite,
use_xpu=self._use_xpu,
precision=self._precision,
use_mkldnn=self._use_mkldnn,
mkldnn_cache_capacity=self._mkldnn_cache_capacity,
mkldnn_op_list=self._mkldnn_op_list,
mkldnn_bf16_op_list=self._mkldnn_bf16_op_list,
use_ascend_cl=self._use_ascend_cl,
min_subgraph_size=self.min_subgraph_size,
dynamic_shape_info=self.dynamic_shape_info,
use_calib=self._use_calib)
return self._local_predictor_client
def get_client_config(self):
return os.path.join(self._model_config, "serving_server_conf.prototxt")
def _prepare_one_server(self, workdir, port, gpuid, thread_num, mem_optim,
ir_optim, precision):
"""
According to self._device_name, generating one Cpu/Gpu/Arm Server, and
setting the model config amd startup params.
Args:
workdir: work directory
port: network port
gpuid: gpu id
thread_num: thread num
mem_optim: use memory/graphics memory optimization
ir_optim: use calculation chart optimization
precision: inference precison, e.g."fp32", "fp16", "int8"
Returns:
server: CpuServer/GpuServer
"""
if self._device_name == "cpu":
from paddle_serving_server import OpMaker, OpSeqMaker, Server
op_maker = OpMaker()
read_op = op_maker.create('general_reader')
general_infer_op = op_maker.create('general_infer')
general_response_op = op_maker.create('general_response')
op_seq_maker = OpSeqMaker()
op_seq_maker.add_op(read_op)
op_seq_maker.add_op(general_infer_op)
op_seq_maker.add_op(general_response_op)
server = Server()
else:
#gpu or arm
from paddle_serving_server import OpMaker, OpSeqMaker, Server
op_maker = OpMaker()
read_op = op_maker.create('general_reader')
general_infer_op = op_maker.create('general_infer')
general_response_op = op_maker.create('general_response')
op_seq_maker = OpSeqMaker()
op_seq_maker.add_op(read_op)
op_seq_maker.add_op(general_infer_op)
op_seq_maker.add_op(general_response_op)
server = Server()
if gpuid >= 0:
server.set_gpuid(gpuid)
# TODO: support arm or arm + xpu later
server.set_device(self._device_name)
if self._use_xpu:
server.set_xpu()
if self._use_lite:
server.set_lite()
if self._use_ascend_cl:
server.set_ascend_cl()
server.set_op_sequence(op_seq_maker.get_op_sequence())
server.set_num_threads(thread_num)
server.set_memory_optimize(mem_optim)
server.set_ir_optimize(ir_optim)
server.set_precision(precision)
server.load_model_config(self._model_config)
server.prepare_server(workdir=workdir,
port=port,
device=self._device_name)
if self._fetch_names is None:
self._fetch_names = server.get_fetch_list()
return server
def _start_one_server(self, service_idx):
"""
Start one server
Args:
service_idx: server index
Returns:
None
"""
self._rpc_service_list[service_idx].run_server()
def prepare_server(self):
"""
Prepare all servers to be started, and append them into list.
"""
for i, device_id in enumerate(self._devices):
if self._workdir != "":
workdir = "{}_{}".format(self._workdir, i)
else:
workdir = _workdir_name_gen.next()
self._rpc_service_list.append(
self._prepare_one_server(workdir,
self._port_list[i],
device_id,
thread_num=self._thread_num,
mem_optim=self._mem_optim,
ir_optim=self._ir_optim,
precision=self._precision))
def start_server(self):
"""
Start multiple processes and start one server in each process
"""
for i, _ in enumerate(self._rpc_service_list):
p = multiprocessing.Process(target=self._start_one_server,
args=(i, ))
p.daemon = True
self._server_pros.append(p)
for p in self._server_pros:
p.start()
class RecallServerServicer(object):
def __init__(self):
self.uv_client = LocalPredictor()
self.uv_client.load_model_config(
"user_vector_model/serving_server_dir")
# milvus_host = '127.0.0.1'
# milvus_port = '19530'
self.milvus_client = RecallByMilvus()
self.collection_name = 'demo_films'
def get_user_vector(self, user_info):
dic = {"userid": [], "gender": [], "age": [], "occupation": []}
lod = [0]
dic["userid"].append(hash2(user_info.user_id))
dic["gender"].append(hash2(user_info.gender))
dic["age"].append(hash2(user_info.age))
dic["occupation"].append(hash2(user_info.job))
lod.append(1)
dic["userid.lod"] = lod
dic["gender.lod"] = lod
dic["age.lod"] = lod
dic["occupation.lod"] = lod
for key in dic:
dic[key] = np.array(dic[key]).astype(np.int64).reshape(
len(dic[key]), 1)
fetch_map = self.uv_client.predict(
feed=dic, fetch=["save_infer_model/scale_0.tmp_0"], batch=True)
return fetch_map["save_infer_model/scale_0.tmp_0"].tolist()[0]
def recall(self, request, context):
'''
message RecallRequest{
string log_id = 1;
user_info.UserInfo user_info = 2;
string recall_type= 3;
uint32 request_num= 4;
}
message RecallResponse{
message Error {
uint32 code = 1;
string text = 2;
}
message ScorePair {
string nid = 1;
float score = 2;
};
Error error = 1;
repeated ScorePair score_pairs = 2;
}
'''
recall_res = recall_pb2.RecallResponse()
user_vector = self.get_user_vector(request.user_info)
status, results = self.milvus_client.search(
collection_name=self.collection_name,
vectors=[user_vector],
partition_tag="Movie")
for entities in results:
if len(entities) == 0:
recall_res.error.code = 500
recall_res.error.text = "Recall server get milvus fail. ({})".format(
str(request))
return recall_res
for topk_film in entities:
# current_entity = topk_film.entity
score_pair = recall_res.score_pairs.add()
score_pair.nid = str(topk_film.id)
score_pair.score = float(topk_film.distance)
recall_res.error.code = 200
return recall_res
import redis
import numpy as np
import codecs
class Movie(object):
def __init__(self):
self.movie_id, self.title, self.genres = "", "", ""
pass
def hash2(a):
return hash(a) % 60000000
ctr_client = LocalPredictor()
ctr_client.load_model_config("serving_server_dir")
with codecs.open("movies.dat", "r", encoding='utf-8', errors='ignore') as f:
lines = f.readlines()
ff = open("movie_vectors.txt", 'w')
for line in lines:
if len(line.strip()) == 0:
continue
tmp = line.strip().split("::")
movie_id = tmp[0]
title = tmp[1]
genre_group = tmp[2]
tmp = genre_group.strip().split("|")
def __init__(self):
self.ctr_client = LocalPredictor()
self.ctr_client.load_model_config("rank_model")
class RankServerServicer(object):
def __init__(self):
self.ctr_client = LocalPredictor()
self.ctr_client.load_model_config("rank_model")
def process_feed_dict(self, user_info, item_infos):
#" userid gender age occupation | movieid title genres"
dic = {
"userid": [],
"gender": [],
"age": [],
"occupation": [],
"movieid": [],
"title": [],
"genres": []
}
batch_size = len(item_infos)
lod = [0]
for i, item_info in enumerate(item_infos):
dic["movieid"].append(hash2(item_info.movie_id))
dic["title"].append(hash2(item_info.title))
dic["genres"].append(hash2(item_info.genre))
dic["userid"].append(hash2(user_info.user_id))
dic["gender"].append(hash2(user_info.gender))
dic["age"].append(hash2(user_info.age))
dic["occupation"].append(hash2(user_info.job))
lod.append(i + 1)
dic["movieid.lod"] = lod
dic["title.lod"] = lod
dic["genres.lod"] = lod
dic["userid.lod"] = lod
dic["gender.lod"] = lod
dic["age.lod"] = lod
dic["occupation.lod"] = lod
for key in dic:
dic[key] = np.array(dic[key]).astype(np.int64).reshape(
len(dic[key]), 1)
return dic
def rank_predict(self, request, context):
'''
message RankRequest {
string log_id = 1;
user_info.UserInfo user_info = 2;
repeated item_info.ItemInfo item_infos = 3;
}
message RankResponse {
message Error {
uint32 code = 1;
string text = 2;
}
message ScorePair {
string nid = 1;
float score = 2;
};
Error error = 1;
repeated ScorePair score_pairs = 2;
};
'''
batch_size = len(request.item_infos)
dic = self.process_feed_dict(request.user_info, request.item_infos)
fetch_map = self.ctr_client.predict(
feed=dic, fetch=["save_infer_model/scale_0.tmp_0"], batch=True)
response = rank_pb2.RankResponse()
#raise ValueError("UM server get user_info from redis fail. ({})".format(str(request)))
response.error.code = 200
for i in range(batch_size):
score_pair = response.score_pairs.add()
score_pair.nid = request.item_infos[i].movie_id
score_pair.score = fetch_map["save_infer_model/scale_0.tmp_0"][i][
0]
response.score_pairs.sort(reverse=True, key=lambda item: item.score)
return response
class WebService(object):
def __init__(self, name="default_service"):
self.name = name
# pipeline
self._server = pipeline.PipelineServer(self.name)
self.gpus = ["-1"] # deprecated
self.rpc_service_list = [] # deprecated
def get_pipeline_response(self, read_op):
return None
def prepare_pipeline_config(self, yml_file=None, yml_dict=None):
# build dag
read_op = pipeline.RequestOp()
last_op = self.get_pipeline_response(read_op)
if not isinstance(last_op, Op):
raise ValueError(
"The return value type of `get_pipeline_response` "
"function is not Op type, please check function "
"`get_pipeline_response`.")
response_op = pipeline.ResponseOp(input_ops=[last_op])
self._server.set_response_op(response_op)
self._server.prepare_server(yml_file=yml_file, yml_dict=yml_dict)
def run_service(self):
self._server.run_server()
def load_model_config(self,
server_config_dir_paths,
client_config_path=None):
if isinstance(server_config_dir_paths, str):
server_config_dir_paths = [server_config_dir_paths]
elif isinstance(server_config_dir_paths, list):
pass
for single_model_config in server_config_dir_paths:
if os.path.isdir(single_model_config):
pass
elif os.path.isfile(single_model_config):
raise ValueError(
"The input of --model should be a dir not file.")
self.server_config_dir_paths = server_config_dir_paths
from .proto import general_model_config_pb2 as m_config
import google.protobuf.text_format
file_path_list = []
for single_model_config in self.server_config_dir_paths:
file_path_list.append(
"{}/serving_server_conf.prototxt".format(single_model_config))
model_conf = m_config.GeneralModelConfig()
f = open(file_path_list[0], 'r')
model_conf = google.protobuf.text_format.Merge(str(f.read()),
model_conf)
self.feed_vars = {var.alias_name: var for var in model_conf.feed_var}
if len(file_path_list) > 1:
model_conf = m_config.GeneralModelConfig()
f = open(file_path_list[-1], 'r')
model_conf = google.protobuf.text_format.Merge(
str(f.read()), model_conf)
self.fetch_vars = {var.alias_name: var for var in model_conf.fetch_var}
if client_config_path == None:
self.client_config_path = file_path_list
# after this function, self.gpus should be a list of str or [].
def set_gpus(self, gpus):
print("This API will be deprecated later. Please do not use it")
self.gpus = format_gpu_to_strlist(gpus)
# this function can be called by user
# or by Function create_rpc_config
# if by user, user can set_gpus or pass the `gpus`
# if `gpus` == None, which means it`s not set at all.
# at this time, we should use self.gpus instead.
# otherwise, we should use the `gpus` first.
# which means if set_gpus and `gpus` is both set.
# `gpus` will be used.
def default_rpc_service(self,
workdir,
port=9292,
gpus=None,
thread_num=4,
mem_optim=True,
use_lite=False,
use_xpu=False,
ir_optim=False,
precision="fp32",
use_calib=False,
use_trt=False,
gpu_multi_stream=False,
runtime_thread_num=None,
batch_infer_size=None):
device = "cpu"
server = Server()
# only when `gpus == None`, which means it`s not set at all
# we will use the self.gpus.
if gpus == None:
gpus = self.gpus
gpus = format_gpu_to_strlist(gpus)
server.set_gpuid(gpus)
if len(gpus) == 0 or gpus == ["-1"]:
if use_lite:
device = "arm"
else:
device = "cpu"
else:
device = "gpu"
op_maker = OpMaker()
op_seq_maker = OpSeqMaker()
read_op = op_maker.create('GeneralReaderOp')
op_seq_maker.add_op(read_op)
for idx, single_model in enumerate(self.server_config_dir_paths):
infer_op_name = "GeneralInferOp"
if len(self.server_config_dir_paths) == 2 and idx == 0:
infer_op_name = "GeneralDetectionOp"
else:
infer_op_name = "GeneralInferOp"
general_infer_op = op_maker.create(infer_op_name)
op_seq_maker.add_op(general_infer_op)
general_response_op = op_maker.create('GeneralResponseOp')
op_seq_maker.add_op(general_response_op)
server.set_op_sequence(op_seq_maker.get_op_sequence())
server.set_num_threads(thread_num)
server.set_memory_optimize(mem_optim)
server.set_ir_optimize(ir_optim)
server.set_device(device)
server.set_precision(precision)
server.set_use_calib(use_calib)
if use_trt and device == "gpu":
server.set_trt()
server.set_ir_optimize(True)
if gpu_multi_stream and device == "gpu":
server.set_gpu_multi_stream()
if runtime_thread_num:
server.set_runtime_thread_num(runtime_thread_num)
if batch_infer_size:
server.set_batch_infer_size(batch_infer_size)
if use_lite:
server.set_lite()
if use_xpu:
server.set_xpu()
server.load_model_config(
self.server_config_dir_paths
) #brpc Server support server_config_dir_paths
server.prepare_server(workdir=workdir, port=port, device=device)
return server
def _launch_rpc_service(self, service_idx):
self.rpc_service_list[service_idx].run_server()
# if use this function, self.gpus must be set before.
# if not, we will use the default value, self.gpus = ["-1"].
# so we always pass the `gpus` = self.gpus.
def create_rpc_config(self):
self.rpc_service_list.append(
self.default_rpc_service(
self.workdir,
self.port_list[0],
self.gpus,
thread_num=self.thread_num,
mem_optim=self.mem_optim,
use_lite=self.use_lite,
use_xpu=self.use_xpu,
ir_optim=self.ir_optim,
precision=self.precision,
use_calib=self.use_calib,
use_trt=self.use_trt,
gpu_multi_stream=self.gpu_multi_stream,
runtime_thread_num=self.runtime_thread_num,
batch_infer_size=self.batch_infer_size))
def prepare_server(self,
workdir,
port=9393,
device="cpu",
precision="fp32",
use_calib=False,
use_lite=False,
use_xpu=False,
ir_optim=False,
thread_num=4,
mem_optim=True,
use_trt=False,
gpu_multi_stream=False,
runtime_thread_num=None,
batch_infer_size=None,
gpuid=None):
print("This API will be deprecated later. Please do not use it")
self.workdir = workdir
self.port = port
self.thread_num = thread_num
# self.device is not used at all.
# device is set by gpuid.
self.precision = precision
self.use_calib = use_calib
self.use_lite = use_lite
self.use_xpu = use_xpu
self.ir_optim = ir_optim
self.mem_optim = mem_optim
self.port_list = []
self.use_trt = use_trt
self.gpu_multi_stream = gpu_multi_stream
self.runtime_thread_num = runtime_thread_num
self.batch_infer_size = batch_infer_size
# record port and pid info for stopping process
dump_pid_file([self.port], "web_service")
# if gpuid != None, we will use gpuid first.
# otherwise, keep the self.gpus unchanged.
# maybe self.gpus is set by the Function set_gpus.
if gpuid != None:
self.gpus = format_gpu_to_strlist(gpuid)
else:
pass
default_port = 12000
for i in range(1000):
if port_is_available(default_port + i):
self.port_list.append(default_port + i)
break
def _launch_web_service(self):
self.client = Client()
self.client.load_client_config(self.client_config_path)
endpoints = ""
endpoints = "127.0.0.1:{}".format(self.port_list[0])
self.client.connect([endpoints])
def get_prediction(self, request):
if not request.json:
abort(400)
if "fetch" not in request.json:
abort(400)
try:
feed, fetch, is_batch = self.preprocess(request.json["feed"],
request.json["fetch"])
if isinstance(feed, dict) and "fetch" in feed:
del feed["fetch"]
if len(feed) == 0:
raise ValueError("empty input")
fetch_map = self.client.predict(feed=feed,
fetch=fetch,
batch=is_batch)
result = self.postprocess(feed=request.json["feed"],
fetch=fetch,
fetch_map=fetch_map)
result = {"result": result}
except ValueError as err:
result = {"result": str(err)}
return result
def run_rpc_service(self):
print("This API will be deprecated later. Please do not use it")
import socket
localIP = socket.gethostbyname(socket.gethostname())
print("web service address:")
print("http://{}:{}/{}/prediction".format(localIP, self.port,
self.name))
server_pros = []
self.create_rpc_config()
for i, service in enumerate(self.rpc_service_list):
p = Process(target=self._launch_rpc_service, args=(i, ))
server_pros.append(p)
for p in server_pros:
p.start()
app_instance = Flask(__name__)
@app_instance.before_first_request
def init():
self._launch_web_service()
service_name = "/" + self.name + "/prediction"
@app_instance.route(service_name, methods=["POST"])
def run():
return self.get_prediction(request)
self.app_instance = app_instance
# TODO: maybe change another API name: maybe run_local_predictor?
def run_debugger_service(self, gpu=False):
print("This API will be deprecated later. Please do not use it")
import socket
localIP = socket.gethostbyname(socket.gethostname())
print("web service address:")
print("http://{}:{}/{}/prediction".format(localIP, self.port,
self.name))
app_instance = Flask(__name__)
@app_instance.before_first_request
def init():
self._launch_local_predictor(gpu)
service_name = "/" + self.name + "/prediction"
@app_instance.route(service_name, methods=["POST"])
def run():
return self.get_prediction(request)
self.app_instance = app_instance
def _launch_local_predictor(self, gpu):
# actually, LocalPredictor is like a server, but it is WebService Request initiator
# for WebService it is a Client.
# local_predictor only support single-Model DirPath - Type:str
# so the input must be self.server_config_dir_paths[0]
from paddle_serving_app.local_predict import LocalPredictor
self.client = LocalPredictor()
if gpu:
# if user forget to call function `set_gpus` to set self.gpus.
# default self.gpus = [0].
if len(self.gpus) == 0 or self.gpus == ["-1"]:
self.gpus = ["0"]
# right now, local Predictor only support 1 card.
# no matter how many gpu_id is in gpus, we only use the first one.
gpu_id = (self.gpus[0].split(","))[0]
self.client.load_model_config(self.server_config_dir_paths[0],
use_gpu=True,
gpu_id=gpu_id)
else:
self.client.load_model_config(self.server_config_dir_paths[0],
use_gpu=False)
def run_web_service(self):
print("This API will be deprecated later. Please do not use it")
self.app_instance.run(host="0.0.0.0", port=self.port, threaded=True)
def get_app_instance(self):
return self.app_instance
def preprocess(self, feed=[], fetch=[]):
print("This API will be deprecated later. Please do not use it")
is_batch = True
feed_dict = {}
for var_name in self.feed_vars.keys():
feed_dict[var_name] = []
for feed_ins in feed:
for key in feed_ins:
feed_dict[key].append(
np.array(feed_ins[key]).reshape(
list(self.feed_vars[key].shape))[np.newaxis, :])
feed = {}
for key in feed_dict:
feed[key] = np.concatenate(feed_dict[key], axis=0)
return feed, fetch, is_batch
def postprocess(self, feed=[], fetch=[], fetch_map=None):
print("This API will be deprecated later. Please do not use it")
for key in fetch_map:
fetch_map[key] = fetch_map[key].tolist()
return fetch_map
def _launch_web_service(self):
self.client = Client()
self.client.load_client_config(self.client_config_path)
endpoints = ""
endpoints = "127.0.0.1:{}".format(self.port_list[0])
self.client.connect([endpoints])