一、基于OpenVINO的“天空分割”回顾
semantic-segmentation-adas-0001模型中包含了天空对象。
只接受batch=1的输入，而它的输出直接是标注label，需要缩放成为原图大小，可能还需要进行一些轮廓处理-但是已经基本上实现了“端到端”的效果。下面的图像中，蓝色区域是天空区域。这里需要注意的是，接口文件(***.py)是需要自己来写的。
The net outputs a blob with the shape [B, H=1024, W=2048]. It can be treated as a one-channel feature map, where each pixel is a label of one of the classes.
原图 | 效果 |
|
|
|
|
|
|
|
|
从上面几张图的效果可以看出，虽然有一定的误差，但是还是在可以接受范围内的。只要我们在融合上面稍微下一点功夫，这些是看不出来的。
经过进一步研究，能够得到以下的“天空替换”结果
二、OpenVINO Model Server服务化要点
最容易出错的地方是 模型文件的准备 ，目前已经验证可行的方法是在本机按照制定的结构安排文件，而后调用“:ro"参数，将文件结构全部复制到docker中。比如：
我们下载了bin+xml,需要 按照以下模式存放
/
models /
├── model1
│ ├── 1
│ │ ├── ir_model.bin
│ │ └── ir_model.xml
│ └── 2
│ ├── ir_model.bin
│ └── ir_model.xml
└── model2
└── 1
├── ir_model.bin
├── ir_model.xml
└── mapping_config.json
这里的models以及下面的级联文件夹，都是在本机创建好的。
而后调用类似下面的命令行，启动Docker
-v /models : /models :ro -p 9000 :
9000 openvino
/model_server
:latest
--model_path
/models
/model1
--model_name face
-detection
--port
9000
--log_level DEBUG
--shape
auto
参数解释
-v 表示的是本机和docker中目录的对应方式， :ro表示是嵌套复制，也就是前面那么多级联的目录”原模原样“的复制过去。本机的文件放在哪里，我们当然知道；docker中的文件放在哪里，其实并不重要。重要的是将这里的文件地址告诉openvino，所以这里的目录地址和后面的 --model_path是一致的 -p 本机和docker的端口镜像关系
openvino /model_server :latest 启动的docker镜像 --model_path 和前面的 -v要保持一致 --model_name openvino调用的model的名称
-d 它的意思就是后台运行，你可以去掉来看调试
其它几个不是太重要， 也不容易写错。
启动成功以后，可以运行
docker ps
来看是否运行成功。
当然你也可以在docker run中去掉 -d 而基于命令行的方法查看，这里还有其他一些相关命令。
ps
sudo docker exec -it 775c7c9ee1e1 /bin /bash
三、基于OpenVINO的道路分割服务化部署
3.1 新建model2,将最新的模型下载下来
: / /download. 01.org /opencv / 2021 /openvinotoolkit /
2021.
1
/open_model_zoo
/models_bin
/
2
/semantic
-segmentation
-adas
-
0001
/FP32
/semantic
-segmentation
-adas
-
0001.bin
wget https : / /download. 01.org /opencv / 2021 /openvinotoolkit
/
2021.
1
/open_model_zoo
/models_bin
/
2
/semantic
-segmentation
-adas
-
0001
/FP32
/semantic
-segmentation
-adas
-
0001.xml
[root@VM - 0 - 13 -centos 1] # cd /models
[root@VM - 0 - 13 -centos models] # tree
.
├── model1
│ └── 1
│ ├── face -detection -retail - 0004.bin
│ └── face -detection -retail - 0004.xml
└── model2
└── 1
├── semantic -segmentation -adas - 0001.bin
└── semantic -segmentation -adas - 0001.xml 4 directories, 4
同时进入image中，将一个图片下载下来
- 0 - 13 -centos images] # wget https://docs.openvinotoolkit.org/2019_R1.1/road-segmentation-adas-0001.png -- 2020 -
10
-
12
19
:
42
:
11
-- https
:
/
/docs.openvinotoolkit.org
/
2019_R1.
1
/road
-segmentation
-adas
-
0001.png
Resolving docs.openvinotoolkit.org (docs.openvinotoolkit.org)... 118. 215. 180. 232, 2600 : 1417 :
76
:
487
:
:
4b21,
2600
:
1417
:
76
:
480
:
:
4b21
Connecting to docs.openvinotoolkit.org (docs.openvinotoolkit.org) | 118. 215. 180. 232 | : 443... connected.
HTTP request sent, awaiting response... 200 OK
Length : 498344 ( 487K) [image /png]
Saving to : ‘road -segmentation -adas - 0001.png’
road -segmentation -adas - 0001.p 100 %[ == ==
==
==
==
==
==
==
==
==
==
==
==
==
==
==
==
==
==
==
==
==
==
>]
486.
66K
219KB
/s
in 2. 2s 2020 - 10 - 12 19
:
42
:
16 (
219 KB
/s)
- ‘road
-segmentation
-adas
-
0001.png’ saved [
498344
/
498344]
[root@VM - 0 - 13 -centos images] # ll
total 696 -rw -r --r -- 1 root root 210765 Oct 11
06
:
44 people1.jpeg
-rw
-r
--r
--
1 root root
498344 Dec
5
2019 road
-segmentation
-adas
-
0001.png
3.2修改几个参数，将服务跑起来：
- 0 - 13 -centos models] # docker run -d -v /models:/models:ro -p 9000:9000 openvino/model_server:latest --model_path /models/ model2
--model_name
semantic-segmentation-adas
--port 9000 --log_level DEBUG --shape auto
27907ca99807fb58184daee3439d821b554199ead70964e6e6bcf233c7ee20f0
[root@VM - 0 - 13 -centos models] # docker ps
CONTAINER ID IMAGE COMMAND CREATED STATUS PORTS NAMES 27907ca99807 openvino /model_server :latest "/ovms/bin/ovms --mo…" 5 seconds ago Up 3 seconds 0. 0.
0.
0
:
9000
-
>
9000
/tcp flamboyant_mahavira
3.3最为困难的是接口文件的编写
在目前情况下，如果直接改写client文件的话，会出现以下问题：
1, 3, 1024, 2048)
( 1, 3, 1024, 2048)
Traceback (most recent call last) :
File "sky_detection.py", line 79, in <module >
result = stub.Predict(request, 10. 0)
File "/usr/local/lib64/python3.6/site-packages/grpc/_channel.py", line 690, in __call__ return _end_unary_response_blocking(state, call, False, None)
File "/usr/local/lib64/python3.6/site-packages/grpc/_channel.py", line 592, in _end_unary_response_blocking
raise _Rendezvous(state, None, None, deadline)
grpc._channel._Rendezvous : <_Rendezvous of RPC that terminated with :
status = StatusCode.RESOURCE_EXHAUSTED
details = "Received message larger than max (8388653 vs. 4194304)"
debug_error_string = "{"created ":"@ 1602672141. 715481155 ","description " :"Received message larger than max (8388653 vs. 4194304)
","file
":"src
/core
/ext
/filters
/message_size
/message_size_filter.cc
","file_line
":190,"grpc_status
":8}"
经过管理员提醒，尝试进行解决
@jsxyhelu The limit on the server side is actually 1GB. Your logs indicate 4MB.
It seems to be client side restriction.
Could you try the following settings :
options = [('grpc.max_receive_message_length', 100 * 1024 * 1024),('grpc.max_send_message_length', 100 * 1024 * 1024)]
channel = grpc.insecure_channel(server_url, options =
尝试服务端采用：
-d -v /models : /models :ro -p 9000 :
9000 openvino
/model_server
:latest
--model_path
/models
/model3
--model_name road
-segmentation
-adas
--port
9000
--log_level DEBUG
--shape
auto
客户端采用
--batch_size 1 --width 1024 --height 2048 --input_images_dir images --output_dir results
python3 road_detection.py --batch_size 1 --width 896 --height 512 --input_images_dir images --output_dir results
具体来说，就是采用这样的修改：
options = [( 'grpc.max_receive_message_length', 100 * 1024 * 1024),( 'grpc.max_send_message_length', 100 * 1024 * 1024)]
# this may make sense
channel = grpc.insecure_channel( "{}:{}".format(args[ 'grpc_address'],args[ 'grpc_port']),options = options)
stub = prediction_service_pb2_grpc.PredictionServiceStub(channel)
在具体调用的时候，除了这里的grpc的问题，还需要注意几个问题，一个是图像的大小，要按照模型的需要进行缩放；还有一个就是需要通过“ get_serving_meta.py”获得输出模型的具体名称，比如：
- 0 - 13 -centos tmp] # python3 get_serving_meta.py --grpc_port 9000 --model_name road-segmentation-adas --model_version 2 2020 - 10
-
16
14
:
20
:
40.
377893
: W tensorflow
/stream_executor
/platform
/default
/dso_loader.cc
:
59] Could
not load dynamic library
'libcudart.so.10.1'; dlerror
: libcudart.so.
10.
1
: cannot open shared object file
: No such file
or directory
2020
-
10
-
16
14
:
20
:
40.
387459
: I tensorflow
/stream_executor
/cuda
/cudart_stub.cc
:
29] Ignore above cudart dlerror
if you do not have a GPU set up on your machine.
Getting model metadata for model : road -segmentation -adas
Inputs metadata :
Input name : data; shape : [ 1, 3, 512, 896]; dtype : DT_FLOAT
Outputs metadata :
Output name : L0317_ReWeight_SoftMax; shape : [ 1, 4, 512, 896]; dtype :
接口文件经过大量改写
#
# Copyright (c) 2019-2020 Intel Corporation
#
# 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.
# update 2020/10/22
import argparse
import cv2
import datetime
import grpc
import numpy as np
import os
from tensorflow import make_tensor_proto, make_ndarray
from tensorflow_serving.apis import predict_pb2
from tensorflow_serving.apis import prediction_service_pb2_grpc
from client_utils import print_statistics
classes_color_map = [
( 150, 150, 150),
( 58, 55, 169),
( 211, 51, 17),
( 157, 80, 44),
( 23, 95, 189),
( 210, 133, 34),
( 76, 226, 202),
( 101, 138, 127),
( 223, 91, 182),
( 80, 128, 113),
( 235, 155, 55),
( 44, 151, 243),
( 159, 80, 170),
( 239, 208, 44),
( 128, 50, 51),
( 82, 141, 193),
( 9, 107, 10),
( 223, 90, 142),
( 50, 248, 83),
( 178, 101, 130),
( 71, 30, 204)
]
def load_image(file_path):
img = cv2.imread(file_path) # BGR color format, shape HWC
img = cv2.resize(img, (args[ 'width'], args[ 'height']))
img = img.transpose( 2, 0, 1).reshape( 1, 3,args[ 'height'],args[ 'width'])
# change shape to NCHW
return img
parser = argparse.ArgumentParser(description= 'Demo for road detection requests via TFS gRPC API.'
'analyses input images and saves with with detected skys.'
'it relies on model semantic-segmentation...')
parser.add_argument( '--input_images_dir', required= False, help= 'Directory with input images', default= "images/people")
parser.add_argument( '--output_dir', required= False, help= 'Directory for staring images with detection results', default= "results")
parser.add_argument( '--batch_size', required= False, help= 'How many images should be grouped in one batch', default= 1, type=int)
parser.add_argument( '--width', required= False, help= 'How the input image width should be resized in pixels', default= 1200, type=int)
parser.add_argument( '--height', required= False, help= 'How the input image width should be resized in pixels', default= 800, type=int)
parser.add_argument( '--grpc_address',required= False, default= 'localhost', help= 'Specify url to grpc service. default:localhost')
parser.add_argument( '--grpc_port',required= False, default= 9000, help= 'Specify port to grpc service. default: 9000')
args = vars(parser.parse_args())
options = [( 'grpc.max_receive_message_length', 100 * 1024 * 1024),( 'grpc.max_send_message_length', 100 * 1024 * 1024)]
# this may make sense
channel = grpc.insecure_channel( "{}:{}".format(args[ 'grpc_address'],args[ 'grpc_port']),options = options)
stub = prediction_service_pb2_grpc.PredictionServiceStub(channel)
files = os.listdir(args[ 'input_images_dir'])
batch_size = args[ 'batch_size']
print(files)
imgs = np.zeros(( 0, 3,args[ 'height'],args[ 'width']), np.dtype( '<f'))
for i in files:
img = load_image(os.path.join(args[ 'input_images_dir'], i))
imgs = np.append(imgs, img, axis= 0) # contains all imported images
print( 'Start processing {} iterations with batch size {}'.format(len(files)//batch_size , batch_size))
iteration = 0
processing_times = np.zeros(( 0),int)
for x in range( 0, imgs.shape[ 0] - batch_size + 1, batch_size):
iteration += 1
request = predict_pb2.PredictRequest()
request.model_spec.name = "road-segmentation-adas"
img = imgs[x:(x + batch_size)]
print( "\nRequest shape", img.shape)
request.inputs[ "data"].CopyFrom(make_tensor_proto(img, shape=(img.shape)))
start_time = datetime.datetime.now()
result = stub.Predict(request, 10.0) # result includes a dictionary with all model outputs print(img.shape)
output = make_ndarray(result.outputs[ "L0317_ReWeight_SoftMax"])
for y in range( 0,img.shape[ 0]): # iterate over responses from all images in the batch
img_out = output[y,:,:,:]
print( "image in batch item",y, ", output shape",img_out.shape)
img_out = img_out.transpose( 1, 2, 0)
print( "saving result to",os.path.join(args[ 'output_dir'],str(iteration)+ "_"+str(y)+ '.jpg'))
out_h, out_w,_ = img_out.shape
print(out_h)
print(out_w)
for batch, data in enumerate(output):
classes_map = np.zeros(shape=(out_h, out_w, 3), dtype=np.int)
for i in range(out_h):
for j in range(out_w):
if len(data[:, i, j]) == 1:
pixel_class = int(data[:, i, j])
else:
pixel_class = np.argmax(data[:, i, j])
classes_map[i, j, :] = classes_color_map[min(pixel_class, 20)]
cv2.imwrite(os.path.join(args[ 'output_dir'],str(iteration)+ "_"+str(batch)+ '.jpg'),classes_map)
附件列表