【PaddlePaddle+OpenVINO】花朵分类部署预测
转自AI Studio，原文链接：o【PaddlePaddle+OpenVINO】花朵分类部署预测 - 飞桨AI Studioo
一、鲜花识别
1.数据集简介
Oxford 102 Flowers Dataset 是一个花卉集合数据集，主要用于图像分类，它分为 102 个类别共计 102 种花，其中每个类别包含 40 到 258 张图像。
该数据集由牛津大学工程科学系于 2008 年发布，相关论文有《Automated flower classification over a large number of classes》。
在文件夹下已经生成用于训练和测试的三个.txt文件：train.txt（训练集，1020张图）、valid.txt（验证集，1020张图）、test.txt(6149)。文件中每行格式：图像相对路径 图像的label_id（注意：中间有空格）。
2.PaddleClas简介
PaddleClas目前已经是 release2.3了，和以前有脱胎换骨的差别，所以需要重新熟悉。
地址: PaddleClas: 飞桨图像分类套件PaddleClas是飞桨为工业界和学术界所准备的一个图像分类任务的工具集，助力使用者训练出更好的视觉模型和应用落地
configs已经移动到了ppcls目录 部署为单独的deploy目录
3.OpenVINO 2022.1 部署支持
OpenVINO™ 是开源的AI预测部署工具箱，支持多种格式，对飞桨支持友好，目前无需转换即可使用。
4.OpenVINO 2022.1 工作流程
In [ ]
# 解压缩数据集
!tar -xvf data/data19852/flowers102.tar -C ./data/ >log.log
二、PaddleClas准备
In [ ]
# 下载最新版
!git clone https://gitee.com/paddlepaddle/PaddleClas/ --depth=1
In [ ]
!pip install -r requirements.txt >log.log
/home/aistudio/PaddleClas
三、模型训练
1.修改imagenet_dataset.py
目录： ppclsdatadataloaderimagenet_dataset.py
修改原因是目录这块存在bug，注释：
assert os.path.exists(self._cls_path)assert os.path.exists(self._img_root)添加
self._cls_path=os.path.join(self._img_root,self._cls_path)否则不能使用相对路径
class ImageNetDataset(CommonDataset):
def _load_anno(self, seed=None):
会对目录进行检测，如果cls_path使用相对目录，就会报错，在此注释掉，并修改为self._cls_path=os.path.join(self._img_root,self._cls_path)
self._cls_path=os.path.join(self._img_root,self._cls_path)
print('self._cls_path',self._cls_path)
self.images = []
self.labels = []
with open(self._cls_path) as fd:
lines = fd.readlines()
if seed is not None:
np.random.RandomState(seed).shuffle(lines)
for l in lines:
l = l.strip().split(" ")
self.images.append(os.path.join(self._img_root, l[0]))
self.labels.append(int(l[1]))
assert os.path.exists(self.images[-1])
2.修改配置文件
# global configs
Global:
checkpoints: null
pretrained_model: null
output_dir: ./output/
# gpu或cpu配置
device: gpu
# 分类数量
class_num: 102
# 保存间隔
save_interval: 5
# 是否再训练立案过程中进行eval
eval_during_train: True
# eval间隔
eval_interval: 5
# 训练轮数
epochs: 20
# 打印batch step设置
print_batch_step: 10
# 是否使用visualdl
use_visualdl: False
# used for static mode and model export
image_shape: [3, 224, 224]
# 保存地址
save_inference_dir: ./inference
# model architecture
Arch:
name: ResNet50_vd
# loss function config for traing/eval process
Loss:
Train:
- CELoss:
weight: 1.0
Eval:
- CELoss:
weight: 1.0
Optimizer:
name: Momentum
momentum: 0.9
lr:
name: Cosine
learning_rate: 0.0125
warmup_epoch: 5
regularizer:
name: 'L2'
coeff: 0.00001
# data loader for train and eval
DataLoader:
Train:
dataset:
name: ImageNetDataset
image_root: /home/aistudio/data/oxford-102-flowers/oxford-102-flowers/
cls_label_path: train.txt
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- RandCropImage:
size: 224
- RandFlipImage:
flip_code: 1
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
sampler:
name: DistributedBatchSampler
batch_size: 256
drop_last: False
shuffle: True
loader:
num_workers: 4
use_shared_memory: True
Eval:
dataset:
name: ImageNetDataset
image_root: /home/aistudio/data/oxford-102-flowers/oxford-102-flowers/
cls_label_path: valid.txt
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
sampler:
name: DistributedBatchSampler
batch_size: 256
drop_last: False
shuffle: False
loader:
num_workers: 4
use_shared_memory: True
Infer:
infer_imgs: /home/aistudio/data/oxford-102-flowers/oxford-102-flowers/
batch_size: 10
transforms:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
resize_short: 256
- CropImage:
size: 224
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- ToCHWImage:
PostProcess:
name: Topk
topk: 5
class_id_map_file: /home/aistudio/data/oxford-102-flowers/oxford-102-flowers/jpg/image_00030.jpg
Metric:
Train:
- TopkAcc:
topk: [1, 5]
Eval:
- TopkAcc:
topk: [1, 5]
-c 参数是指定训练的配置文件路径，训练的具体超参数可查看yaml文件yaml文Global.device 参数设置为cpu，即使用CPU进行训练（若不设置，此参数默认为True）yaml文件中epochs参数设置为20，说明对整个数据集进行20个epoch迭代，预计训练20分钟左右（不同CPU，训练时间略有不同），此时训练模型不充分。若提高训练模型精度，请将此参数设大，如40，训练时间也会相应延长3.配置说明
3.1 全局配置(Global) 参数名字具体含义默认值可选值checkpoints断点模型路径，用于恢复训练nullstrpretrained_model预训练模型路径nullstroutput_dir保存模型路径"./output/"strsave_interval每隔多少个epoch保存模型1inteval_during_train是否在训练时进行评估Truebooleval_interval每隔多少个epoch进行模型评估1intepochs训练总epoch数intprint_batch_step每隔多少个mini-batch打印输出10intuse_visualdl是否是用visualdl可视化训练过程Falseboolimage_shape图片大小[3，224，224]list, shape: (3,)save_inference_dirinference模型的保存路径"./inference"streval_modeeval的模式"classification""retrieval" 3.2 结构(Arch) 参数名字具体含义默认值可选值name模型结构名字ResNet50PaddleClas提供的模型结构class_num分类数1000intpretrained预训练模型Falsebool， str 3.3 损失函数（Loss） 参数名字具体含义默认值可选值CELoss交叉熵损失函数————CELoss.weightCELoss的在整个Loss中的权重1.0floatCELoss.epsilonCELoss中label_smooth的epsilon值0.1float，0-1之间 3.4 优化器(Optimizer) 参数名字具体含义默认值可选值name优化器方法名"Momentum""RmsProp"等其他优化器momentummomentum值0.9floatlr.name学习率下降方式"Cosine""Linear"、"Piecewise"等其他下降方式lr.learning_rate学习率初始值0.1floatlr.warmup_epochwarmup轮数0int，如5regularizer.name正则化方法名"L2"["L1", "L2"]regularizer.coeff正则化系数0.00007float4.训练
In [ ]
!pwd
!cp ~/ResNet50_vd.yaml ./ppcls/configs/quick_start/ResNet50_vd.yaml
!cp ~/imagenet_dataset.py ./ppcls/data/dataloader/imagenet_dataset.py
/home/aistudio/PaddleClas
In [ ]
# GPU设置
!export CUDA_VISIBLE_DEVICES=0
# -o Arch.pretrained=True 使用预训练模型，当选择为True时，预训练权重会自动下载到本地
!python tools/train.py -c ./ppcls/configs/quick_start/ResNet50_vd.yaml -o Arch.pretrained=True
训练日志如下
[2021/10/31 01:53:47] root INFO: [Train][Epoch 16/20][Iter: 0/4]lr: 0.00285, top1: 0.93750, top5: 0.96484, CELoss: 0.36489, loss: 0.36489, batch_cost: 1.48066s, reader_cost: 0.68550, ips: 172.89543 images/sec, eta: 0:00:29
[2021/10/31 01:53:49] root INFO: [Train][Epoch 16/20][Avg]top1: 0.95098, top5: 0.97745, CELoss: 0.31581, loss: 0.31581
[2021/10/31 01:53:53] root INFO: [Train][Epoch 17/20][Iter: 0/4]lr: 0.00183, top1: 0.94531, top5: 0.97656, CELoss: 0.32916, loss: 0.32916, batch_cost: 1.47958s, reader_cost: 0.68473, ips: 173.02266 images/sec, eta: 0:00:23
[2021/10/31 01:53:55] root INFO: [Train][Epoch 17/20][Avg]top1: 0.95686, top5: 0.98137, CELoss: 0.29560, loss: 0.29560
[2021/10/31 01:53:58] root INFO: [Train][Epoch 18/20][Iter: 0/4]lr: 0.00101, top1: 0.93750, top5: 0.98047, CELoss: 0.31542, loss: 0.31542, batch_cost: 1.47524s, reader_cost: 0.68058, ips: 173.53117 images/sec, eta: 0:00:17
[2021/10/31 01:54:01] root INFO: [Train][Epoch 18/20][Avg]top1: 0.94608, top5: 0.98627, CELoss: 0.29086, loss: 0.29086
[2021/10/31 01:54:04] root INFO: [Train][Epoch 19/20][Iter: 0/4]lr: 0.00042, top1: 0.97266, top5: 0.98438, CELoss: 0.24642, loss: 0.24642, batch_cost: 1.47376s, reader_cost: 0.67916, ips: 173.70590 images/sec, eta: 0:00:11
[2021/10/31 01:54:07] root INFO: [Train][Epoch 19/20][Avg]top1: 0.94608, top5: 0.97941, CELoss: 0.30998, loss: 0.30998
[2021/10/31 01:54:10] root INFO: [Train][Epoch 20/20][Iter: 0/4]lr: 0.00008, top1: 0.98047, top5: 0.98438, CELoss: 0.20209, loss: 0.20209, batch_cost: 1.47083s, reader_cost: 0.67647, ips: 174.05180 images/sec, eta: 0:00:05
[2021/10/31 01:54:13] root INFO: [Train][Epoch 20/20][Avg]top1: 0.95784, top5: 0.98922, CELoss: 0.25974, loss: 0.25974
[2021/10/31 01:54:16] root INFO: [Eval][Epoch 20][Iter: 0/4]CELoss: 0.47912, loss: 0.47912, top1: 0.91797, top5: 0.96094, batch_cost: 3.26175s, reader_cost: 3.02034, ips: 78.48538 images/sec
[2021/10/31 01:54:17] root INFO: [Eval][Epoch 20][Avg]CELoss: 0.54982, loss: 0.54982, top1: 0.88922, top5: 0.96667
[2021/10/31 01:54:18] root INFO: Already save model in ./output/ResNet50_vd/best_model
[2021/10/31 01:54:18] root INFO: [Eval][Epoch 20][best metric: 0.8892156844045601]
[2021/10/31 01:54:18] root INFO: Already save model in ./output/ResNet50_vd/epoch_20
[2021/10/31 01:54:18] root INFO: Already save model in ./output/ResNet50_vd/latest
可见日志输出比较混乱，没有以前那么清晰，最好使用visualdl来查看训练情况
四、模型导出
In [ ]
!python tools/export_model.py
-c ./ppcls/configs/quick_start/ResNet50_vd.yaml
-o Global.pretrained_model=./output/ResNet50_vd/best_model
-o Global.save_inference_dir=./deploy/models/class_ResNet50_vd_ImageNet_infer
[2022/04/04 18:13:38] root INFO: =========================================================== == PaddleClas is powered by PaddlePaddle ! == =========================================================== == == == For more info please go to the following website. == == == == https://github.com/PaddlePaddle/PaddleClas == =========================================================== [2022/04/04 18:13:38] root INFO: Arch : [2022/04/04 18:13:38] root INFO: name : ResNet50_vd [2022/04/04 18:13:38] root INFO: DataLoader : [2022/04/04 18:13:38] root INFO: Eval : [2022/04/04 18:13:38] root INFO: dataset : [2022/04/04 18:13:38] root INFO: cls_label_path : valid.txt [2022/04/04 18:13:38] root INFO: image_root : /home/aistudio/data/oxford-102-flowers/oxford-102-flowers/ [2022/04/04 18:13:38] root INFO: name : ImageNetDataset [2022/04/04 18:13:38] root INFO: transform_ops : [2022/04/04 18:13:38] root INFO: DecodeImage : [2022/04/04 18:13:38] root INFO: channel_first : False [2022/04/04 18:13:38] root INFO: to_rgb : True [2022/04/04 18:13:38] root INFO: ResizeImage : [2022/04/04 18:13:38] root INFO: resize_short : 256 [2022/04/04 18:13:38] root INFO: CropImage : [2022/04/04 18:13:38] root INFO: size : 224 [2022/04/04 18:13:38] root INFO: NormalizeImage : [2022/04/04 18:13:38] root INFO: mean : [0.485, 0.456, 0.406] [2022/04/04 18:13:38] root INFO: order : [2022/04/04 18:13:38] root INFO: scale : 1.0/255.0 [2022/04/04 18:13:38] root INFO: std : [0.229, 0.224, 0.225] [2022/04/04 18:13:38] root INFO: loader : [2022/04/04 18:13:38] root INFO: num_workers : 4 [2022/04/04 18:13:38] root INFO: use_shared_memory : True [2022/04/04 18:13:38] root INFO: sampler : [2022/04/04 18:13:38] root INFO: batch_size : 128 [2022/04/04 18:13:38] root INFO: drop_last : False [2022/04/04 18:13:38] root INFO: name : DistributedBatchSampler [2022/04/04 18:13:38] root INFO: shuffle : False [2022/04/04 18:13:38] root INFO: Train : [2022/04/04 18:13:38] root INFO: dataset : [2022/04/04 18:13:38] root INFO: cls_label_path : train.txt [2022/04/04 18:13:38] root INFO: image_root : /home/aistudio/data/oxford-102-flowers/oxford-102-flowers/ [2022/04/04 18:13:38] root INFO: name : ImageNetDataset [2022/04/04 18:13:38] root INFO: transform_ops : [2022/04/04 18:13:38] root INFO: DecodeImage : [2022/04/04 18:13:38] root INFO: channel_first : False [2022/04/04 18:13:38] root INFO: to_rgb : True [2022/04/04 18:13:38] root INFO: RandCropImage : [2022/04/04 18:13:38] root INFO: size : 224 [2022/04/04 18:13:38] root INFO: RandFlipImage : [2022/04/04 18:13:38] root INFO: flip_code : 1 [2022/04/04 18:13:38] root INFO: NormalizeImage : [2022/04/04 18:13:38] root INFO: mean : [0.485, 0.456, 0.406] [2022/04/04 18:13:38] root INFO: order : [2022/04/04 18:13:38] root INFO: scale : 1.0/255.0 [2022/04/04 18:13:38] root INFO: std : [0.229, 0.224, 0.225] [2022/04/04 18:13:38] root INFO: loader : [2022/04/04 18:13:38] root INFO: num_workers : 4 [2022/04/04 18:13:38] root INFO: use_shared_memory : True [2022/04/04 18:13:38] root INFO: sampler : [2022/04/04 18:13:38] root INFO: batch_size : 128 [2022/04/04 18:13:38] root INFO: drop_last : False [2022/04/04 18:13:38] root INFO: name : DistributedBatchSampler [2022/04/04 18:13:38] root INFO: shuffle : True [2022/04/04 18:13:38] root INFO: Global : [2022/04/04 18:13:38] root INFO: checkpoints : None [2022/04/04 18:13:38] root INFO: class_num : 102 [2022/04/04 18:13:38] root INFO: device : gpu [2022/04/04 18:13:38] root INFO: epochs : 20 [2022/04/04 18:13:38] root INFO: eval_during_train : True [2022/04/04 18:13:38] root INFO: eval_interval : 5 [2022/04/04 18:13:38] root INFO: image_shape : [3, 224, 224] [2022/04/04 18:13:38] root INFO: output_dir : ./output/ [2022/04/04 18:13:38] root INFO: pretrained_model : ./output/ResNet50_vd/best_model [2022/04/04 18:13:38] root INFO: print_batch_step : 10 [2022/04/04 18:13:38] root INFO: save_inference_dir : ./deploy/models/class_ResNet50_vd_ImageNet_infer [2022/04/04 18:13:38] root INFO: save_interval : 5 [2022/04/04 18:13:38] root INFO: use_visualdl : False [2022/04/04 18:13:38] root INFO: Infer : [2022/04/04 18:13:38] root INFO: PostProcess : [2022/04/04 18:13:38] root INFO: class_id_map_file : /home/aistudio/data/oxford-102-flowers/oxford-102-flowers/jpg/image_00030.jpg [2022/04/04 18:13:38] root INFO: name : Topk [2022/04/04 18:13:38] root INFO: topk : 5 [2022/04/04 18:13:38] root INFO: batch_size : 10 [2022/04/04 18:13:38] root INFO: infer_imgs : /home/aistudio/data/oxford-102-flowers/oxford-102-flowers/ [2022/04/04 18:13:38] root INFO: transforms : [2022/04/04 18:13:38] root INFO: DecodeImage : [2022/04/04 18:13:38] root INFO: channel_first : False [2022/04/04 18:13:38] root INFO: to_rgb : True [2022/04/04 18:13:38] root INFO: ResizeImage : [2022/04/04 18:13:38] root INFO: resize_short : 256 [2022/04/04 18:13:38] root INFO: CropImage : [2022/04/04 18:13:38] root INFO: size : 224 [2022/04/04 18:13:38] root INFO: NormalizeImage : [2022/04/04 18:13:38] root INFO: mean : [0.485, 0.456, 0.406] [2022/04/04 18:13:38] root INFO: order : [2022/04/04 18:13:38] root INFO: scale : 1.0/255.0 [2022/04/04 18:13:38] root INFO: std : [0.229, 0.224, 0.225] [2022/04/04 18:13:38] root INFO: ToCHWImage : None [2022/04/04 18:13:38] root INFO: Loss : [2022/04/04 18:13:38] root INFO: Eval : [2022/04/04 18:13:38] root INFO: CELoss : [2022/04/04 18:13:38] root INFO: weight : 1.0 [2022/04/04 18:13:38] root INFO: Train : [2022/04/04 18:13:38] root INFO: CELoss : [2022/04/04 18:13:38] root INFO: weight : 1.0 [2022/04/04 18:13:38] root INFO: Metric : [2022/04/04 18:13:38] root INFO: Eval : [2022/04/04 18:13:38] root INFO: TopkAcc : [2022/04/04 18:13:38] root INFO: topk : [1, 5] [2022/04/04 18:13:38] root INFO: Train : [2022/04/04 18:13:38] root INFO: TopkAcc : [2022/04/04 18:13:38] root INFO: topk : [1, 5] [2022/04/04 18:13:38] root INFO: Optimizer : [2022/04/04 18:13:38] root INFO: lr : [2022/04/04 18:13:38] root INFO: learning_rate : 0.0125 [2022/04/04 18:13:38] root INFO: name : Cosine [2022/04/04 18:13:38] root INFO: warmup_epoch : 5 [2022/04/04 18:13:38] root INFO: momentum : 0.9 [2022/04/04 18:13:38] root INFO: name : Momentum [2022/04/04 18:13:38] root INFO: regularizer : [2022/04/04 18:13:38] root INFO: coeff : 1e-05 [2022/04/04 18:13:38] root INFO: name : L2 [2022/04/04 18:13:38] root INFO: train with paddle 2.1.2 and device CUDAPlace(0) [2022/04/04 18:13:38] root WARNING: The Global.class_num will be deprecated. Please use Arch.class_num instead. Arch.class_num has been set to 102. W0404 18:13:38.957692 2099 device_context.cc:404] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 10.1, Runtime API Version: 10.1 W0404 18:13:38.962862 2099 device_context.cc:422] device: 0, cuDNN Version: 7.6. /opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/utils.py:77: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working return (isinstance(seq, collections.Sequence) and
In [ ]
!ls ./deploy/models/class_ResNet50_vd_ImageNet_infer -la
total 93944 drwxr-xr-x 2 aistudio aistudio 4096 Apr 4 18:13 . drwxr-xr-x 3 aistudio aistudio 4096 Apr 4 18:13 .. -rw-r--r-- 1 aistudio aistudio 95165295 Apr 4 18:13 inference.pdiparams -rw-r--r-- 1 aistudio aistudio 23453 Apr 4 18:13 inference.pdiparams.info -rw-r--r-- 1 aistudio aistudio 996386 Apr 4 18:13 inference.pdmodel
五、OpenVINO预测
鉴于AiStudio无法使用最新版OpenVINO，在本地跑完后上传
1.OpenVINO安装
此处要注意，使用最新版的OpenVINO，目前最新版为2022.1.0
!pip install OpenVINO
Looking in indexes: https://pypi.tuna.tsinghua.edu.cn/simple
Requirement already satisfied: OpenVINO in c:miniconda3envsp2glibsite-packages (2022.1.0)
Requirement already satisfied: numpy<1.20,>=1.16.6 in c:miniconda3envsp2glibsite-packages (from OpenVINO) (1.19.3)
2.Import
导入必须的OpenVINO库
# model download
from pathlib import Path
import os
import urllib.request
import tarfile
# inference
from openvino.runtime import Core
# preprocessing
import cv2
import numpy as np
from openvino.preprocess import PrePostProcessor, ResizeAlgorithm
from openvino.runtime import Layout, Type, AsyncInferQueue, PartialShape
# results visualization
import time
import json
from IPython.display import Image
3.预处理
3.1生成花分类字典flowers_classes={}
for i in range(102):
flowers_classes[str(i)]='flower_'+ str(i)
print(flowers_classes)
{'0': 'flower_0', '1': 'flower_1', '2': 'flower_2', '3': 'flower_3', '4': 'flower_4', '5': 'flower_5', '6': 'flower_6', '7': 'flower_7', '8': 'flower_8', '9': 'flower_9', '10': 'flower_10', '11': 'flower_11', '12': 'flower_12', '13': 'flower_13', '14': 'flower_14', '15': 'flower_15', '16': 'flower_16', '17': 'flower_17', '18': 'flower_18', '19': 'flower_19', '20': 'flower_20', '21': 'flower_21', '22': 'flower_22', '23': 'flower_23', '24': 'flower_24', '25': 'flower_25', '26': 'flower_26', '27': 'flower_27', '28': 'flower_28', '29': 'flower_29', '30': 'flower_30', '31': 'flower_31', '32': 'flower_32', '33': 'flower_33', '34': 'flower_34', '35': 'flower_35', '36': 'flower_36', '37': 'flower_37', '38': 'flower_38', '39': 'flower_39', '40': 'flower_40', '41': 'flower_41', '42': 'flower_42', '43': 'flower_43', '44': 'flower_44', '45': 'flower_45', '46': 'flower_46', '47': 'flower_47', '48': 'flower_48', '49': 'flower_49', '50': 'flower_50', '51': 'flower_51', '52': 'flower_52', '53': 'flower_53', '54': 'flower_54', '55': 'flower_55', '56': 'flower_56', '57': 'flower_57', '58': 'flower_58', '59': 'flower_59', '60': 'flower_60', '61': 'flower_61', '62': 'flower_62', '63': 'flower_63', '64': 'flower_64', '65': 'flower_65', '66': 'flower_66', '67': 'flower_67', '68': 'flower_68', '69': 'flower_69', '70': 'flower_70', '71': 'flower_71', '72': 'flower_72', '73': 'flower_73', '74': 'flower_74', '75': 'flower_75', '76': 'flower_76', '77': 'flower_77', '78': 'flower_78', '79': 'flower_79', '80': 'flower_80', '81': 'flower_81', '82': 'flower_82', '83': 'flower_83', '84': 'flower_84', '85': 'flower_85', '86': 'flower_86', '87': 'flower_87', '88': 'flower_88', '89': 'flower_89', '90': 'flower_90', '91': 'flower_91', '92': 'flower_92', '93': 'flower_93', '94': 'flower_94', '95': 'flower_95', '96': 'flower_96', '97': 'flower_97', '98': 'flower_98', '99': 'flower_99', '100': 'flower_100', '101': 'flower_101'} 3.2预处理callback定义
def callback(infer_request, i) -> None:
"""
Define the callback function for postprocessing
:param: infer_request: the infer_request object
i: the iteration of inference
:retuns:
None
"""
# flowers_classes
predictions = next(iter(infer_request.results.values()))
indices = np.argsort(-predictions[0])
if (i == 0):
# Calculate the first inference time
latency = time.time() - start
print(f"latency: {latency}")
for n in range(5):
print(
"class name: {}, probability: {:.5f}"
.format(flowers_classes[str(list(indices)[n])], predictions[0][list(indices)[n]])
)
3.3读取模型# Intialize Inference Engine with Core()
ie = Core()
# model_path
model_path="inference/inference.pdmodel"
model = ie.read_model(model_path)
# get the information of intput and output layer
input_layer = model.input(0)
output_layer = model.output(0)
4.调用API进行预处理
如果输入数据不完全符合模型输入张量，则需要额外的操作/步骤将数据转换为模型所期望的格式。这些操作被称为“预处理”。预处理步骤被集成到执行图中，并在选定的设备(CPU/GPU/VPU/等)上执行，而不是总是在CPU上执行。这将提高所选设备的利用率。相关 API: https://docs.openvino.ai/latest/openvino_docs_OV_Runtime_UG_Preprocessing_Overview.html
# 待预测图片
filename = "myflower.jpg"
test_image = cv2.imread(filename)
test_image = np.expand_dims(test_image, 0) / 255
_, h, w, _ = test_image.shape
# 调整模型输入图片尺寸
model.reshape({input_layer.any_name: PartialShape([1, 3, 224, 224])})
ppp = PrePostProcessor(model)
# 设置输入 tensor 信息:
# - input() 提供模型的输入
# - 数据格式 "NHWC"
# - 设置静态模型输入维度
ppp.input().tensor()
.set_spatial_static_shape(h, w)
.set_layout(Layout("NHWC"))
inputs = model.inputs
# 设模型有“NCHW”布局作为输入
ppp.input().model().set_layout(Layout("NCHW"))
# 处理操作:
# - tensor RESIZE_LINEAR 缩放设置
# - 每个通道的归一化
# - 将每个像素数据划分为适当的比例值
ppp.input().preprocess()
.resize(ResizeAlgorithm.RESIZE_LINEAR, 224, 224)
.mean([0.485, 0.456, 0.406])
.scale([0.229, 0.224, 0.225])
# 设置输出张量信息:
# - 张量精度设置为 'f32'
ppp.output().tensor().set_element_type(Type.f32)
# Apply preprocessing to modify the original 'model'
model = ppp.build()
5.预测
使用“AUTO”作为设备名，委托OpenVINO选择设备。自动设备插件内部识别和选择设备从英特尔的CPU和GPU之间依赖于设备功能和模型(例如，精度)的特点。然后，它将推理请求分配给最佳设备。
AUTO立即在CPU上启动推理，然后在准备好后透明地转移到GPU(或VPU)，大大减少了第一次推理的时间。
# 检查可用设备
devices = ie.available_devices
for device in devices:
device_name = ie.get_property(device_name=device, name="FULL_DEVICE_NAME")
print(f"{device}: {device_name}")
# 将模型加载到由AUTO从可用设备列表中选择的设备
compiled_model = ie.compile_model(model=model, device_name="AUTO")
# 创建请求队列
infer_queue = AsyncInferQueue(compiled_model)
infer_queue.set_callback(callback)
start = time.time()
# 开始预测
infer_queue.start_async({input_layer.any_name: test_image}, 0)
infer_queue.wait_all()
Image(filename=filename)
CPU: Intel(R) Core(TM) i5-9400F CPU @ 2.90GHz
latency: 0.02329254150390625
class name: flower_76, probability: 0.40075
class name: flower_81, probability: 0.15170
class name: flower_91, probability: 0.03979
class name: flower_12, probability: 0.03356
class name: flower_17, probability: 0.02347
6.性能技巧:延迟和吞吐量
吞吐量和延迟是一些最广泛使用的度量应用程序整体性能的指标。
OpenVINO性能提示是在考虑可移植性的情况下配置性能的新方法。性能提示将允许设备自己配置，而不是将应用程序需要映射到低级别的性能设置，并保持一个相关的应用程序逻辑来分别配置每个可能的设备。
高级技巧: https://docs.openvino.ai/latest/openvino_docs_OV_UG_Performance_Hints.html
6.1延迟计算可以通过配置调整应用程序的性能设置，让设备调整以实现更好的面向延迟的性能。
loop = 100
# AUTO sets device config based on hints
compiled_model = ie.compile_model(model=model, device_name="AUTO", config={"PERFORMANCE_HINT": "LATENCY"})
infer_queue = AsyncInferQueue(compiled_model)
# implement AsyncInferQueue Python API to boost the performance in Async mode
infer_queue.set_callback(callback)
start = time.time()
# run infernce for 100 times to get the average FPS
for i in range(loop):
infer_queue.start_async({input_layer.any_name: test_image}, i)
infer_queue.wait_all()
end = time.time()
# Calculate the average FPS
fps = loop / (end - start)
print(f"fps: {fps}")
latency: 0.018686771392822266
class name: flower_76, probability: 0.40075
class name: flower_81, probability: 0.15170
class name: flower_91, probability: 0.03979
class name: flower_12, probability: 0.03356
class name: flower_17, probability: 0.02347
fps: 50.20953840260486
6.2吞吐量计算可以使用配置设置应用程序的性能设置，让设备调整以实现更好的吞吐量性能。
# AUTO sets device config based on hints
compiled_model = ie.compile_model(model=model, device_name="AUTO", config={"PERFORMANCE_HINT": "THROUGHPUT"})
infer_queue = AsyncInferQueue(compiled_model)
infer_queue.set_callback(callback)
start = time.time()
for i in range(loop):
infer_queue.start_async({input_layer.any_name: test_image}, i)
infer_queue.wait_all()
end = time.time()
# Calculate the average FPS
fps = loop / (end - start)
print(f"fps: {fps}")
latency: 0.04830741882324219
class name: flower_76, probability: 0.40075
class name: flower_81, probability: 0.15170
class name: flower_91, probability: 0.03979
class name: flower_12, probability: 0.03356
class name: flower_17, probability: 0.02347
fps: 57.274455164002134
!benchmark_app -m $model_path -data_shape [1,3,224,224] -hint "latency"
[Step 1/11] Parsing and validating input arguments
[ WARNING ] -nstreams default value is determined automatically for a device. Although the automatic selection usually provides a reasonable performance, but it still may be non-optimal for some cases, for more information look at README.
[Step 2/11] Loading OpenVINO
[ INFO ] OpenVINO:
API version............. 2022.1.0-7019-cdb9bec7210-releases/2022/1
[ INFO ] Device info
CPU
openvino_intel_cpu_plugin version 2022.1
Build................... 2022.1.0-7019-cdb9bec7210-releases/2022/1
[Step 3/11] Setting device configuration
[Step 4/11] Reading network files
[ INFO ] Read model took 144.60 ms
[Step 5/11] Resizing network to match image sizes and given batch
[ INFO ] Network batch size: ?
[Step 6/11] Configuring input of the model
[ INFO ] Model input 'x' precision u8, dimensions ([N,C,H,W]): ? 3 224 224
[ INFO ] Model output 'save_infer_model/scale_0.tmp_1' precision f32, dimensions ([...]): ? 102
[Step 7/11] Loading the model to the device
[ INFO ] Compile model took 232.62 ms
[Step 8/11] Querying optimal runtime parameters
[ INFO ] DEVICE: CPU
[ INFO ] AVAILABLE_DEVICES , ['']
[ INFO ] RANGE_FOR_ASYNC_INFER_REQUESTS , (1, 1, 1)
[ INFO ] RANGE_FOR_STREAMS , (1, 6)
[ INFO ] FULL_DEVICE_NAME , Intel(R) Core(TM) i5-9400F CPU @ 2.90GHz
[ INFO ] OPTIMIZATION_CAPABILITIES , ['FP32', 'FP16', 'INT8', 'BIN', 'EXPORT_IMPORT']
[ INFO ] CACHE_DIR ,
[ INFO ] NUM_STREAMS , 1
[ INFO ] INFERENCE_NUM_THREADS , 0
[ INFO ] PERF_COUNT , False
[ INFO ] PERFORMANCE_HINT_NUM_REQUESTS , 0
[Step 9/11] Creating infer requests and preparing input data
[ INFO ] Create 1 infer requests took 0.00 ms
[ WARNING ] No input files were given for input 'x'!. This input will be filled with random values!
[ INFO ] Fill input 'x' with random values
[Step 10/11] Measuring performance (Start inference asynchronously, 1 inference requests, inference only: False, limits: 60000 ms duration)
[ INFO ] Benchmarking in full mode (inputs filling are included in measurement loop).
[ INFO ] First inference took 31.26 ms
[Step 11/11] Dumping statistics report
Count: 2793 iterations
Duration: 60018.45 ms
Latency:
AVG: 21.40 ms
MIN: 17.28 ms
MAX: 73.23 ms
Throughput: 46.54 FPS
相关知识
Python教程之使用 Gradio 部署 Keras 花卉分类模型
python机器学习 鸢尾花分类预测 详细教程 数据集+源码+远程部署
EfficientNet实现农业病害识别(FastDeploy部署和安卓端部署)
使用PyTorch实现对花朵的分类
构建、训练和部署 102 种花卉类型的真实花卉分类器
机器学习实战：鸢尾花分类
花卉分类识别模型设计
深度学习：花朵图像生成模型训练数据集
“花朵分类“ 手把手搭建【卷积神经网络】
对鸢尾花进行分类预测
网址: 【PaddlePaddle+OpenVINO】花朵分类部署预测 https://www.huajiangbk.com/newsview1985350.html
| 上一篇: 【科研骨干】李鲜花 |
下一篇: 东大田磊研究员“噬菌体微米级仿生 |
推荐分享
- 1君子兰什么品种最名贵 十大名 4012
- 2世界上最名贵的10种兰花图片 3364
- 3花圈挽联怎么写？ 3286
- 4迷信说家里不能放假花 家里摆 1878
- 5香山红叶什么时候红 1493
- 6花的意思,花的解释,花的拼音 1210
- 7教师节送什么花最合适 1167
- 8勿忘我花图片 1103
- 9橄榄枝的象征意义 1093
- 10洛阳的市花 1039
