本文围绕交通标志分类识别竞赛展开，介绍了赛题背景、任务及数据集情况。还阐述了利用PaddleClas套件，通过数据格式化划分、配置MobileNetV3模型、训练评估、导出模型并预测，最终生成结果提交的完整流程与注意事项。

一、【Paddle打比赛】 交通标志分类识别

1.赛题名称

【交通标志分类识别 竞赛 - DataFountain 】?https://www.datafountain.cn/competitions/553

2.赛题背景

目标检测在自动驾驶等方面有着广泛的应用前景，在自动驾驶场景中，需要对交通标志（如限速标志等）进行识别以采取不同的驾驶策略。此赛题旨在使用Oneflow框架识别确定图像中显示的交通标志的类型，并且对不同的环境（如光线、障碍物或标志距离）具有鲁棒性。

3.赛题任务

此赛题的数据集由云测数据提供。比赛数据集中包含6358张真实场景下行车记录仪采集的图片，我们选取其中95%的图片作为训练集，5%的图片作为测试集。参赛者需基于Oneflow框架在训练集上进行训练，对测试集中的交通标志进行分类识别。数据集中一共有10种细分的交通标志标签。

4.数据简介

本赛题的数据集包含6358张人工标注的类别标签。标签包括以下10种类别：“GuideSign”, “M1”, “M4, “M5”, “M6”, “M7”, “P1”, “P10_50”, “P12”, “W1”，分别对应十个不同的交通标识类别。所有数据已经按比例人工划分为了训练集和测试集，并且提供了相关的训练集标注的json文件。

5.数据说明

数据集包含train和test两个文件夹，以及一个统一的train.json文件用来保存所有训练数据集的标注，图片标签将字典以json格式序列化进行保存，所有的标注都保存在train.json文件中的名为annotations的key下，其中每个训练文件都包含以下属性：

• P12： 礼让标志
• P10_50：限速50
• GuideSing：路牌
• W1：警告
• P1：禁停
• M1：方向指示
• M4：直行
• M6：机动车道
• M6：自行车道
• M7：人行横道

6.提交要求

参赛者使用Oneflow框架对数据集进行训练后对测试集图片进行推理后，

• 1.将测试集图片的目标检测和识别结果以与训练集格式保持一致的json文件序列化保存，并上传至参赛平台由参赛平台自动评测返回结果。
• 2.在提交时的备注附上自己的模型github仓库链接。

{ “annotations”: [ { “filename”: “test/W1/00123.jpg”, # filepath: str “label”: 0 # label: int from 0 to 9 }] }

二、环境准备

飞桨图像识别套件PaddleClas是飞桨为工业界和学术界所准备的一个图像识别任务的工具集，助力使用者训练出更好的视觉模型和应用落地。此次计划使用端到端的PaddleClas图像分类套件来快速完成分类。

# git下载!git clone https://gitee.com/paddlepaddle/PaddleClas.git --depth=1登录后复制

Cloning into 'PaddleClas'...remote: Enumerating objects: 1037, done.remote: Counting objects: 100% (1037/1037), done.remote: Compressing objects: 100% (918/918), done.remote: Total 1037 (delta 332), reused 478 (delta 95), pack-reused 0Receiving objects: 100% (1037/1037), 59.94 MiB | 7.11 MiB/s, done.Resolving deltas: 100% (332/332), done.Checking connectivity... done.登录后复制In [?]

# PaddleClas安装%cd ~/PaddleClas/!pip install -U pip!pip install -r requirements.txt!pip install -e ./登录后复制

三、数据准备

1.数据解压缩

%cd ~!unzip -oq /home/aistudio/data/data125044/road_sign.zip登录后复制In [?]

!unzip -qoa train_dataset.zip!unzip -qoa test_dataset.zip登录后复制In [?]

! rm test_dataset.zip! rm train_dataset.zip登录后复制In [2]

from PIL import Imageimg=Image.open(”train/P12/00014.jpg“)img登录后复制

<PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=224x224 at 0x7F008C68FE10>登录后复制In [3]

img=Image.open(”train/P10_50/00014.jpg“)img登录后复制

<PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=224x224 at 0x7F008C66F290>登录后复制In [4]

img=Image.open(”train/GuideSign/00452.jpg“)img登录后复制

<PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=224x224 at 0x7F008C686690>登录后复制

2.数据集格式化 AND 划分

%cd ~import jsonimport ostrain_json={}with open('train.json','r') as f: train_json=json.load(f)print(len(train_json[”annotations“]))train_f = open('train_list.txt', 'w') eval_f = open('eval_list.txt', 'w') for i in range(len(train_json[”annotations“])): if i%5==0: eval_f.write(train_json[”annotations“][i]['filename']+' '+ str(train_json[”annotations“][i]['label']) +'n') else: train_f.write(train_json[”annotations“][i]['filename']+' '+ str(train_json[”annotations“][i]['label']) +'n')登录后复制

/home/aistudio6024登录后复制

四、模型训练

1.配置

以?PaddleClas/ppcls/configs/quick_start/MobileNetV3_large_x1_0.yaml?为基础进行配置，具体见下：

# global configsGlobal: checkpoints: null pretrained_model: null output_dir: ./output/ device: gpu save_interval: 1 eval_during_train: True eval_interval: 1 epochs: 20 print_batch_step: 10 use_visualdl: True # used for static mode and model export image_shape: [3, 224, 224] save_inference_dir: ./inference# model architectureArch: name: MobileNetV3_large_x1_0 class_num: 102 # loss function config for traing/eval processLoss: Train: - CELoss: weight: 1.0 Eval: - CELoss: weight: 1.0Optimizer: name: Momentum momentum: 0.9 lr: name: Cosine learning_rate: 0.00375 warmup_epoch: 5 last_epoch: -1 regularizer: name: 'L2' coeff: 0.000001# data loader for train and evalDataLoader: Train: dataset: name: ImageNetDataset image_root: /home/aistudio/ cls_label_path: /home/aistudio/train_list.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: 512 drop_last: False shuffle: True loader: num_workers: 4 use_shared_memory: True Eval: dataset: name: ImageNetDataset image_root: /home/aistudio/ cls_label_path: /home/aistudio/eval_list.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: 128 drop_last: False shuffle: False loader: num_workers: 4 use_shared_memory: TrueInfer: infer_imgs: docs/images/whl/demo.jpg 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: ./dataset/flowers102/flowers102_label_list.txtMetric: Train: - TopkAcc: topk: [1, 5] Eval: - TopkAcc: topk: [1, 5]登录后复制

2.开始训练

# 覆盖配置%cd ~!cp -f ~/MobileNetV3_large_x1_0.yaml PaddleClas/ppcls/configs/quick_start/MobileNetV3_large_x1_0.yaml登录后复制In [?]

%cd ~/PaddleClas/# 进入 PaddleClas 根目录，执行此命令# Arch.pretrained=True使用预训练模型!python tools/train.py -c ./ppcls/configs/quick_start/MobileNetV3_large_x1_0.yaml -o Arch.pretrained=True登录后复制

训练情况：

3.开始评估

%cd ~/PaddleClas/!python tools/eval.py -c ./ppcls/configs/quick_start/MobileNetV3_large_x1_0.yaml -o Global.pretrained_model=./output/MobileNetV3_large_x1_0/best_model登录后复制

[2022/01/10 23:08:17] root INFO: [Eval][Epoch 0][Iter: 0/9]CELoss: 0.36174, loss: 0.36174, top1: 0.90625, top5: 1.00000, batch_cost: 1.31483s, reader_cost: 1.22580, ips: 97.35107 images/sec[2022/01/10 23:08:17] root INFO: [Eval][Epoch 0][Avg]CELoss: 0.25337, loss: 0.25337, top1: 0.93923, top5: 0.99732登录后复制

五、预测

1.模型导出

# 模型导出%cd ~/PaddleClas/!python tools/export_model.py -c ./ppcls/configs/quick_start/MobileNetV3_large_x1_0.yaml -o Global.pretrained_model=./output/MobileNetV3_large_x1_0/best_model登录后复制

2.开始预测

编辑?PaddleClas/deploy/python/predict_cls.py，按提交格式输出预测结果到文件。

import osimport sys__dir__ = os.path.dirname(os.path.abspath(__file__))sys.path.append(os.path.abspath(os.path.join(__dir__, '../')))import cv2import numpy as npfrom utils import loggerfrom utils import configfrom utils.predictor import Predictor# from utils.get_image_list import get_image_listfrom python.preprocess import create_operatorsfrom python.postprocess import build_postprocessclass ClsPredictor(Predictor): def __init__(self, config): super().__init__(config[”Global“]) self.preprocess_ops = [] self.postprocess = None if ”PreProcess“ in config: if ”transform_ops“ in config[”PreProcess“]: self.preprocess_ops = create_operators(config[”PreProcess“][ ”transform_ops“]) if ”PostProcess“ in config: self.postprocess = build_postprocess(config[”PostProcess“]) # for whole_chain project to test each repo of paddle self.benchmark = config[”Global“].get(”benchmark“, False) if self.benchmark: import auto_log import os pid = os.getpid() self.auto_logger = auto_log.AutoLogger( model_name=config[”Global“].get(”model_name“, ”cls“), model_precision='fp16' if config[”Global“][”use_fp16“] else 'fp32', batch_size=config[”Global“].get(”batch_size“, 1), data_shape=[3, 224, 224], save_path=config[”Global“].get(”save_log_path“, ”./auto_log.log“), inference_config=self.config, pids=pid, process_name=None, gpu_ids=None, time_keys=[ 'preprocess_time', 'inference_time', 'postprocess_time' ], warmup=2) def predict(self, images): input_names = self.paddle_predictor.get_input_names() input_tensor = self.paddle_predictor.get_input_handle(input_names[0]) output_names = self.paddle_predictor.get_output_names() output_tensor = self.paddle_predictor.get_output_handle(output_names[ 0]) if self.benchmark: self.auto_logger.times.start() if not isinstance(images, (list, )): images = [images] for idx in range(len(images)): for ops in self.preprocess_ops: images[idx] = ops(images[idx]) image = np.array(images) if self.benchmark: self.auto_logger.times.stamp() input_tensor.copy_from_cpu(image) self.paddle_predictor.run() batch_output = output_tensor.copy_to_cpu() if self.benchmark: self.auto_logger.times.stamp() if self.postprocess is not None: batch_output = self.postprocess(batch_output) if self.benchmark: self.auto_logger.times.end(stamp=True) return batch_outputdef get_image_list(img_file_path): imgs_lists = [] # if img_file is None or not os.path.exists(img_file): # raise Exception(”not found any img file in {}“.format(img_file)) img_end = ['jpg', 'png', 'jpeg', 'JPEG', 'JPG', 'bmp'] import os from os.path import join, getsize for root, dirs, files in os.walk(img_file_path): for file in files: img_file=os.path.join(root, file) imgs_lists.append(img_file) if len(imgs_lists) == 0: raise Exception(”not found any img file in {}“.format(img_file)) imgs_lists = sorted(imgs_lists) return imgs_listsdef main(config): cls_predictor = ClsPredictor(config) image_list = get_image_list(config[”Global“][”infer_imgs“]) batch_imgs = [] batch_names = [] cnt = 0 # dic import json mydic=[] for idx, img_path in enumerate(image_list): img = cv2.imread(img_path) if img is None: logger.warning( ”Image file failed to read and has been skipped. The path: {}“. format(img_path)) else: img = img[:, :, ::-1] batch_imgs.append(img) img_name = os.path.basename(img_path) batch_names.append(img_name) cnt += 1 if cnt % config[”Global“][”batch_size“] == 0 or (idx + 1 ) == len(image_list): if len(batch_imgs) == 0: continue batch_results = cls_predictor.predict(batch_imgs) for number, result_dict in enumerate(batch_results): filename = batch_names[number] clas_ids = result_dict[”class_ids“] scores_str = ”[{}]“.format(”, “.join(”{:.2f}“.format( r) for r in result_dict[”scores“])) label_names = result_dict[”label_names“] print(”{}:tclass id(s): {}, score(s): {}, label_name(s): {}“. format(filename, clas_ids, scores_str, label_names)) temp_dic={} temp_dic[”filename“]=”test/“+img_path.strip('/home/aistudio') temp_dic[”label“]=clas_ids[0] mydic.append(temp_dic) batch_imgs = [] batch_names = [] # 保存到文件 with open('/home/aistudio/result.json', 'w') as f: myjson={”annotations“:mydic} json.dump(myjson,f) if cls_predictor.benchmark: cls_predictor.auto_logger.report() returnif __name__ == ”__main__“: args = config.parse_args() config = config.get_config(args.config, overrides=args.override, show=True) main(config)登录后复制In [?]

# 自定义predict_cls.py覆盖%cd ~!cp -f ~/predict_cls.py ~/PaddleClas/deploy/python/predict_cls.py登录后复制In [77]

# 开始预测%cd /home/aistudio/PaddleClas/deploy!python3 python/predict_cls.py -c configs/inference_cls.yaml -o Global.infer_imgs=/home/aistudio/test -o Global.inference_model_dir=../inference/ -o PostProcess.Topk.class_id_map_file=None登录后复制

预测日志

01220.jpg:class id(s): [0, 3, 2, 4, 5], score(s): [1.00, 0.00, 0.00, 0.00, 0.00], label_name(s): []01229.jpg:class id(s): [0, 2, 3, 4, 5], score(s): [0.97, 0.02, 0.00, 0.00, 0.00], label_name(s): []01232.jpg:class id(s): [0, 3, 2, 8, 5], score(s): [0.98, 0.01, 0.00, 0.00, 0.00], label_name(s): []00004.jpg:class id(s): [2, 3, 1, 0, 5], score(s): [0.79, 0.10, 0.03, 0.02, 0.01], label_name(s): []00013.jpg:class id(s): [1, 5, 2, 3, 0], score(s): [0.80, 0.08, 0.05, 0.02, 0.01], label_name(s): []00052.jpg:class id(s): [1, 5, 3, 4, 6], score(s): [0.91, 0.03, 0.02, 0.01, 0.01], label_name(s): []00061.jpg:class id(s): [3, 0, 4, 2, 5], score(s): [0.60, 0.14, 0.09, 0.06, 0.03], label_name(s): []00092.jpg:class id(s): [1, 2, 3, 5, 4], score(s): [0.65, 0.18, 0.08, 0.04, 0.01], label_name(s): []登录后复制

六、提交

1.提交结果

提交当前目录下的result.json，即刻得到成绩，如下图：

大家来试试吧！

2.其他注意事项

生成版本时提示存在无效软链接无法保存?，可以在终端?PaddleClas?下运行下列代码清理即可。

for a in `find . -type l`do stat -L $a >/dev/null 2>/dev/null if [ $? -gt 0 ] then rm $a fidone登录后复制