转载自AI Studio
项目链接https://aistudio.baidu.com/aistudio/projectdetail/3169455

天气以及时间分类

赛题名称

天气以及时间分类

https://www.datafountain.cn/competitions/555

赛题背景

在自动驾驶场景中，天气和时间（黎明、早上、下午、黄昏、夜晚）会对传感器的精度造成影响，比如雨天和夜晚会对视觉传感器的精度造成很大的影响。此赛题旨在使用Oneflow框架对拍摄的照片天气和时间进行分类，从而在不同的天气和时间使用不同的自动驾驶策略。

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3000张真实场景下行车记录仪采集的图片，其中训练集包含2600张带有天气和时间类别标签的图片，测试集包含400张不带有标签的图片。

本赛题的数据集包含2600张人工标注的天气和时间标签。天气类别包含多云、晴天、雨天、雪天和雾天5个类别；时间包含黎明、早上、下午、黄昏、夜晚5个类别。
部分数据可视化及标签如下：


数据说明

数据集包含anno和image两个文件夹，anno文件夹中包含2600个标签json文件，image文件夹中包含3000张行车记录仪拍摄的JPEG编码照片。图片标签将字典以json格式序列化进行保存：

列名取值范围作用
Period黎明、早上、下午、黄昏、夜晚图片拍摄时间
Weather多云、晴天、雨天、雪天、雾天图片天气
!echo y | unzip test_dataset.zip > log.log
!echo y | unzip train_dataset.zip > log.log

replace test_images/00008.jpg? [y]es, [n]o, [A]ll, [N]one, [r]ename: replace test_images/00018.jpg? [y]es, [n]o, [A]ll, [N]one, [r]ename:  NULL
(EOF or read error, treating as "[N]one" ...)
replace train_images/00001.jpg? [y]es, [n]o, [A]ll, [N]one, [r]ename: replace train_images/00002.jpg? [y]es, [n]o, [A]ll, [N]one, [r]ename:  NULL
(EOF or read error, treating as "[N]one" ...)

import io
import math, json
import numpy as np
import pandas as pd
from PIL import Image

import matplotlib.pyplot as plt
import paddle
import paddle.nn.functional as F
import paddle.vision.transforms as T
from paddle.io import DataLoader, Dataset

import warnings
warnings.filterwarnings("ignore")

paddle.__version__

'2.2.0'

数据读取

# 读取数据集标注，提取标注信息中的关键信息
train_json = pd.read_json('train.json')
train_json['filename'] = train_json['annotations'].apply(lambda x: x['filename'].replace('\\', '/'))
train_json['period'] = train_json['annotations'].apply(lambda x: x['period'])
train_json['weather'] = train_json['annotations'].apply(lambda x: x['weather'])

train_json.head()
annotationsfilenameperiodweather
0{'filename': 'train_images\00001.jpg', 'period...train_images/00001.jpgMorningCloudy
1{'filename': 'train_images\00002.jpg', 'period...train_images/00002.jpgAfternoonCloudy
2{'filename': 'train_images\00003.jpg', 'period...train_images/00003.jpgMorningCloudy
3{'filename': 'train_images\00004.jpg', 'period...train_images/00004.jpgMorningSunny
4{'filename': 'train_images\00005.jpg', 'period...train_images/00005.jpgAfternoonCloudy

标签处理

# 将标签进行编码，这里需要记住编码的次序。
# 可以手动使用dict来实现，这里用factorize
train_json['period'], period_dict = pd.factorize(train_json['period'])
train_json['weather'], weather_dict = pd.factorize(train_json['weather'])

统计标签

train_json['period'].value_counts()

0    1613
1     829
3     124
2      34
Name: period, dtype: int64

train_json['weather'].value_counts()

0    1119
1     886
2     595
Name: weather, dtype: int64

自定义数据集

# 自定义数据集
class WeatherDataset(Dataset):
    def __init__(self, df):
        super(WeatherDataset, self).__init__()
        self.df = df
    
        # 定义数据扩增方法
        self.transform = T.Compose([
            T.Resize(size=(340,340)),
            T.RandomCrop(size=(256, 256)),
            T.RandomRotation(10),
            T.RandomHorizontalFlip(),
            T.RandomVerticalFlip(),
            T.ToTensor(),
            T.Normalize(mean=0.5, std=0.5)
        ])

    def __getitem__(self, index):
        file_name = self.df['filename'].iloc[index]
        img = Image.open(file_name)
        img = self.transform(img)
        return img,\
                paddle.to_tensor(self.df['period'].iloc[index]),\
                paddle.to_tensor(self.df['weather'].iloc[index])

    def __len__(self):
        return len(self.df)

# 训练集
train_dataset = WeatherDataset(train_json.iloc[:-500])
train_loader = DataLoader(train_dataset, batch_size=64, shuffle=True)

# 验证集
val_dataset = WeatherDataset(train_json.iloc[-500:])
val_loader = DataLoader(val_dataset, batch_size=64, shuffle=True)

搭建模型

from paddle.vision.models import resnet18

# 自定义模型，模型有两个head
class WeatherModel(paddle.nn.Layer):
    def __init__(self):
        super(WeatherModel, self).__init__()
        backbone = resnet18(pretrained=True)
        backbone.fc = paddle.nn.Identity()
        self.backbone = backbone

        # 分类1
        self.fc1 = paddle.nn.Linear(512, 4)

        # 分类2
        self.fc2 = paddle.nn.Linear(512, 3)

    def forward(self, x):
        out = self.backbone(x)

        # 同时完成类别1 和 类别2 分类
        logits1 = self.fc1(out)
        logits2 = self.fc2(out)
        return logits1, logits2

model = WeatherModel()
model(paddle.to_tensor(np.random.rand(10, 3, 256, 256).astype(np.float32)))

W1203 00:28:16.542893  7854 device_context.cc:447] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 10.1, Runtime API Version: 10.1
W1203 00:28:16.546566  7854 device_context.cc:465] device: 0, cuDNN Version: 7.6.
INFO:paddle.utils.download:unique_endpoints {''}
INFO:paddle.utils.download:File /home/aistudio/.cache/paddle/hapi/weights/resnet18.pdparams md5 checking...
INFO:paddle.utils.download:Found /home/aistudio/.cache/paddle/hapi/weights/resnet18.pdparams





(Tensor(shape=[10, 4], dtype=float32, place=CUDAPlace(0), stop_gradient=False,
        [[ 0.82836318,  1.35641491,  0.46156025, -1.45563996],
         [ 0.08160630,  1.42760110,  0.87870872, -2.34320068],
         [ 0.33396342,  1.35725105, -0.14363584, -2.19671369],
         [ 0.07748616, -0.12090680, -0.29871672, -1.20805240],
         [ 0.11444393,  1.95767570,  0.52079809, -1.79601681],
         [-1.35744286,  0.13195890, -0.26686692, -1.47781038],
         [ 0.99778777, -0.46945763,  1.13994253, -0.39733961],
         [-0.46440995, -0.54322171, -0.02341729, -1.50987113],
         [-0.85322177,  1.23348832, -0.56562370, -2.12171173],
         [-0.74074465,  2.26537490,  1.03538465, -2.16883540]]),
 Tensor(shape=[10, 3], dtype=float32, place=CUDAPlace(0), stop_gradient=False,
        [[ 0.27792573,  0.37752092,  0.09397808],
         [-1.53032672,  1.26350784,  0.16554639],
         [-0.11922672,  2.85570526, -1.09816480],
         [ 0.15814775, -0.65982032, -0.02658810],
         [-0.84676772,  1.20923686, -0.41986156],
         [-0.72880727,  0.98558438, -1.09161079],
         [-0.65401202,  1.84503722, -0.01031107],
         [-0.10907726,  0.86592245,  0.49232167],
         [-0.59027153,  0.95352286, -0.89762455],
         [-0.68332863,  1.41990232, -1.38094914]]))

训练与验证

# 定义损失函数和优化器
optimizer = paddle.optimizer.Adam(parameters=model.parameters(), learning_rate=0.0001)
criterion = paddle.nn.CrossEntropyLoss()

for epoch in range(0, 1):
    Train_Loss, Val_Loss = [], []
    Train_ACC1, Train_ACC2 = [], []
    Val_ACC1, Val_ACC2 = [], []
    
    # 模型训练
    model.train()
    for i, (x, y1, y2) in enumerate(train_loader):
        pred1, pred2 = model(x)

        # 类别1 loss + 类别2 loss 为总共的loss
        loss = criterion(pred1, y1) + criterion(pred2, y2)
        Train_Loss.append(loss.item())
        loss.backward()
        optimizer.step()
        optimizer.clear_grad()

        Train_ACC1.append((pred1.argmax(1) == y1.flatten()).numpy().mean())
        Train_ACC2.append((pred2.argmax(1) == y2.flatten()).numpy().mean())

    # 模型验证
    model.eval()
    for i, (x, y1, y2) in enumerate(val_loader):
        pred1, pred2 = model(x)
        loss = criterion(pred1, y1) + criterion(pred2, y2)
        Val_Loss.append(loss.item())
        Val_ACC1.append((pred1.argmax(1) == y1.flatten()).numpy().mean())
        Val_ACC2.append((pred2.argmax(1) == y2.flatten()).numpy().mean())

    if epoch % 1 == 0:
        print(f'\nEpoch: {epoch}')
        print(f'Loss {np.mean(Train_Loss):3.5f}/{np.mean(Val_Loss):3.5f}')
        print(f'period.ACC {np.mean(Train_ACC1):3.5f}/{np.mean(Val_ACC1):3.5f}')
        print(f'weather.ACC {np.mean(Train_ACC2):3.5f}/{np.mean(Val_ACC2):3.5f}')

Epoch: 0
Loss 1.60270/1.57734
period.ACC 0.64383/0.72972
weather.ACC 0.72986/0.62680

预测与提交

import glob

# 测试集数据路径
test_df = pd.DataFrame({'filename': glob.glob('./test_images/*.jpg')})
test_df['period'] = 0
test_df['weather'] = 0
test_df = test_df.sort_values(by='filename')

test_dataset = WeatherDataset(test_df)
test_loader = DataLoader(test_dataset, batch_size=64, shuffle=False)

model.eval()
period_pred = []
weather_pred = []

# 测试集进行预测
for i, (x, y1, y2) in enumerate(test_loader):
    pred1, pred2 = model(x)
    period_pred += period_dict[pred1.argmax(1).numpy()].tolist()
    weather_pred += weather_dict[pred2.argmax(1).numpy()].tolist()

test_df['period'] = period_pred
test_df['weather'] = weather_pred

submit_json = {
    'annotations':[]
}

# 生成测试集结果
for row in test_df.iterrows():
    submit_json['annotations'].append({
        'filename': 'test_images\\' + row[1].filename.split('/')[-1],
        'period': row[1].period,
        'weather': row[1].weather,
    })

with open('submit.json', 'w') as up:
    json.dump(submit_json, up)

总结和展望

  1. 项目使用预训练模型 + 多分类分支完成天气分类和时间分类操作。
  2. 可以考虑在项目中加入额外的数据扩增方法和交叉验证来改建模型精度。
# 深度学习 # 分类 # paddlepaddle # 人工智能
Logo
百度飞桨AI Studio社区

学大模型，用大模型上飞桨星河社区！每天8点V100G算力免费领！免费领取ERNIE 4.0 100w Token >>>

更多推荐

cover

利用Amazon Bedrock畅玩Claude 3等多种领先模型，抢占AI高地(体验倒计时4小时)

avatar 百度飞桨星河社区
cover

LangChain的RAG实践

avatar 百度飞桨星河社区
cover

RAG综述汇总

avatar 百度飞桨星河社区