CCF BDCI人工智能比赛-天气以及时间分类赛题baseline

来源：投稿 作者：livingbody
编辑：学姐

文章较长，请留意阅读时间

关注公众号【学姐带你玩AI】

回复“天气”领取完整代码

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。参赛者需基于Oneflow框架在训练集上进行训练，对测试集中照片的天气和时间进行分类。

数据简介

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：多云、晴天、雨天、雪天和雾天5个类别

时间：黎明、早上、下午、黄昏、夜晚5个类别

下午 多云

早上 雨天

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：

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

提交示例

{ “annotations”: [ { “filename”: “test_images\00008.jpg”, “period”: “Morning”, “weather”: “Cloudy” }] }

解题思路

总体上看，该任务可以分为2个：一个是预测时间、一个是预测天气。

具体如下：

• 预测时间、天气数据标签列表生成

• 数据集划分

• 数据均衡（数据很不均衡）

• 分别预测

• 合并预测结果

数据集准备

数据下载

# 直接下载，速度超快
!wget https://awscdn.datafountain.cn/cometition_data2/Files/BDCI2021/555/train_dataset.zip
!wget https://awscdn.datafountain.cn/cometition_data2/Files/BDCI2021/555/test_dataset.zip
!wget https://awscdn.datafountain.cn/cometition_data2/Files/BDCI2021/555/submit_example.json

--2022-01-17 11:05:49--  https://awscdn.datafountain.cn/cometition_data2/Files/BDCI2021/555/train_dataset.zip
Resolving awscdn.datafountain.cn (awscdn.datafountain.cn)... 210.51.40.148, 210.51.40.156, 210.51.40.133, ...
Connecting to awscdn.datafountain.cn (awscdn.datafountain.cn)|210.51.40.148|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 419324853 (400M) [application/octet-stream]
Saving to: ‘train_dataset.zip’

train_dataset.zip   100%[===================>] 399.90M   102MB/s    in 4.3s    

2022-01-17 11:05:54 (94.0 MB/s) - ‘train_dataset.zip’ saved [419324853/419324853]

--2022-01-17 11:05:54--  https://awscdn.datafountain.cn/cometition_data2/Files/BDCI2021/555/test_dataset.zip
Resolving awscdn.datafountain.cn (awscdn.datafountain.cn)... 210.51.40.156, 210.51.40.150, 210.51.40.133, ...
Connecting to awscdn.datafountain.cn (awscdn.datafountain.cn)|210.51.40.156|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 62269247 (59M) [application/octet-stream]
Saving to: ‘test_dataset.zip’

test_dataset.zip    100%[===================>]  59.38M  65.4MB/s    in 0.9s    

2022-01-17 11:05:55 (65.4 MB/s) - ‘test_dataset.zip’ saved [62269247/62269247]

--2022-01-17 11:05:55--  https://awscdn.datafountain.cn/cometition_data2/Files/BDCI2021/555/submit_example.json
Resolving awscdn.datafountain.cn (awscdn.datafountain.cn)... 210.51.40.156, 210.51.40.148, 210.51.40.150, ...
Connecting to awscdn.datafountain.cn (awscdn.datafountain.cn)|210.51.40.156|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 593 [application/octet-stream]
Saving to: ‘submit_example.json’

submit_example.json 100%[===================>]     593  --.-KB/s    in 0s      

2022-01-17 11:05:56 (187 MB/s) - ‘submit_example.json’ saved [593/593]

数据解压缩

# 解压缩数据集
!unzip -qoa test_dataset.zip 
!unzip -qoa train_dataset.zip

按时间制作标签

注意事项：虽然数据描述说时间** Period 为 黎明、早上、下午、黄昏、夜晚**，但是经过遍历发现只有4类。。。。。，故如下制作标签

# 标签修改
%cd ~
import json
import os

train = {}
with open('train.json', 'r') as f:
    train = json.load(f)

period_list = {'Dawn': 0, 'Dusk': 1, 'Morning': 2, 'Afternoon': 3}
f_period=open('train_period.txt','w')
for item in train["annotations"]:
    label = period_list[item['period']] 
    file_name=os.path.join(item['filename'].split('\\')[0], item['filename'].split('\\')[1])
    f_period.write(file_name +' '+ str(label) +'\n')
f_period.close()
print("写入train_period.txt完成！！！")

/home/aistudio
写入train_period.txt完成！！！

数据集划分并数据均衡

# 数据分析
%cd ~
import pandas as pd
from matplotlib import pyplot as plt
%matplotlib inline

data=pd.read_csv('train_period.txt', header=None, sep=' ')
print(data[1].value_counts())
data[1].value_counts().plot(kind="bar")

/home/aistudio
2    1613
3     829
1     124
0      34
Name: 1, dtype: int64





<matplotlib.axes._subplots.AxesSubplot at 0x7feffe438b50>

# 训练集、测试集划分
import pandas as pd
import os
from sklearn.model_selection import train_test_split

def split_dataset(data_file):
    # 展示不同的调用方式
    data = pd.read_csv(data_file, header=None, sep=' ')
    train_dataset, eval_dataset = train_test_split(data, test_size=0.2, random_state=42)
    print(f'train dataset len: {train_dataset.size}')
    print(f'eval dataset len: {eval_dataset.size}')
    train_filename='train_' + data_file.split('.')[0]+'.txt'
    eval_filename='eval_' + data_file.split('.')[0]+'.txt'
    train_dataset.to_csv(train_filename, index=None, header=None, sep=' ')
    eval_dataset.to_csv(eval_filename, index=None, header=None, sep=' ')
    

data_file='train_period.txt'
split_dataset(data_file)

train dataset len: 4160
eval dataset len: 1040

# pip更新或安装包后需要重启notebook
!pip install -U scikit-learn
# 数据均衡用
!pip install -U imblearn

# 数据均衡
import pandas as pd
from collections import Counter
from imblearn.over_sampling import SMOTE
import numpy as np

def upsampleing(filename):
    print(50 * '*')
    data = pd.read_csv(filename, header=None, sep=' ')
    print(data[1].value_counts())
    # 查看各个标签的样本量
    print(Counter(data[1]))
    print(50 * '*')
    # 数据均衡
    X = np.array(data[0].index.tolist()).reshape(-1, 1)
    y = data[1]
    ros = SMOTE(random_state=0)
    X_resampled, y_resampled = ros.fit_resample(X, y)
    print(Counter(y_resampled))
    print(len(y_resampled))
    print(50 * '*')
    img_list=[]
    for i in range(len(X_resampled)):
        img_list.append(data.loc[X_resampled[i]][0].tolist()[0])
    dict_weather={'0':img_list, '1':y_resampled.values}
    newdata=pd.DataFrame(dict_weather)
    print(len(newdata))
    new_filename=filename.split('.')[0]+'_imblearn'+'.txt'
    newdata.to_csv(new_filename, header=None, index=None, sep=' ')
    
    

filename='train_train_period.txt'
upsampleing(filename)
filename='eval_train_period.txt'
upsampleing(filename)

**************************************************
2    1304
3     653
1      95
0      28
Name: 1, dtype: int64
Counter({2: 1304, 3: 653, 1: 95, 0: 28})
**************************************************
Counter({2: 1304, 3: 1304, 1: 1304, 0: 1304})
5216
**************************************************
5216
**************************************************
2    309
3    176
1     29
0      6
Name: 1, dtype: int64
Counter({2: 309, 3: 176, 1: 29, 0: 6})
**************************************************
Counter({2: 309, 3: 309, 1: 309, 0: 309})
1236
**************************************************
1236

按天气分制作标签

import json
import os

train = {}
with open('train.json', 'r') as f:
    train = json.load(f)

weather_list =  {'Cloudy': 0, 'Rainy': 1, 'Sunny': 2}
f_weather=open('train_weather.txt','w')
for item in train["annotations"]:
    label = weather_list[item['weather']] 
    file_name=os.path.join(item['filename'].split('\\')[0], item['filename'].split('\\')[1])
    f_weather.write(file_name +' '+ str(label) +'\n')
f_weather.close()
print("写入train_weather.txt完成！！！")

写入train_weather.txt完成！！！

数据集划分并均衡

import pandas as pd
from matplotlib import pyplot as plt

data=pd.read_csv('train_weather.txt', header=None, sep=' ')
print(data[1].value_counts())
data[1].value_counts().plot(kind="bar")

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





<matplotlib.axes._subplots.AxesSubplot at 0x7feffe82d190>

# 训练集、测试集划分

import pandas as pd
import os
from sklearn.model_selection import train_test_split

def split_dataset(data_file):
    # 展示不同的调用方式
    data = pd.read_csv(data_file, header=None, sep=' ')
    train_dataset, eval_dataset = train_test_split(data, test_size=0.2, random_state=42)
    print(f'train dataset len: {train_dataset.size}')
    print(f'eval dataset len: {eval_dataset.size}')
    train_filename='train_' + data_file.split('.')[0]+'.txt'
    eval_filename='eval_' + data_file.split('.')[0]+'.txt'
    train_dataset.to_csv(train_filename, index=None, header=None, sep=' ')
    eval_dataset.to_csv(eval_filename, index=None, header=None, sep=' ')
    

data_file='train_weather.txt'
split_dataset(data_file)

train dataset len: 4160
eval dataset len: 1040

# 数据均衡
import pandas as pd
from collections import Counter
from imblearn.over_sampling import SMOTE
import numpy as np

def upsampleing(filename):
    print(50 * '*')
    data = pd.read_csv(filename, header=None, sep=' ')
    print(data[1].value_counts())
    # 查看各个标签的样本量
    print(Counter(data[1]))
    print(50 * '*')
    # 数据均衡
    X = np.array(data[0].index.tolist()).reshape(-1, 1)
    y = data[1]
    ros = SMOTE(random_state=0)
    X_resampled, y_resampled = ros.fit_resample(X, y)
    print(Counter(y_resampled))
    print(len(y_resampled))
    print(50 * '*')
    img_list=[]
    for i in range(len(X_resampled)):
        img_list.append(data.loc[X_resampled[i]][0].tolist()[0])
    dict_weather={'0':img_list, '1':y_resampled.values}
    newdata=pd.DataFrame(dict_weather)
    print(len(newdata))
    new_filename=filename.split('.')[0]+'_imblearn'+'.txt'
    newdata.to_csv(new_filename, header=None, index=None, sep=' ')
    
    

filename='train_train_weather.txt'
upsampleing(filename)
filename='eval_train_weather.txt'
upsampleing(filename)

**************************************************
0    892
2    715
1    473
Name: 1, dtype: int64
Counter({0: 892, 2: 715, 1: 473})
**************************************************
Counter({0: 892, 2: 892, 1: 892})
2676
**************************************************
2676
**************************************************
0    227
2    171
1    122
Name: 1, dtype: int64
Counter({0: 227, 2: 171, 1: 122})
**************************************************
Counter({0: 227, 2: 227, 1: 227})
681
**************************************************
681

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。此次使用PaddleClas框架完成比赛。

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

fatal: destination path 'PaddleClas' already exists and is not an empty directory.

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

模型训练 and 评估

时间训练

以PaddleClas/ppcls/configs/ImageNet/VisionTransformer/ViT_small_patch16_224.yaml 为基础进行修改

# global configs
Global:
  checkpoints: null
  pretrained_model: null
  output_dir: ./output/
  device: gpu
  save_interval: 1
  eval_during_train: True
  eval_interval: 1
  epochs: 120
  print_batch_step: 10
  use_visualdl: False
  # used for static mode and model export
  image_shape: [3, 224, 224]
  save_inference_dir: ./inference

# model architecture
Arch:
  name: ViT_small_patch16_224
  class_num: 1000
 
# 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: Piecewise
    learning_rate: 0.1
    decay_epochs: [30, 60, 90]
    values: [0.1, 0.01, 0.001, 0.0001]
  regularizer:
    name: 'L2'
    coeff: 0.0001


# data loader for train and eval
DataLoader:
  Train:
    dataset:
      name: ImageNetDataset
      image_root: ./dataset/ILSVRC2012/
      cls_label_path: ./dataset/ILSVRC2012/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: [0.5, 0.5, 0.5]
            std: [0.5, 0.5, 0.5]
            order: ''

    sampler:
      name: DistributedBatchSampler
      batch_size: 64
      drop_last: False
      shuffle: True
    loader:
      num_workers: 4
      use_shared_memory: True

  Eval:
    dataset: 
      name: ImageNetDataset
      image_root: ./dataset/ILSVRC2012/
      cls_label_path: ./dataset/ILSVRC2012/val_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: [0.5, 0.5, 0.5]
            std: [0.5, 0.5, 0.5]
            order: ''
    sampler:
      name: DistributedBatchSampler
      batch_size: 64
      drop_last: False
      shuffle: False
    loader:
      num_workers: 4
      use_shared_memory: True

Infer:
  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: [0.5, 0.5, 0.5]
        std: [0.5, 0.5, 0.5]
        order: ''
    - ToCHWImage:
  PostProcess:
    name: Topk
    topk: 5
    class_id_map_file: ppcls/utils/imagenet1k_label_list.txt

Metric:
  Train:
    - TopkAcc:
        topk: [1, 5]
  Eval:
    - TopkAcc:
        topk: [1, 5]

# 覆盖配置
%cd ~
!cp -f ~/ViT_small_patch16_224.yaml ~/ppcls/configs/ImageNet/VisionTransformer/ViT_small_patch16_224.yaml

/home/aistudio

# 开始训练
%cd ~/PaddleClas/

!python3 tools/train.py \
    -c ./ppcls/configs/ImageNet/VisionTransformer/ViT_base_patch16_224.yaml \
    -o Arch.pretrained=True \
    -o Global.pretrained_model=./output/ViT_base_patch16_224/epoch_21 \
    -o Global.device=gpu

/home/aistudio/PaddleClas

# 模型评估
%cd ~/PaddleClas/

!python  tools/eval.py \
        -c  ./ppcls/configs/ImageNet/VisionTransformer/ViT_base_patch16_224.yaml \
        -o Global.pretrained_model=./output/ViT_base_patch16_224/best_model

/home/aistudio/PaddleClas
[2022/01/18 00:16:10] root INFO: 
===========================================================
==        PaddleClas is powered by PaddlePaddle !        ==
===========================================================
==                                                       ==
==   For more info please go to the following website.   ==
==                                                       ==
==       https://github.com/PaddlePaddle/PaddleClas      ==
===========================================================

[2022/01/18 00:16:10] root INFO: Arch : 
[2022/01/18 00:16:10] root INFO:     class_num : 4
[2022/01/18 00:16:10] root INFO:     name : ViT_base_patch16_224
[2022/01/18 00:16:10] root INFO: DataLoader : 
[2022/01/18 00:16:10] root INFO:     Eval : 
[2022/01/18 00:16:10] root INFO:         dataset : 
[2022/01/18 00:16:10] root INFO:             cls_label_path : /home/aistudio/eval_train_period_imblearn.txt
[2022/01/18 00:16:10] root INFO:             image_root : /home/aistudio/
[2022/01/18 00:16:10] root INFO:             name : ImageNetDataset
[2022/01/18 00:16:10] root INFO:             transform_ops : 
[2022/01/18 00:16:10] root INFO:                 DecodeImage : 
[2022/01/18 00:16:10] root INFO:                     channel_first : False
[2022/01/18 00:16:10] root INFO:                     to_rgb : True
[2022/01/18 00:16:10] root INFO:                 ResizeImage : 
[2022/01/18 00:16:10] root INFO:                     resize_short : 256
[2022/01/18 00:16:10] root INFO:                 CropImage : 
[2022/01/18 00:16:10] root INFO:                     size : 224
[2022/01/18 00:16:10] root INFO:                 NormalizeImage : 
[2022/01/18 00:16:10] root INFO:                     mean : [0.5, 0.5, 0.5]
[2022/01/18 00:16:10] root INFO:                     order : 
[2022/01/18 00:16:10] root INFO:                     scale : 1.0/255.0
[2022/01/18 00:16:10] root INFO:                     std : [0.5, 0.5, 0.5]
[2022/01/18 00:16:10] root INFO:         loader : 
[2022/01/18 00:16:10] root INFO:             num_workers : 4
[2022/01/18 00:16:10] root INFO:             use_shared_memory : True
[2022/01/18 00:16:10] root INFO:         sampler : 
[2022/01/18 00:16:10] root INFO:             batch_size : 128
[2022/01/18 00:16:10] root INFO:             drop_last : False
[2022/01/18 00:16:10] root INFO:             name : DistributedBatchSampler
[2022/01/18 00:16:10] root INFO:             shuffle : False
[2022/01/18 00:16:10] root INFO:     Train : 
[2022/01/18 00:16:10] root INFO:         dataset : 
[2022/01/18 00:16:10] root INFO:             cls_label_path : /home/aistudio/train_train_period_imblearn.txt
[2022/01/18 00:16:10] root INFO:             image_root : /home/aistudio
[2022/01/18 00:16:10] root INFO:             name : ImageNetDataset
[2022/01/18 00:16:10] root INFO:             transform_ops : 
[2022/01/18 00:16:10] root INFO:                 DecodeImage : 
[2022/01/18 00:16:10] root INFO:                     channel_first : False
[2022/01/18 00:16:10] root INFO:                     to_rgb : True
[2022/01/18 00:16:10] root INFO:                 RandCropImage : 
[2022/01/18 00:16:10] root INFO:                     size : 224
[2022/01/18 00:16:10] root INFO:                 RandFlipImage : 
[2022/01/18 00:16:10] root INFO:                     flip_code : 1
[2022/01/18 00:16:10] root INFO:                 NormalizeImage : 
[2022/01/18 00:16:10] root INFO:                     mean : [0.5, 0.5, 0.5]
[2022/01/18 00:16:10] root INFO:                     order : 
[2022/01/18 00:16:10] root INFO:                     scale : 1.0/255.0
[2022/01/18 00:16:10] root INFO:                     std : [0.5, 0.5, 0.5]
[2022/01/18 00:16:10] root INFO:         loader : 
[2022/01/18 00:16:10] root INFO:             num_workers : 4
[2022/01/18 00:16:10] root INFO:             use_shared_memory : True
[2022/01/18 00:16:10] root INFO:         sampler : 
[2022/01/18 00:16:10] root INFO:             batch_size : 160
[2022/01/18 00:16:10] root INFO:             drop_last : False
[2022/01/18 00:16:10] root INFO:             name : DistributedBatchSampler
[2022/01/18 00:16:10] root INFO:             shuffle : True
[2022/01/18 00:16:10] root INFO: Global : 
[2022/01/18 00:16:10] root INFO:     checkpoints : None
[2022/01/18 00:16:10] root INFO:     device : gpu
[2022/01/18 00:16:10] root INFO:     epochs : 120
[2022/01/18 00:16:10] root INFO:     eval_during_train : True
[2022/01/18 00:16:10] root INFO:     eval_interval : 1
[2022/01/18 00:16:10] root INFO:     image_shape : [3, 224, 224]
[2022/01/18 00:16:10] root INFO:     output_dir : ./output/
[2022/01/18 00:16:10] root INFO:     pretrained_model : ./output/ViT_base_patch16_224/best_model
[2022/01/18 00:16:10] root INFO:     print_batch_step : 10
[2022/01/18 00:16:10] root INFO:     save_inference_dir : ./inference
[2022/01/18 00:16:10] root INFO:     save_interval : 1
[2022/01/18 00:16:10] root INFO:     use_visualdl : False
[2022/01/18 00:16:10] root INFO: Infer : 
[2022/01/18 00:16:10] root INFO:     PostProcess : 
[2022/01/18 00:16:10] root INFO:         class_id_map_file : ppcls/utils/imagenet1k_label_list.txt
[2022/01/18 00:16:10] root INFO:         name : Topk
[2022/01/18 00:16:10] root INFO:         topk : 5
[2022/01/18 00:16:10] root INFO:     batch_size : 10
[2022/01/18 00:16:10] root INFO:     infer_imgs : docs/images/whl/demo.jpg
[2022/01/18 00:16:10] root INFO:     transforms : 
[2022/01/18 00:16:10] root INFO:         DecodeImage : 
[2022/01/18 00:16:10] root INFO:             channel_first : False
[2022/01/18 00:16:10] root INFO:             to_rgb : True
[2022/01/18 00:16:10] root INFO:         ResizeImage : 
[2022/01/18 00:16:10] root INFO:             resize_short : 256
[2022/01/18 00:16:10] root INFO:         CropImage : 
[2022/01/18 00:16:10] root INFO:             size : 224
[2022/01/18 00:16:10] root INFO:         NormalizeImage : 
[2022/01/18 00:16:10] root INFO:             mean : [0.5, 0.5, 0.5]
[2022/01/18 00:16:10] root INFO:             order : 
[2022/01/18 00:16:10] root INFO:             scale : 1.0/255.0
[2022/01/18 00:16:10] root INFO:             std : [0.5, 0.5, 0.5]
[2022/01/18 00:16:10] root INFO:         ToCHWImage : None
[2022/01/18 00:16:10] root INFO: Loss : 
[2022/01/18 00:16:10] root INFO:     Eval : 
[2022/01/18 00:16:10] root INFO:         CELoss : 
[2022/01/18 00:16:10] root INFO:             weight : 1.0
[2022/01/18 00:16:10] root INFO:     Train : 
[2022/01/18 00:16:10] root INFO:         CELoss : 
[2022/01/18 00:16:10] root INFO:             weight : 1.0
[2022/01/18 00:16:10] root INFO: Metric : 
[2022/01/18 00:16:10] root INFO:     Eval : 
[2022/01/18 00:16:10] root INFO:         TopkAcc : 
[2022/01/18 00:16:10] root INFO:             topk : [1, 2]
[2022/01/18 00:16:10] root INFO:     Train : 
[2022/01/18 00:16:10] root INFO:         TopkAcc : 
[2022/01/18 00:16:10] root INFO:             topk : [1, 2]
[2022/01/18 00:16:10] root INFO: Optimizer : 
[2022/01/18 00:16:10] root INFO:     lr : 
[2022/01/18 00:16:10] root INFO:         decay_epochs : [10, 22, 30]
[2022/01/18 00:16:10] root INFO:         learning_rate : 0.01
[2022/01/18 00:16:10] root INFO:         name : Piecewise
[2022/01/18 00:16:10] root INFO:         values : [0.01, 0.001, 0.0001, 1e-05]
[2022/01/18 00:16:10] root INFO:     momentum : 0.9
[2022/01/18 00:16:10] root INFO:     name : Momentum
[2022/01/18 00:16:10] root INFO:     regularizer : 
[2022/01/18 00:16:10] root INFO:         coeff : 0.0001
[2022/01/18 00:16:10] root INFO:         name : L2
[2022/01/18 00:16:10] root INFO: train with paddle 2.2.1 and device CUDAPlace(0)
W0118 00:16:10.219051  1924 device_context.cc:447] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 10.1, Runtime API Version: 10.1
W0118 00:16:10.224006  1924 device_context.cc:465] device: 0, cuDNN Version: 7.6.
[2022/01/18 00:16:21] root INFO: [Eval][Epoch 0][Iter: 0/10]CELoss: 0.68425, loss: 0.68425, top1: 0.79688, top2: 0.92969, batch_cost: 6.03284s, reader_cost: 5.55918, ips: 21.21719 images/sec
[2022/01/18 00:16:30] root INFO: [Eval][Epoch 0][Avg]CELoss: 1.24542, loss: 1.24542, top1: 0.47006, top2: 0.73463

天气训练

配置文件为：

** PaddleClas/ppcls/configs/ImageNet/VisionTransformer/ViT_base_patch16_224_weather.yaml**

# global configs
Global:
  checkpoints: null
  pretrained_model: null
  output_dir: ./output_weather/
  device: gpu
  save_interval: 1
  eval_during_train: True
  eval_interval: 1
  epochs: 120
  print_batch_step: 10
  use_visualdl: False
  # used for static mode and model export
  image_shape: [3, 224, 224]
  save_inference_dir: ./inference_weather

# model architecture
Arch:
  name: ViT_base_patch16_224
  class_num: 3
 
# 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: Piecewise
    learning_rate: 0.01
    decay_epochs: [10, 22, 30]
    values: [0.01, 0.001, 0.0001, 0.00001]
  regularizer:
    name: 'L2'
    coeff: 0.0001


# data loader for train and eval
DataLoader:
  Train:
    dataset:
      name: ImageNetDataset
      image_root: /home/aistudio
      cls_label_path: /home/aistudio/train_train_weather_imblearn.txt
      transform_ops:
        - DecodeImage:
            to_rgb: True
            channel_first: False
        - RandCropImage:
            size: 224
        - RandFlipImage:
            flip_code: 1
        - NormalizeImage:
            scale: 1.0/255.0
            mean: [0.5, 0.5, 0.5]
            std: [0.5, 0.5, 0.5]
            order: ''

    sampler:
      name: DistributedBatchSampler
      batch_size: 160
      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_train_weather_imblearn.txt
      transform_ops:
        - DecodeImage:
            to_rgb: True
            channel_first: False
        - ResizeImage:
            resize_short: 256
        - CropImage:
            size: 224
        - NormalizeImage:
            scale: 1.0/255.0
            mean: [0.5, 0.5, 0.5]
            std: [0.5, 0.5, 0.5]
            order: ''
    sampler:
      name: DistributedBatchSampler
      batch_size: 128
      drop_last: False
      shuffle: False
    loader:
      num_workers: 4
      use_shared_memory: True

Infer:
  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: [0.5, 0.5, 0.5]
        std: [0.5, 0.5, 0.5]
        order: ''
    - ToCHWImage:
  PostProcess:
    name: Topk
    topk: 5
    class_id_map_file: ppcls/utils/imagenet1k_label_list.txt

Metric:
  Train:
    - TopkAcc:
        topk: [1, 2]
  Eval:
    - TopkAcc:
        topk: [1, 2]

# 覆盖配置
%cd ~
!cp -f  ~/ViT_small_patch16_224_weather.yaml ~/ppcls/configs/ImageNet/VisionTransformer/ViT_small_patch16_224_weather.yaml

/home/aistudio
cp: cannot stat 'PaddleClas/ppcls/configs/ImageNet/VisionTransformer/ViT_small_patch16_224_weather.yaml': No such file or directory

# 模型训练
%cd ~/PaddleClas/

!python3 tools/train.py \
    -c ./ppcls/configs/ImageNet/VisionTransformer/ViT_base_patch16_224_weather.yaml \
    -o Arch.pretrained=True \
    -o Global.device=gpu

# 模型评估
%cd ~/PaddleClas/

!python  tools/eval.py \
        -c  ./ppcls/configs/ImageNet/VisionTransformer/ViT_base_patch16_224_weather.yaml \
        -o Global.pretrained_model=./output_weather/ViT_base_patch16_224/best_model

/home/aistudio/PaddleClas
[2022/01/18 00:17:24] root INFO: 
===========================================================
==        PaddleClas is powered by PaddlePaddle !        ==
===========================================================
==                                                       ==
==   For more info please go to the following website.   ==
==                                                       ==
==       https://github.com/PaddlePaddle/PaddleClas      ==
===========================================================

[2022/01/18 00:17:24] root INFO: Arch : 
[2022/01/18 00:17:24] root INFO:     class_num : 3
[2022/01/18 00:17:24] root INFO:     name : ViT_base_patch16_224
[2022/01/18 00:17:24] root INFO: DataLoader : 
[2022/01/18 00:17:24] root INFO:     Eval : 
[2022/01/18 00:17:24] root INFO:         dataset : 
[2022/01/18 00:17:24] root INFO:             cls_label_path : /home/aistudio/eval_train_weather_imblearn.txt
[2022/01/18 00:17:24] root INFO:             image_root : /home/aistudio/
[2022/01/18 00:17:24] root INFO:             name : ImageNetDataset
[2022/01/18 00:17:24] root INFO:             transform_ops : 
[2022/01/18 00:17:24] root INFO:                 DecodeImage : 
[2022/01/18 00:17:24] root INFO:                     channel_first : False
[2022/01/18 00:17:24] root INFO:                     to_rgb : True
[2022/01/18 00:17:24] root INFO:                 ResizeImage : 
[2022/01/18 00:17:24] root INFO:                     resize_short : 256
[2022/01/18 00:17:24] root INFO:                 CropImage : 
[2022/01/18 00:17:24] root INFO:                     size : 224
[2022/01/18 00:17:24] root INFO:                 NormalizeImage : 
[2022/01/18 00:17:24] root INFO:                     mean : [0.5, 0.5, 0.5]
[2022/01/18 00:17:24] root INFO:                     order : 
[2022/01/18 00:17:24] root INFO:                     scale : 1.0/255.0
[2022/01/18 00:17:24] root INFO:                     std : [0.5, 0.5, 0.5]
[2022/01/18 00:17:24] root INFO:         loader : 
[2022/01/18 00:17:24] root INFO:             num_workers : 4
[2022/01/18 00:17:24] root INFO:             use_shared_memory : True
[2022/01/18 00:17:24] root INFO:         sampler : 
[2022/01/18 00:17:24] root INFO:             batch_size : 128
[2022/01/18 00:17:24] root INFO:             drop_last : False
[2022/01/18 00:17:24] root INFO:             name : DistributedBatchSampler
[2022/01/18 00:17:24] root INFO:             shuffle : False
[2022/01/18 00:17:24] root INFO:     Train : 
[2022/01/18 00:17:24] root INFO:         dataset : 
[2022/01/18 00:17:24] root INFO:             cls_label_path : /home/aistudio/train_train_weather_imblearn.txt
[2022/01/18 00:17:24] root INFO:             image_root : /home/aistudio
[2022/01/18 00:17:24] root INFO:             name : ImageNetDataset
[2022/01/18 00:17:24] root INFO:             transform_ops : 
[2022/01/18 00:17:24] root INFO:                 DecodeImage : 
[2022/01/18 00:17:24] root INFO:                     channel_first : False
[2022/01/18 00:17:24] root INFO:                     to_rgb : True
[2022/01/18 00:17:24] root INFO:                 RandCropImage : 
[2022/01/18 00:17:24] root INFO:                     size : 224
[2022/01/18 00:17:24] root INFO:                 RandFlipImage : 
[2022/01/18 00:17:24] root INFO:                     flip_code : 1
[2022/01/18 00:17:24] root INFO:                 NormalizeImage : 
[2022/01/18 00:17:24] root INFO:                     mean : [0.5, 0.5, 0.5]
[2022/01/18 00:17:24] root INFO:                     order : 
[2022/01/18 00:17:24] root INFO:                     scale : 1.0/255.0
[2022/01/18 00:17:24] root INFO:                     std : [0.5, 0.5, 0.5]
[2022/01/18 00:17:24] root INFO:         loader : 
[2022/01/18 00:17:24] root INFO:             num_workers : 4
[2022/01/18 00:17:24] root INFO:             use_shared_memory : True
[2022/01/18 00:17:24] root INFO:         sampler : 
[2022/01/18 00:17:24] root INFO:             batch_size : 160
[2022/01/18 00:17:24] root INFO:             drop_last : False
[2022/01/18 00:17:24] root INFO:             name : DistributedBatchSampler
[2022/01/18 00:17:24] root INFO:             shuffle : True
[2022/01/18 00:17:24] root INFO: Global : 
[2022/01/18 00:17:24] root INFO:     checkpoints : None
[2022/01/18 00:17:24] root INFO:     device : gpu
[2022/01/18 00:17:24] root INFO:     epochs : 120
[2022/01/18 00:17:24] root INFO:     eval_during_train : True
[2022/01/18 00:17:24] root INFO:     eval_interval : 1
[2022/01/18 00:17:24] root INFO:     image_shape : [3, 224, 224]
[2022/01/18 00:17:24] root INFO:     output_dir : ./output_weather/
[2022/01/18 00:17:24] root INFO:     pretrained_model : ./output_weather/ViT_base_patch16_224/best_model
[2022/01/18 00:17:24] root INFO:     print_batch_step : 10
[2022/01/18 00:17:24] root INFO:     save_inference_dir : ./inference
[2022/01/18 00:17:24] root INFO:     save_interval : 1
[2022/01/18 00:17:24] root INFO:     use_visualdl : False
[2022/01/18 00:17:24] root INFO: Infer : 
[2022/01/18 00:17:24] root INFO:     PostProcess : 
[2022/01/18 00:17:24] root INFO:         class_id_map_file : ppcls/utils/imagenet1k_label_list.txt
[2022/01/18 00:17:24] root INFO:         name : Topk
[2022/01/18 00:17:24] root INFO:         topk : 5
[2022/01/18 00:17:24] root INFO:     batch_size : 10
[2022/01/18 00:17:24] root INFO:     infer_imgs : docs/images/whl/demo.jpg
[2022/01/18 00:17:24] root INFO:     transforms : 
[2022/01/18 00:17:24] root INFO:         DecodeImage : 
[2022/01/18 00:17:24] root INFO:             channel_first : False
[2022/01/18 00:17:24] root INFO:             to_rgb : True
[2022/01/18 00:17:24] root INFO:         ResizeImage : 
[2022/01/18 00:17:24] root INFO:             resize_short : 256
[2022/01/18 00:17:24] root INFO:         CropImage : 
[2022/01/18 00:17:24] root INFO:             size : 224
[2022/01/18 00:17:24] root INFO:         NormalizeImage : 
[2022/01/18 00:17:24] root INFO:             mean : [0.5, 0.5, 0.5]
[2022/01/18 00:17:24] root INFO:             order : 
[2022/01/18 00:17:24] root INFO:             scale : 1.0/255.0
[2022/01/18 00:17:24] root INFO:             std : [0.5, 0.5, 0.5]
[2022/01/18 00:17:24] root INFO:         ToCHWImage : None
[2022/01/18 00:17:24] root INFO: Loss : 
[2022/01/18 00:17:24] root INFO:     Eval : 
[2022/01/18 00:17:24] root INFO:         CELoss : 
[2022/01/18 00:17:24] root INFO:             weight : 1.0
[2022/01/18 00:17:24] root INFO:     Train : 
[2022/01/18 00:17:24] root INFO:         CELoss : 
[2022/01/18 00:17:24] root INFO:             weight : 1.0
[2022/01/18 00:17:24] root INFO: Metric : 
[2022/01/18 00:17:24] root INFO:     Eval : 
[2022/01/18 00:17:24] root INFO:         TopkAcc : 
[2022/01/18 00:17:24] root INFO:             topk : [1, 2]
[2022/01/18 00:17:24] root INFO:     Train : 
[2022/01/18 00:17:24] root INFO:         TopkAcc : 
[2022/01/18 00:17:24] root INFO:             topk : [1, 2]
[2022/01/18 00:17:24] root INFO: Optimizer : 
[2022/01/18 00:17:24] root INFO:     lr : 
[2022/01/18 00:17:24] root INFO:         decay_epochs : [10, 22, 30]
[2022/01/18 00:17:24] root INFO:         learning_rate : 0.01
[2022/01/18 00:17:24] root INFO:         name : Piecewise
[2022/01/18 00:17:24] root INFO:         values : [0.01, 0.001, 0.0001, 1e-05]
[2022/01/18 00:17:24] root INFO:     momentum : 0.9
[2022/01/18 00:17:24] root INFO:     name : Momentum
[2022/01/18 00:17:24] root INFO:     regularizer : 
[2022/01/18 00:17:24] root INFO:         coeff : 0.0001
[2022/01/18 00:17:24] root INFO:         name : L2
[2022/01/18 00:17:24] root INFO: train with paddle 2.2.1 and device CUDAPlace(0)
W0118 00:17:24.490124  2035 device_context.cc:447] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 10.1, Runtime API Version: 10.1
W0118 00:17:24.495002  2035 device_context.cc:465] device: 0, cuDNN Version: 7.6.
[2022/01/18 00:17:35] root INFO: [Eval][Epoch 0][Iter: 0/6]CELoss: 0.17755, loss: 0.17755, top1: 0.97656, top2: 1.00000, batch_cost: 5.94345s, reader_cost: 5.49357, ips: 21.53630 images/sec
[2022/01/18 00:17:39] root INFO: [Eval][Epoch 0][Avg]CELoss: 0.54770, loss: 0.54770, top1: 0.83700, top2: 0.95154

预测

时间模型导出

# 模型导出
%cd ~/PaddleClas/
!python tools/export_model.py -c ./ppcls/configs/ImageNet/VisionTransformer/ViT_base_patch16_224.yaml -o Global.pretrained_model=./output/ViT_base_patch16_224/best_model

/home/aistudio/PaddleClas
[2022/01/18 00:34:27] root INFO: 
===========================================================
==        PaddleClas is powered by PaddlePaddle !        ==
===========================================================
==                                                       ==
==   For more info please go to the following website.   ==
==                                                       ==
==       https://github.com/PaddlePaddle/PaddleClas      ==
===========================================================

[2022/01/18 00:34:27] root INFO: Arch : 
[2022/01/18 00:34:27] root INFO:     class_num : 4
[2022/01/18 00:34:27] root INFO:     name : ViT_base_patch16_224
[2022/01/18 00:34:27] root INFO: DataLoader : 
[2022/01/18 00:34:27] root INFO:     Eval : 
[2022/01/18 00:34:27] root INFO:         dataset : 
[2022/01/18 00:34:27] root INFO:             cls_label_path : /home/aistudio/eval_train_period_imblearn.txt
[2022/01/18 00:34:27] root INFO:             image_root : /home/aistudio/
[2022/01/18 00:34:27] root INFO:             name : ImageNetDataset
[2022/01/18 00:34:27] root INFO:             transform_ops : 
[2022/01/18 00:34:27] root INFO:                 DecodeImage : 
[2022/01/18 00:34:27] root INFO:                     channel_first : False
[2022/01/18 00:34:27] root INFO:                     to_rgb : True
[2022/01/18 00:34:27] root INFO:                 ResizeImage : 
[2022/01/18 00:34:27] root INFO:                     resize_short : 256
[2022/01/18 00:34:27] root INFO:                 CropImage : 
[2022/01/18 00:34:27] root INFO:                     size : 224
[2022/01/18 00:34:27] root INFO:                 NormalizeImage : 
[2022/01/18 00:34:27] root INFO:                     mean : [0.5, 0.5, 0.5]
[2022/01/18 00:34:27] root INFO:                     order : 
[2022/01/18 00:34:27] root INFO:                     scale : 1.0/255.0
[2022/01/18 00:34:27] root INFO:                     std : [0.5, 0.5, 0.5]
[2022/01/18 00:34:27] root INFO:         loader : 
[2022/01/18 00:34:27] root INFO:             num_workers : 4
[2022/01/18 00:34:27] root INFO:             use_shared_memory : True
[2022/01/18 00:34:27] root INFO:         sampler : 
[2022/01/18 00:34:27] root INFO:             batch_size : 128
[2022/01/18 00:34:27] root INFO:             drop_last : False
[2022/01/18 00:34:27] root INFO:             name : DistributedBatchSampler
[2022/01/18 00:34:27] root INFO:             shuffle : False
[2022/01/18 00:34:27] root INFO:     Train : 
[2022/01/18 00:34:27] root INFO:         dataset : 
[2022/01/18 00:34:27] root INFO:             cls_label_path : /home/aistudio/train_train_period_imblearn.txt
[2022/01/18 00:34:27] root INFO:             image_root : /home/aistudio
[2022/01/18 00:34:27] root INFO:             name : ImageNetDataset
[2022/01/18 00:34:27] root INFO:             transform_ops : 
[2022/01/18 00:34:27] root INFO:                 DecodeImage : 
[2022/01/18 00:34:27] root INFO:                     channel_first : False
[2022/01/18 00:34:27] root INFO:                     to_rgb : True
[2022/01/18 00:34:27] root INFO:                 RandCropImage : 
[2022/01/18 00:34:27] root INFO:                     size : 224
[2022/01/18 00:34:27] root INFO:                 RandFlipImage : 
[2022/01/18 00:34:27] root INFO:                     flip_code : 1
[2022/01/18 00:34:27] root INFO:                 NormalizeImage : 
[2022/01/18 00:34:27] root INFO:                     mean : [0.5, 0.5, 0.5]
[2022/01/18 00:34:27] root INFO:                     order : 
[2022/01/18 00:34:27] root INFO:                     scale : 1.0/255.0
[2022/01/18 00:34:27] root INFO:                     std : [0.5, 0.5, 0.5]
[2022/01/18 00:34:27] root INFO:         loader : 
[2022/01/18 00:34:27] root INFO:             num_workers : 4
[2022/01/18 00:34:27] root INFO:             use_shared_memory : True
[2022/01/18 00:34:27] root INFO:         sampler : 
[2022/01/18 00:34:27] root INFO:             batch_size : 160
[2022/01/18 00:34:27] root INFO:             drop_last : False
[2022/01/18 00:34:27] root INFO:             name : DistributedBatchSampler
[2022/01/18 00:34:27] root INFO:             shuffle : True
[2022/01/18 00:34:27] root INFO: Global : 
[2022/01/18 00:34:27] root INFO:     checkpoints : None
[2022/01/18 00:34:27] root INFO:     device : gpu
[2022/01/18 00:34:27] root INFO:     epochs : 120
[2022/01/18 00:34:27] root INFO:     eval_during_train : True
[2022/01/18 00:34:27] root INFO:     eval_interval : 1
[2022/01/18 00:34:27] root INFO:     image_shape : [3, 224, 224]
[2022/01/18 00:34:27] root INFO:     output_dir : ./output/
[2022/01/18 00:34:27] root INFO:     pretrained_model : ./output/ViT_base_patch16_224/best_model
[2022/01/18 00:34:27] root INFO:     print_batch_step : 10
[2022/01/18 00:34:27] root INFO:     save_inference_dir : ./inference
[2022/01/18 00:34:27] root INFO:     save_interval : 1
[2022/01/18 00:34:27] root INFO:     use_visualdl : False
[2022/01/18 00:34:27] root INFO: Infer : 
[2022/01/18 00:34:27] root INFO:     PostProcess : 
[2022/01/18 00:34:27] root INFO:         class_id_map_file : ppcls/utils/imagenet1k_label_list.txt
[2022/01/18 00:34:27] root INFO:         name : Topk
[2022/01/18 00:34:27] root INFO:         topk : 5
[2022/01/18 00:34:27] root INFO:     batch_size : 10
[2022/01/18 00:34:27] root INFO:     infer_imgs : docs/images/whl/demo.jpg
[2022/01/18 00:34:27] root INFO:     transforms : 
[2022/01/18 00:34:27] root INFO:         DecodeImage : 
[2022/01/18 00:34:27] root INFO:             channel_first : False
[2022/01/18 00:34:27] root INFO:             to_rgb : True
[2022/01/18 00:34:27] root INFO:         ResizeImage : 
[2022/01/18 00:34:27] root INFO:             resize_short : 256
[2022/01/18 00:34:27] root INFO:         CropImage : 
[2022/01/18 00:34:27] root INFO:             size : 224
[2022/01/18 00:34:27] root INFO:         NormalizeImage : 
[2022/01/18 00:34:27] root INFO:             mean : [0.5, 0.5, 0.5]
[2022/01/18 00:34:27] root INFO:             order : 
[2022/01/18 00:34:27] root INFO:             scale : 1.0/255.0
[2022/01/18 00:34:27] root INFO:             std : [0.5, 0.5, 0.5]
[2022/01/18 00:34:27] root INFO:         ToCHWImage : None
[2022/01/18 00:34:27] root INFO: Loss : 
[2022/01/18 00:34:27] root INFO:     Eval : 
[2022/01/18 00:34:27] root INFO:         CELoss : 
[2022/01/18 00:34:27] root INFO:             weight : 1.0
[2022/01/18 00:34:27] root INFO:     Train : 
[2022/01/18 00:34:27] root INFO:         CELoss : 
[2022/01/18 00:34:27] root INFO:             weight : 1.0
[2022/01/18 00:34:27] root INFO: Metric : 
[2022/01/18 00:34:27] root INFO:     Eval : 
[2022/01/18 00:34:27] root INFO:         TopkAcc : 
[2022/01/18 00:34:27] root INFO:             topk : [1, 2]
[2022/01/18 00:34:27] root INFO:     Train : 
[2022/01/18 00:34:27] root INFO:         TopkAcc : 
[2022/01/18 00:34:27] root INFO:             topk : [1, 2]
[2022/01/18 00:34:27] root INFO: Optimizer : 
[2022/01/18 00:34:27] root INFO:     lr : 
[2022/01/18 00:34:27] root INFO:         decay_epochs : [10, 22, 30]
[2022/01/18 00:34:27] root INFO:         learning_rate : 0.01
[2022/01/18 00:34:27] root INFO:         name : Piecewise
[2022/01/18 00:34:27] root INFO:         values : [0.01, 0.001, 0.0001, 1e-05]
[2022/01/18 00:34:27] root INFO:     momentum : 0.9
[2022/01/18 00:34:27] root INFO:     name : Momentum
[2022/01/18 00:34:27] root INFO:     regularizer : 
[2022/01/18 00:34:27] root INFO:         coeff : 0.0001
[2022/01/18 00:34:27] root INFO:         name : L2
[2022/01/18 00:34:27] root INFO: train with paddle 2.2.1 and device CUDAPlace(0)
W0118 00:34:27.483398  3428 device_context.cc:447] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 10.1, Runtime API Version: 10.1
W0118 00:34:27.488457  3428 device_context.cc:465] 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
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmpg1ueymey.py:8
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmpg1ueymey.py:11
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmp64fnbu8d.py:8
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmp64fnbu8d.py:11
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmp_h9tn_a7.py:8
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmp_h9tn_a7.py:11
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmpjl9_2_tk.py:8
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmpjl9_2_tk.py:11
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmpql2k50z9.py:8
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmpql2k50z9.py:11
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmpknifbg79.py:8
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmpknifbg79.py:11
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmpiv1rr45m.py:8
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmpiv1rr45m.py:11
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmpa99vj9b7.py:8
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmpa99vj9b7.py:11
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmp5tt116i3.py:8
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmp5tt116i3.py:11
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmpo6kgrf5v.py:8
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmpo6kgrf5v.py:11
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmp919bum5s.py:8
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmp919bum5s.py:11
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmpbiilsbhv.py:8
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/fluid/layers/math_op_patch.py:341: UserWarning: /tmp/tmpbiilsbhv.py:11
The behavior of expression A + B has been unified with elementwise_add(X, Y, axis=-1) from Paddle 2.0. If your code works well in the older versions but crashes in this version, try to use elementwise_add(X, Y, axis=0) instead of A + B. This transitional warning will be dropped in the future.
  op_type, op_type, EXPRESSION_MAP[method_name]))

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。

def main(config):
    cls_predictor = ClsPredictor(config)
    image_list = get_image_list(config["Global"]["infer_imgs"])

    batch_imgs = []
    batch_names = []
    cnt = 0

    # 保存到文件
    f=open('/home/aistudio/result.txt', 'w')

    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"]
                f.write("{} {}\n".format(filename, clas_ids[0]))
                print("{}:\tclass id(s): {}, score(s): {}, label_name(s): {}".
                      format(filename, clas_ids, scores_str, label_names))
            batch_imgs = []
            batch_names = []
 

    if cls_predictor.benchmark:
        cls_predictor.auto_logger.report()
    return

# 覆盖预测文件
!cp -f ~/predict_cls.py ~/deploy/python/predict_cls.py

# 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_images -o Global.inference_model_dir=../inference/  -o PostProcess.Topk.class_id_map_file=None

%cd ~
!mv result.txt result_period.txt

/home/aistudio
mv: cannot stat 'result.txt': No such file or directory

天气模型导出

# 模型导出
%cd ~/PaddleClas/
!python tools/export_model.py -c ./ppcls/configs/ImageNet/VisionTransformer/ViT_base_patch16_224_weather.yaml -o Global.pretrained_model=./output_weather/ViT_base_patch16_224/best_model

天气预测

# 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_images -o Global.inference_model_dir=../inference_weather/  -o PostProcess.Topk.class_id_map_file=None

%cd ~
!mv result.txt result_weather.txt

/home/aistudio

合并并提交

预测结果合并

period_list = { 0:'Dawn', 1:'Dusk', 2:'Morning', 3:'Afternoon'}
weather_list =  {0:'Cloudy', 1:'Rainy', 2:'Sunny'}

import pandas as pd
import json
data_period= pd.read_csv('result_period.txt', header=None, sep=' ')
data_weather= pd.read_csv('result_weather.txt', header=None, sep=' ')
annotations_list=[]
for i in range(len(data_period)):
    temp={}
    temp["filename"]="test_images"+"\\"+data_weather.loc[i][0]
    temp["period"]=period_list[data_period.loc[i][1]]
    temp["weather"]=weather_list[data_weather.loc[i][1]]
    annotations_list.append(temp)
myresult={}
myresult["annotations"]=annotations_list

with open('result.json','w') as f:
    json.dump(myresult, f)
    print("结果生成完毕")

结果生成完毕

提交并获取成绩 下载result.json并提交，即可获得成绩

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
    fi
done

关注公众号【学姐带你玩AI】

回复“天气”领取完整代码