Kaggle实战-迪拜卫星航拍多类别分类

俺有几个问题写在前面

一、配置文件是哪个

pspnet_r50-d8_4xb2-40k_DubaiDataset.py 2、pspnet-DubaiDataset_20230612.py

二、这里说载入config配置文件pspnet_r50-d8_4xb2-40k_DubaiDataset.py再修改config配置文件，最后又说保存config配置文件到pspnet-DubaiDataset_20230612.py，那么我修改的是pspnet_r50-d8_4xb2-40k_DubaiDataset.py还是pspnet-DubaiDataset_20230612.py呢

三、运行语句应该咋写 哪个才是配置文件（救命

 

 

1. 视频链接：MMSegmentation代码课
   教程链接：MMSegmentation_Tutorials/20230612
2. 数据集：

Dubai-dataset

结构： ann_dir 

               ——train

              |

               ——val

      img_dir

               ——train

             |

              ——val

1. 数据可视化：

数据可视化

import numpy as np

import matplotlib.pyplot as plt

from PIL import Image

# Load the image and annotation

image_path = r'C:\Users\dell\Desktop\Dubai-dataset\img_dir\train\14.jpg'

annotation_path = r'C:\Users\dell\Desktop\Dubai-dataset\ann_dir\train\14.png'

image = np.array(Image.open(image_path))

annotation = np.array(Image.open(annotation_path))

# Visualize the image and annotation

fig, ax = plt.subplots(1, 2, figsize=(10, 5))

ax[0].imshow(image)

ax[0].set_title("Image")

ax[1].imshow(annotation)

ax[1].set_title("Annotation")

plt.show()

在原图上叠加显示：

import numpy as np

import matplotlib.pyplot as plt

from PIL import Image

# Load the image and annotation

image_path = r'C:\Users\dell\Desktop\Dubai-dataset\img_dir\train\14.jpg'

annotation_path = r'C:\Users\dell\Desktop\Dubai-dataset\ann_dir\train\14.png'

image = np.array(Image.open(image_path))

annotation = np.array(Image.open(annotation_path))

# Create a colormap for visualization

cmap = plt.get_cmap("tab20")

norm = plt.Normalize(vmin=0, vmax=20)

# Convert the annotation to a color image

annotation_color = norm(annotation)

annotation_color = cmap(annotation_color)

annotation_color = (annotation_color[:, :, :3] * 255).astype(np.uint8)

# Merge the image and annotation

merged = cv2.addWeighted(image, 0.7, annotation_color, 0.3, 0)

# Display the merged image

plt.imshow(merged)

plt.show()

批量可视化：

import os

import numpy as np

import matplotlib.pyplot as plt

from PIL import Image

# Set the directory paths                                           

image_dir = r'C:\Users\dell\Desktop\Dubai-dataset\img_dir\train'

annotation_dir = r'C:\Users\dell\Desktop\Dubai-dataset\ann_dir\train'

# Get the file names

image_files = os.listdir(image_dir)

annotation_files = os.listdir(annotation_dir)

# Sort the file names

image_files.sort()

annotation_files.sort()

# Create a colormap for visualization

cmap = plt.get_cmap("tab20")

norm = plt.Normalize(vmin=0, vmax=20)

# Loop through the files

for i, (image_file, annotation_file) in enumerate(zip(image_files, annotation_files)):

    # Load the image and annotation

    image_path = os.path.join(image_dir, image_file)

    annotation_path = os.path.join(annotation_dir, annotation_file)

    image = np.array(Image.open(image_path))

    annotation = np.array(Image.open(annotation_path))

    # Convert the annotation to a color image

    annotation_color = norm(annotation)

    annotation_color = cmap(annotation_color)

    annotation_color = (annotation_color[:, :, :3] * 255).astype(np.uint8)

    # Merge the image and annotation

    merged = cv2.addWeighted(image, 0.7, annotation_color, 0.3, 0)

    # Display the merged image

    plt.imshow(merged)

    plt.title(f"Image {i+1}")

plt.show()

显示标签名称和种类

import numpy as np

import matplotlib.pyplot as plt

from PIL import Image

# Load the annotation image

annotation_path = r'C:\Users\dell\Desktop\Dubai-dataset\ann_dir\train\14.png'

annotation = np.array(Image.open(annotation_path))

# Determine the number of unique labels in the annotation

unique_labels = np.unique(annotation)

# Convert the unique labels to a list of integers

label_values = unique_labels.astype(int).tolist()

# Load the label names from a dictionary

label_names = {

    0: 'background',

    1: 'sky',

    2: 'building',

    3: 'vegetation',

    4: 'sidewalk',

    5: 'car',

    6: 'pedestrian',

    7: 'fence',

    8: 'pole',

    9: 'traffic light',

    10: 'traffic sign'

}

# Create a list of label names for the unique labels

label_names_list = [label_names[label_value] for label_value in label_values]

# Print the number of unique labels and their names

print(f"The annotation contains {len(unique_labels)} unique labels:")

for label_value, label_name in zip(label_values, label_names_list):

    print(f"Label {label_value}: {label_name}")

# Select a label to display (e.g., the first unique label)

label_to_display = unique_labels[0]

# Create a binary mask for the selected label

mask = np.zeros_like(annotation)

mask[annotation == label_to_display] = 1

# Apply the binary mask to the original image

image_path = r'C:\Users\dell\Desktop\Dubai-dataset\img_dir\train\14.jpg'

image = np.array(Image.open(image_path))

masked_image = np.zeros_like(image)

for i in range(3):

    masked_image[:, :, i] = image[:, :, i] * mask

# Display the masked image

plt.imshow(masked_image)

plt.title(f"Label {label_to_display}: {label_names[label_to_display]}")

plt.show()

十全大补丸

import os

import cv2

import numpy as np

import matplotlib.pyplot as plt

from PIL import Image

# Set the directory paths

# 设置目录路径 和下面是否批量处理图像要对应

image_dir = r'C:\Users\dell\Desktop\Dubai-dataset\img_dir\train'

annotation_dir = r'C:\Users\dell\Desktop\Dubai-dataset\ann_dir\train'

# Load the label names from a dictionary 

# 从字典中加载标签名称

label_names = {

    0: 'background',

    1: 'sky',

    2: 'building',

    3: 'vegetation',

    4: 'sidewalk',

    5: 'car',

    6: 'pedestrian',

    7: 'fence',

    8: 'pole',

    9: 'traffic light',

    10: 'traffic sign'

}

# Set the label to display 

# 设置要显示的标签想显示单个标签直接写数字，想显示多个标签[2,3]

label_to_display = 2

# Set whether to process images in batch 

# 设置是否批量处理图像 整个文件夹True，单个文件False

process_in_batch = True

# Process images

if process_in_batch:

    # Get the file names

    image_files = os.listdir(image_dir)

    annotation_files = os.listdir(annotation_dir)

    # Sort the file names

    image_files.sort()

    annotation_files.sort()

    # Create a colormap for visualization

    cmap = plt.get_cmap("tab20")

    norm = plt.Normalize(vmin=0, vmax=20)

    # Loop through the files

    for i, (image_file, annotation_file) in enumerate(zip(image_files, annotation_files)):

        # Load the image and annotation

        image_path = os.path.join(image_dir, image_file)

        annotation_path = os.path.join(annotation_dir, annotation_file)

        image = np.array(Image.open(image_path))

        annotation = np.array(Image.open(annotation_path))

        # Determine the number of unique labels in the annotation

        unique_labels = np.unique(annotation)

        # Convert the unique labels to a list of integers

        label_values = unique_labels.astype(int).tolist()

        # Create a list of label names for the unique labels

        label_names_list = [label_names[label_value] for label_value in label_values]

        # Print the number of unique labels and their names

        print(f"Image {i+1} contains {len(unique_labels)} unique labels:")

        for label_value, label_name in zip(label_values, label_names_list):

            print(f"Label {label_value}: {label_name}")

        # Create a binary mask for the selected label

        mask = np.zeros_like(annotation)

        mask[annotation == label_to_display] = 1

        # Convert the annotation to a color image

        annotation_color = norm(annotation)

        annotation_color = cmap(annotation_color)

        annotation_color = (annotation_color[:, :, :3] * 255).astype(np.uint8)

        # Apply the binary mask to the color annotation

        masked_annotation = np.zeros_like(annotation_color)

        for i in range(3):

            masked_annotation[:, :, i] = annotation_color[:, :, i] * mask

        # Merge the image and masked annotation#

#合并图像和屏蔽注释 权重

        merged = cv2.addWeighted(image, 0.7, masked_annotation, 0.3, 0)

        # Display the merged image

        plt.imshow(merged)

        plt.title(f"Image {i+1}")

        plt.show()

else:

    # Select an image to process

    image_path = r'C:\Users\dell\Desktop\Dubai-dataset\img_dir\train\14.jpg'

    annotation_path = r'C:\Users\dell\Desktop\Dubai-dataset\ann_dir\train\14.png'

    # Load the image and annotation

    image = np.array(Image.open(image_path))

    annotation = np.array(Image.open(annotation_path))

    # Determine the number of unique labels in the annotation

    unique_labels = np.unique(annotation)

    # Convert the unique labels to a list of integers

    label_values = unique_labels.astype(int).tolist()

    # Create a list of label names for the unique labels

    label_names_list = [label_names[label_value] for label_value in label_values]

    # Print the number of unique labels and their names

    print(f"The annotation contains {len(unique_labels)} unique labels:")

    for label_value, label_name in zip(label_values, label_names_list):

        print(f"Label {label_value}: {label_name}")

    # Create a binary mask for the selected label

    mask = np.zeros_like(annotation)

    mask[annotation == label_to_display] = 1

    # Convert the annotation to a color image

    cmap = plt.get_cmap("tab20")

    norm = plt.Normalize(vmin=0, vmax=20)

    annotation_color = norm(annotation)

    annotation_color = cmap(annotation_color)

    annotation_color = (annotation_color[:, :, :3] * 255).astype(np.uint8)

    # Apply the binary mask to the color annotation

    masked_annotation = np.zeros_like(annotation_color)

    for i in range(3):

        masked_annotation[:, :, i] = annotation_color[:, :, i] * mask

    # Merge the image and masked annotation

    merged = cv2.addWeighted(image, 0.7, masked_annotation, 0.3, 0)

    # Display the merged image

    plt.imshow(merged)

    plt.title(f"Label {label_to_display}: {label_names[label_to_display]}")

    plt.show()

1. 配置

素材位置：

/zhangxl/models_main/mmsegmentation-main/mmsegmentation-main/data                     Dubai-dataset

配置：

（1）定义数据集类：

/zhangxl/models_main/mmsegmentation-main/mmsegmentation-main/mmseg/datasets

DubaiDataset.py

代码

# 同济子豪兄 2023-2-15

from mmseg.registry import DATASETS

from .basesegdataset import BaseSegDataset

@DATASETS.register_module()

class DubaiDataset(BaseSegDataset):

    # 类别和对应的可视化配色

    METAINFO = {

        'classes':['Land', 'Road', 'Building', 'Vegetation', 'Water', 'Unlabeled'],

        'palette':[[132,41,246], [228,193,110], [152,16,60], [58,221,254], [41,169,226], [155,155,155]]

    }

    

    # 指定图像扩展名、标注扩展名

    def __init__(self,

                 img_suffix='.jpg',

                 seg_map_suffix='.png',

                 reduce_zero_label=False, # 类别ID为0的类别是否需要除去

                 **kwargs) -> None:

        super().__init__(

            img_suffix=img_suffix,

            seg_map_suffix=seg_map_suffix,

            reduce_zero_label=reduce_zero_label,

            **kwargs)

（2）注册数据集

/zhangxl/models_main/mmsegmentation-main/mmsegmentation-main/mmseg/datasets

__init__.py

改动

from .DubaiDataset import DubaiDataset

 'DubaiDataset'

（3）定义训练及测试pipeline

/zhangxl/models_main/mmsegmentation-main/mmsegmentation-main/configs/_base_/datasets

 DubaiDataset_pipeline.py

代码

# dataset settings

#3.4行需要改动一下

dataset_type = 'DubaiDataset' # 数据集类名

data_root = 'Dubai-dataset/' # 数据集路径（相对于mmsegmentation主目录）

crop_size = (64, 64)

train_pipeline = [

    dict(type='LoadImageFromFile'),

    dict(type='LoadAnnotations'),

    dict(

        type='RandomResize',

        scale=(2048, 1024),

        ratio_range=(0.5, 2.0),

        keep_ratio=True),

    dict(type='RandomCrop', crop_size=crop_size, cat_max_ratio=0.75),

    dict(type='RandomFlip', prob=0.5),

    dict(type='PhotoMetricDistortion'),

    dict(type='PackSegInputs')

]

test_pipeline = [

    dict(type='LoadImageFromFile'),

    dict(type='Resize', scale=(2048, 1024), keep_ratio=True),

    # add loading annotation after ``Resize`` because ground truth

    # does not need to do resize data transform

    dict(type='LoadAnnotations'),

    dict(type='PackSegInputs')

]

img_ratios = [0.5, 0.75, 1.0, 1.25, 1.5, 1.75]

tta_pipeline = [

    dict(type='LoadImageFromFile', file_client_args=dict(backend='disk')),

    dict(

        type='TestTimeAug',

        transforms=[

            [

                dict(type='Resize', scale_factor=r, keep_ratio=True)

                for r in img_ratios

            ],

            [

                dict(type='RandomFlip', prob=0., direction='horizontal'),

                dict(type='RandomFlip', prob=1., direction='horizontal')

            ], [dict(type='LoadAnnotations')], [dict(type='PackSegInputs')]

        ])

]

train_dataloader = dict(

    batch_size=2,

    num_workers=2,

    persistent_workers=True,

    sampler=dict(type='InfiniteSampler', shuffle=True),

    dataset=dict(

        type=dataset_type,

        data_root=data_root,

        data_prefix=dict(

            img_path='img_dir/train', seg_map_path='ann_dir/train'),

        pipeline=train_pipeline))

val_dataloader = dict(

    batch_size=1,

    num_workers=4,

    persistent_workers=True,

    sampler=dict(type='DefaultSampler', shuffle=False),

    dataset=dict(

        type=dataset_type,

        data_root=data_root,

        data_prefix=dict(

            img_path='img_dir/val', seg_map_path='ann_dir/val'),

        pipeline=test_pipeline))

test_dataloader = val_dataloader

val_evaluator = dict(type='IoUMetric', iou_metrics=['mIoU'])

test_evaluator = val_evaluator

1. 配置文件： /zhangxl/models_main/mmsegmentation-main/mmsegmentation-main/configs/pspnet      pspnet_r50-d8_4xb2-40k_DubaiDataset.py

代码

_base_ = [

    '../_base_/models/pspnet_r50-d8.py', '../_base_/datasets/DubaiDataset_pipeline.py',

    '../_base_/default_runtime.py', '../_base_/schedules/schedule_40k.py'

]

crop_size = (64, 64) # 输入图像尺寸，根据自己数据集情况修改

data_preprocessor = dict(size=crop_size)

model = dict(data_preprocessor=data_preprocessor)

1. 修改配置文件

/zhangxl/models_main/mmsegmentation-main/mmsegmentation-main

pspnet-DubaiDataset_20230612.py

代码

# 载入config配置文件

from mmengine import Config

cfg = Config.fromfile('./configs/pspnet/pspnet_r50-d8_4xb2-40k_DubaiDataset.py')

# 根据个人数据集和分割任务，修改config配置文件

cfg.norm_cfg = dict(type='BN', requires_grad=True) # 只使用GPU时，BN取代SyncBN

cfg.crop_size = (256, 256)

cfg.model.data_preprocessor.size = cfg.crop_size

cfg.model.backbone.norm_cfg = cfg.norm_cfg

cfg.model.decode_head.norm_cfg = cfg.norm_cfg

cfg.model.auxiliary_head.norm_cfg = cfg.norm_cfg

# modify num classes of the model in decode/auxiliary head

# 模型 decode/auxiliary 输出头，指定为类别个数

cfg.model.decode_head.num_classes = 6

cfg.model.auxiliary_head.num_classes = 6

cfg.train_dataloader.batch_size = 8

cfg.test_dataloader = cfg.val_dataloader

# 结果保存目录

cfg.work_dir = './work_dirs/DubaiDataset'

# 训练迭代次数

cfg.train_cfg.max_iters = 3000

# 评估模型间隔

cfg.train_cfg.val_interval = 400

# 日志记录间隔

cfg.default_hooks.logger.interval = 100

# 模型权重保存间隔

cfg.default_hooks.checkpoint.interval = 1500

# 随机数种子

cfg['randomness'] = dict(seed=0)

# 查看完整config配置文件

print(cfg.pretty_text)

# 保存config配置文件

cfg.dump('pspnet-DubaiDataset_20230612.py')