MMSegmentation跑自己的数据进行二分类

我的数据是RGB影像为.tif，标签影像为单通道（0-255）的.png影像，背景像素值为0，建筑物像素值为255
制作自己的数据集进行二分类步骤如下：
1、在mmsegmentation\mmseg\datasets文件夹下创建mydataset.py
注意：将reduce_zero_label=False，ignore_index随便设置一个数不在0-255即可。这两个变量的作用参看源代码mmsegmentation\mmseg\datasets\custom.py。个人理解reduce_zero_label为Ture就是减少0像素在标签中的数量，因为我们的0为背景，所有设置为False，ignore_index应该指的是忽略像素值为ignore_index所表示的类别

# Copyright (c) OpenMMLab. All rights reserved.
import os.path as osp
from .builder import DATASETS
from .custom import CustomDataset

@DATASETS.register_module()
class MydataDataset(CustomDataset):
    CLASSES = ('background','building')
    PALETTE = [[0,0,0],[255,255,255]]

    def __init__(self, **kwargs):
        super(MydataDataset, self).__init__(
            img_suffix='.tif',
            seg_map_suffix='.png',
            reduce_zero_label=False,
            ignore_index=300,
            classes = ('background','building'),
            palette = [[0,0,0],[255,255,255]],
            **kwargs)
        assert osp.exists(self.img_dir)

mmsegmentation\mmseg\core\evaluation\metrics.py中intersect_and_union函数————reduce_zero_label对于评价指标的影响

    if reduce_zero_label:
        label[label == 0] = 255
        label = label - 1
        label[label == 254] = 255
    mask = (label != ignore_index)
    pred_label = pred_label[mask]
    label = label[mask]

2、在mmsegmentation\mmseg\core\evaluation\class_names.py文件后面添加代码

def mydata_classes():
    """shengteng class names for external use."""
    return [
         'background','building'
    ]


def mydata_palette():
    return [[0,0,0][255,255,255]]

3、在mmseg/datasets/init.py文件中添加自己的数据类别
注意：from .mydataset import MydataDataset一定要添加

from .mydataset import MydataDataset
__all__ = [
    'CustomDataset', 'build_dataloader', 'ConcatDataset', 'RepeatDataset',
    'DATASETS', 'build_dataset', 'PIPELINES', 'CityscapesDataset',
    'PascalVOCDataset', 'ADE20KDataset', 'PascalContextDataset',
    'PascalContextDataset59', 'ChaseDB1Dataset', 'DRIVEDataset', 'HRFDataset',
    'STAREDataset', 'DarkZurichDataset', 'NightDrivingDataset',
    'COCOStuffDataset', 'LoveDADataset', 'MultiImageMixDataset',
    'iSAIDDataset', 'ISPRSDataset', 'PotsdamDataset', 'MydataDataset',
]

4、在mmsegmentation\configs_base_\datasets中创建mydata.py，设置路径、尺寸等相关参数

# dataset settings
dataset_type = 'MydataDataset'
data_root = '/media/vge/DataA/lcw/Sanyuan_Buildings'
img_norm_cfg = dict(
    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
crop_size = (512, 512)
train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', reduce_zero_label=False),
    dict(type='Resize', img_scale=(512, 512), ratio_range=(0.5, 2.0)),
    dict(type='RandomCrop', crop_size=crop_size, cat_max_ratio=0.75),
    dict(type='RandomFlip', prob=0.5),
    dict(type='PhotoMetricDistortion'),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='Pad', size=crop_size, pad_val=0, seg_pad_val=255),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_semantic_seg']),
]
test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale=(512, 512),
        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
        flip=False,
        transforms=[
            dict(type='Resize', keep_ratio=True),
            dict(type='RandomFlip'),
            dict(type='Normalize', **img_norm_cfg),
            dict(type='ImageToTensor', keys=['img']),
            dict(type='Collect', keys=['img']),
        ])
]
data = dict(
    samples_per_gpu=4, #设置每个GPU的batch_size
    workers_per_gpu=6, #设置每个GPU对应读取线程数
    train=dict(
        type=dataset_type,
        data_root=data_root,
        img_dir='train/image',
        ann_dir='train/label',
        pipeline=train_pipeline),
    val=dict(
        type=dataset_type,
        data_root=data_root,
        img_dir='val/image',
        ann_dir='val/label',
        pipeline=test_pipeline),
    test=dict(
        type=dataset_type,
        data_root=data_root,
        img_dir='test/image',
        ann_dir='test/label',
        pipeline=test_pipeline))

5、我使用的模型是mmsegmentation\configs\swin\upernet_swin_base_patch4_window7_512x512_160k_ade20k_pretrain_224x224_1K.py
在这个文件夹中设置mydata.py、upernet_swin.py、default_runtime.py、schedule_80k.py的路径，以及预训练模型参数.pth文件的路径位置。这里我将num_classes=2设置类别数为2。

_base_ = [
    '../_base_/models/upernet_swin.py', '../_base_/datasets/mydata.py',
    '../_base_/default_runtime.py', '../_base_/schedules/schedule_80k.py'
]
checkpoint_file = '/home/vge/Documents/lcw/mmsegmentation/checkpoints/swin_base_patch4_window7_224_20220317-e9b98025.pth'
#checkpoint_file = 'https://download.openmmlab.com/mmsegmentation/v0.5/pretrain/swin/swin_base_patch4_window7_224_20220317-e9b98025.pth'  # noqa
model = dict(
    backbone=dict(
        init_cfg=dict(type='Pretrained', checkpoint=checkpoint_file),
        embed_dims=128,
        depths=[2, 2, 18, 2],
        num_heads=[4, 8, 16, 32]),
    decode_head=dict(in_channels=[128, 256, 512, 1024], num_classes=2),
    auxiliary_head=dict(in_channels=512, num_classes=2)
)

6、其他相关设置
评价指标mIoU、mFscore在mmsegmentation\configs_base_\schedules文件夹中进行设置
默认GPU id，设置mmsegmentation\tools\train.py中47行代码

    group_gpus.add_argument(
        '--gpu-id',
        type=int,
        default=2,
        help='id of gpu to use '
        '(only applicable to non-distributed training)')

为了在pycharm中进行调试，设置mmsegmentation\tools\train.py中的
的第23行代码

 parser.add_argument('--config', default='../configs/swin/upernet_swin_base_patch4_window7_512x512_160k_ade20k_pretrain_224x224_1K.py', help='train config file path')  

7、一个行坑的地方
由于我的数据是8位单通道的数据，所有做了上述步骤以后还是报错
经过调试发现读取label数据的时候，对于单通道数据，像素值要和类别的id
相等。比如我的数’background’,‘building’在单通道的label数据中要像素值分别要为background’=0，‘building’= 1，因为我的数据是background’=0，‘building’= 255,所有我修改了源代码。
C:\Users\76760\Desktop\mmsegmentation\mmseg\datasets\pipelines\loading.py143行代码

gt_semantic_seg[gt_semantic_seg_copy == old_id*255] = new_id