MMSegmentation跑自己的数据进行二分类_mmseg如何处理二分类

官方文档：https://mmsegmentation.readthedocs.io/
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。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=10,
            classes = ('background','building'),
            palette = [[0,0,0],[255,255,255]],
            **kwargs)
        assert osp.exists(self.img_dir)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20

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]
1
2
3
4
5
6
7

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]]
1
2
3
4
5
6
7
8
9
10

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',
]

1
2
3
4
5
6
7
8
9
10
11

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))
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54

5、我使用的模型是mmsegmentation\configs\segformer\segformer_mit-b0_512x512_160k_ade20k.py
在这个文件夹中设置mydata.py、segformer_mit-b0.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。我去掉了辅助分类器。

_base_ = [
    '../_base_/models/segformer_mit-b0.py', '../_base_/datasets/mydata.py',
    '../_base_/default_runtime.py', '../_base_/schedules/schedule_160k.py'
]

model = dict(
    pretrained='/home/vge/Documents/lcw/mmsegmentation/checkpoints/segformer_mit_b0_convert.pth', decode_head=dict(num_classes=2))

# optimizer
optimizer = dict(
    _delete_=True,
    type='AdamW',
    lr=0.00006,
    betas=(0.9, 0.999),
    weight_decay=0.01,
    paramwise_cfg=dict(
        custom_keys={
            'pos_block': dict(decay_mult=0.),
            'norm': dict(decay_mult=0.),
            'head': dict(lr_mult=10.)
        }))

lr_config = dict(
    _delete_=True,
    policy='poly',
    warmup='linear',
    warmup_iters=1500,
    warmup_ratio=1e-6,
    power=1.0,
    min_lr=0.0,
    by_epoch=False)

data = dict(samples_per_gpu=16, workers_per_gpu=12) # 这里在设置每个GPU的batch_size和读取数据的线程数
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33

6、一个需要注意的地方，将官方预训练模型转换到 MMSegmentation。
利用mmsegmentation\tools\model_converters\mit2mmseg.py对模型进行转换，因为mmsegmentation中Segformer代码的一些参数名字和segformer_mit_b0预训练模型参数名字不一样。因此需要利用mit2mmseg.py代码将预训练模型中的参数名字进行一些修改。

    parser.add_argument('--src',default='XXX/mmsegmentation/checkpoints/segformer_mit_b0.pth', help='src model path or url')
    # The dst path must be a full path of the new checkpoint.
    parser.add_argument('--dst',default='XXX/mmsegmentation/checkpoints/segformer_mit_b0_convert.pth', help='save path')
1
2
3

7、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)')
1
2
3
4
5
6

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/segformer/segformer_mit-b0_512x512_160k_ade20k.py', help='train config file path')  
1

8、一个很坑的地方
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,所以我修改了源代码。
mmsegmentation\mmseg\datasets\pipelines\loading.py143行代码

gt_semantic_seg[gt_semantic_seg_copy == old_id*255] = new_id
1

9、遇到的评价指标为0或者100%的问题
修改LOSS函数应该可以解决此问题。在mmsegmentation\configs_base_\models\upernet_swin.py中我将主分类器LOSS函数都由CrossEntropyLoss修改为DiceLoss。评价指标就正常了。我没用辅助分类器。

# model settings
norm_cfg = dict(type='SyncBN', requires_grad=True)
backbone_norm_cfg = dict(type='LN', requires_grad=True)
model = dict(
    type='EncoderDecoder',
    pretrained=None,
    backbone=dict(
        type='SwinTransformer',
        pretrain_img_size=224,
        embed_dims=96,
        patch_size=4,
        window_size=7,
        mlp_ratio=4,
        depths=[2, 2, 6, 2],
        num_heads=[3, 6, 12, 24],
        strides=(4, 2, 2, 2),
        out_indices=(0, 1, 2, 3),
        qkv_bias=True,
        qk_scale=None,
        patch_norm=True,
        drop_rate=0.,
        attn_drop_rate=0.,
        drop_path_rate=0.3,
        use_abs_pos_embed=False,
        act_cfg=dict(type='GELU'),
        norm_cfg=backbone_norm_cfg),
    decode_head=dict(
        type='UPerHead',
        in_channels=[96, 192, 384, 768],
        in_index=[0, 1, 2, 3],
        pool_scales=(1, 2, 3, 6),
        channels=512,
        dropout_ratio=0.1,
        num_classes=2,
        norm_cfg=norm_cfg,
        align_corners=False,
        loss_decode=dict(
        type='DiceLoss', loss_weight=1.0)),
    train_cfg=dict(),
    test_cfg=dict(mode='whole'))

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41

10、输出mFscore的同时输出mIoU，在mmseg/core/evaluation/metrics.py第360-380行代码

    for metric in metrics:
        if metric == 'mIoU':
            iou = total_area_intersect / total_area_union
            acc = total_area_intersect / total_area_label
            ret_metrics['IoU'] = iou
            ret_metrics['Acc'] = acc
        elif metric == 'mDice':
            dice = 2 * total_area_intersect / (
                total_area_pred_label + total_area_label)
            acc = total_area_intersect / total_area_label
            ret_metrics['Dice'] = dice
            ret_metrics['Acc'] = acc
        elif metric == 'mFscore':
            precision = total_area_intersect / total_area_pred_label
            recall = total_area_intersect / total_area_label
            f_value = torch.tensor(
                [f_score(x[0], x[1], beta) for x in zip(precision, recall)])
            ret_metrics['Fscore'] = f_value
            ret_metrics['Precision'] = precision
            ret_metrics['Recall'] = recall
            iou = total_area_intersect / total_area_union
            ret_metrics['IoU'] = iou
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

参考：mmsegmentation教程1:自定义数据集、config文件修改、训练教程