步骤 1. 创建一个 conda 环境，并激活
conda create --name openmmlab python=3.8 -y
conda activate openmmlab
步骤 2. 安装 PyTorch
安装低版本的pytorch，在继续安装MMSegmentation，才能避免报错。因为显卡是4090，CUDA Toolkit必须安装11.8以上的。（只有本机系统的CUDA是合适的，在环境中的是向下兼容的，个人的理解）CUDA Toolkit
conda安装：conda install pytorch==1.13.0 torchvision==0.14.0 torchaudio==0.13.0 pytorch-cuda=11.7 -c pytorch -c nvidia
pip安装：pip install torch==1.13.0+cu117 torchvision==0.14.0+cu117 torchaudio==0.13.0 --extra-index-url https://download.pytorch.org/whl/cu117
以下的步骤就是为安装MMSegmentation做准备。
最佳实践
步骤 0. 使用 MIM 安装 MMCV
pip install -U openmim
mim install mmengine
mim install "mmcv>=2.0.0"
步骤 1. 安装 MMSegmentation
情况 a: 如果您想立刻开发和运行 mmsegmentation，您可通过源码安装：
git clone -b main https://github.com/open-mmlab/mmsegmentation.git
cd mmsegmentation
pip install -v -e .
'-v' 表示详细模式，更多的输出
'-e' 表示以可编辑模式安装工程，
因此对代码所做的任何修改都生效，无需重新安装
情况 b: 如果您把 mmsegmentation 作为依赖库或者第三方库，可以通过 pip 安装：
pip install "mmsegmentation>=1.0.0"
然后运行测试代码
步骤1.我们需要下载配置和检查点文件。
mim download mmsegmentation --config pspnet_r50-d8_4xb2-40k_cityscapes-512x1024 --dest .
下载过程将需要几秒钟或更长时间，具体取决于您的网络环境。下载完成后，您将在当前文件夹中找到两个文件pspnet_r50-d8_4xb2-40k_cityscapes-512x1024.py和。pspnet_r50-d8_512x1024_40k_cityscapes_20200605_003338-2966598c.pth
步骤2.验证推理演示。
选项（a）。如果从源安装 mmsegmentation，只需运行以下命令。
python demo/image_demo.py demo/demo.png configs/pspnet/pspnet_r50-d8_4xb2-40k_cityscapes-512x1024.py pspnet_r50-d8_512x1024_40k_cityscapes_20200605_003338-2966598c.pth --device cuda:0 --out-file result.jpg
您将在当前文件夹中看到一个新图像result.jpg，其中所有对象都覆盖有分割蒙版。
下面可能会进行的报错，以及我解决的方式。
报错一：
Traceback (most recent call last):
File "demo/image_demo.py", line 6, in
from mmseg.apis import inference_model, init_model, show_result_pyplot
File "/home/ubuntu/mmsegmentation/mmseg/init.py", line 61, in
assert (mmcv_min_version <= mmcv_version < mmcv_max_version),
AssertionError: MMCV==2.2.0 is used but incompatible. Please install mmcv>=2.0.0rc4.、
报错二：
ModuleNotFoundError: No module named 'mmcv._ext'
有两种解决方式：
pip install mmcv-lite==2.0.0rc4
如果还是报错可以使用下面的这个命令
mim install mmcv==2.1.0
在运行测试代码，即可。
在训练的初始阶段，需要更多的评价指标，以下是除了默认的mIoU以外,添加别的指标。
可视化代码：多个模型的可视化输出
from mmseg.apis import inference_model, init_model, show_result_pyplot
import mmcv
import os
import glob
# 定义所有模型配置
MODEL_CONFIGS = [
# {
# 'name': 'danet',
# 'config': '/home/lsl/Workspace/mmsegmentation/work_dirs0/danet_r50-d8_4xb4-40k_voc12aug-512x512/danet_r50-d8_4xb4-40k_voc12aug-512x512.py',
# 'checkpoint': '/home/lsl/Workspace/mmsegmentation/work_dirs0/danet_r50-d8_4xb4-40k_voc12aug-512x512/iter_37000.pth'
# },
# {
# 'name': 'ocrnet',
# 'config': '/home/lsl/Workspace/mmsegmentation/work_dirs/ocrnet_hr18_4xb4-40k_voc12aug-512x512/ocrnet_hr18_4xb4-40k_voc12aug-512x512.py',
# 'checkpoint': '/home/lsl/Workspace/mmsegmentation/work_dirs/ocrnet_hr18_4xb4-40k_voc12aug-512x512/iter_37000.pth'
# },
# {
# 'name': 'deeplabv3_r50',
# 'config': '/home/lsl/Workspace/mmsegmentation/work_dirs/deeplabv3_r50-d8_4xb4-20k_voc12aug-512x512/deeplabv3_r50-d8_4xb4-20k_voc12aug-512x512.py',
# 'checkpoint': '/home/lsl/Workspace/mmsegmentation/work_dirs/deeplabv3_r50-d8_4xb4-20k_voc12aug-512x512/iter_37000.pth'
# },
# {
# 'name': 'deeplabv3plus_r50',
# 'config': '/home/lsl/Workspace/mmsegmentation/work_dirs/deeplabv3plus_r50-d8_4xb4-40k_voc12aug-512x512/deeplabv3plus_r50-d8_4xb4-40k_voc12aug-512x512.py',
# 'checkpoint': '/home/lsl/Workspace/mmsegmentation/work_dirs/deeplabv3plus_r50-d8_4xb4-40k_voc12aug-512x512/iter_37000.pth'
# },
# {
# 'name': 'pspnet_r50',
# 'config': '/home/lsl/Workspace/mmsegmentation/work_dirs/pspnet_r50-d8_4xb4-40k_voc12aug-512x512/pspnet_r50-d8_4xb4-40k_voc12aug-512x512.py',
# 'checkpoint': '/home/lsl/Workspace/mmsegmentation/work_dirs/pspnet_r50-d8_4xb4-40k_voc12aug-512x512/iter_37000.pth'
# },
# {
# 'name': 'segformer',
# 'config': '/home/lsl/Workspace/mmsegmentation/work_dirs/segformer0_mit-b0_8xb2-160k_ade20k-512x512/segformer_mit-b0_8xb2-160k_ade20k-512x512.py',
# 'checkpoint': '/home/lsl/Workspace/mmsegmentation/work_dirs/segformer0_mit-b0_8xb2-160k_ade20k-512x512/iter_37000.pth'
# },
{
'name': 'segmenter',
'config': '/home/lsl/Workspace/mmsegmentation/work_dirs/segmenter_vit-t_mask_8xb1-160k_ade20k-512x512/segmenter_vit-t_mask_8xb1-160k_ade20k-512x512.py',
'checkpoint': '/home/lsl/Workspace/mmsegmentation/work_dirs/segmenter_vit-t_mask_8xb1-160k_ade20k-512x512/iter_37000.pth'
},
# {
# 'name': 'segnext',
# 'config': '/home/lsl/Workspace/mmsegmentation/work_dirs/segnext_mscan-t_1xb16-adamw-160k_ade20k-512x512/segnext_mscan-t_1xb16-adamw-160k_ade20k-512x512.py',
# 'checkpoint': '/home/lsl/Workspace/mmsegmentation/work_dirs/segnext_mscan-t_1xb16-adamw-160k_ade20k-512x512/iter_37000.pth'
# },
{
'name': 'upernet',
'config': '/home/lsl/Workspace/mmsegmentation/work_dirs/upernet_r50_4xb4-40k_voc12aug-512x512/upernet_r50_4xb4-40k_voc12aug-512x512.py',
'checkpoint': '/home/lsl/Workspace/mmsegmentation/work_dirs/upernet_r50_4xb4-40k_voc12aug-512x512/iter_37000.pth'
}
]
def process_images(model_config, input_dir, base_output_dir):
"""
使用指定的模型处理输入目录中的所有图像
"""
# 创建模型特定的输出目录
output_dir = os.path.join(base_output_dir, f"{model_config['name']}Vis")
os.makedirs(output_dir, exist_ok=True)
# 初始化模型
print(f"\nInitializing {model_config['name']} model...")
model = init_model(model_config['config'], model_config['checkpoint'], device='cuda:0')
# 获取所有PNG图片
image_files = glob.glob(os.path.join(input_dir, '*.jpg'))
total_images = len(image_files)
print(f"Processing {total_images} images with {model_config['name']}...")
# 处理每张图片
for idx, img_path in enumerate(image_files, 1):
img_name = os.path.basename(img_path)
img_name_without_ext = os.path.splitext(img_name)[0]
# 构建输出文件路径
output_path = os.path.join(output_dir, f"{model_config['name']}_{img_name_without_ext}.jpg")
# 推理并保存结果
result = inference_model(model, img_path)
show_result_pyplot(model, img_path, result, show=False, out_file=output_path, opacity=1,with_labels=False)
print(f"[{idx}/{total_images}] Processed {img_name} -> {os.path.basename(output_path)}")
def main():
# 设置输入输出目录
input_dir = '/home/lsl/Workspace/mmsegmentation/data/test-org-img'
base_output_dir = '/home/lsl/Workspace/mmsegmentation/work_dirs/flood-Vis'
# 确保基础输出目录存在
os.makedirs(base_output_dir, exist_ok=True)
# 处理每个模型
for model_config in MODEL_CONFIGS:
try:
print(f"\n{'='*50}")
print(f"Processing with {model_config['name'].upper()} model")
print(f"{'='*50}")
process_images(model_config, input_dir, base_output_dir)
print(f"\nCompleted processing with {model_config['name']} model")
except Exception as e:
print(f"\nError processing {model_config['name']} model: {str(e)}")
continue
if __name__ == '__main__':
main()在进行预测时默认带有类别标签，去除如下：
show_result_pyplot(model, img_path, result, show=False, out_file=output_path, opacity=1,with_labels=False)with_labels=False, # 不显示标签
保存最优权重的：save_best='mIoU'，可以自定义选择你需要的评价指标。
train_cfg = dict(type='IterBasedTrainLoop', max_iters=80000, val_interval=4000)
val_cfg = dict(type='ValLoop')
test_cfg = dict(type='TestLoop')
default_hooks = dict(
timer=dict(type='IterTimerHook'),
logger=dict(type='LoggerHook', interval=50),
param_scheduler=dict(type='ParamSchedulerHook'),
sampler_seed=dict(type='DistSamplerSeedHook'),
checkpoint=dict(
type='CheckpointHook',
by_epoch=False,
interval=4000, # 每 4000 iter 保存一次
save_best='mIoU',
rule='greater',
max_keep_ckpts=3
)
)resume模型一直卡住，也不报错
卡在这里 Advance dataloader 102456 steps to skip data that has already been trained
解决方法：降低mmengine 版本
原本为：
mim install mmengine==0.10.2
如果在子系统上出现训练中断，多次出现在权重保存方面，可能是yapf 版本与 Python 3.8 或 mmengine 不兼容
pip install -U yapf # 升级到最新版
or
pip install yapf==0.32.0 # 安装官方推荐版本
报错：
AssertionError: only one of size and size_divisor should be valid
加上
crop_size = (512, 512)
data_preprocessor = dict(size=crop_size)
base = [
'../base/models/upernet_r50.py',
'../base/datasets/pascal_voc12.py', '../base/default_runtime.py',
'../base/schedules/schedule_40k.py'
]
crop_size = (512, 512)
data_preprocessor = dict(size=crop_size)
model = dict(
data_preprocessor=data_preprocessor,
decode_head=dict(num_classes=10),
auxiliary_head=dict(num_classes=10))