1. 创建openmmlab环境
conda create --name openmmlab python=3.9 -y conda activate openmmlab
2. 安装pytorch
pip install torch==1.11.0+cu102 torchvision==0.12.0+cu102 torchaudio==0.11.0 --extra-index-url https://download.pytorch.org/whl/cu102
在官方资料中写者，10.2版本的cuda有更好的兼容性，所以选择安装cuda10.2版本
对于基于 Ampere 的 NVIDIA GPU，例如 GeForce 30 系列和 NVIDIA A100，CUDA 11 是必须的。
对于较旧的 NVIDIA GPU，CUDA 11 向后兼容，但 CUDA 10.2 提供更好的兼容性并且更轻量级。
3. 安装mmcv-full
pip install -U openmim pip install mmcv-full=={mmcv_version} -f https://download.openmmlab.com/mmcv/dist/cu102/torch1.11.0/index.html # 我使用的安装命令： pip install mmcv-full==1.5.3 -f https://download.openmmlab.com/mmcv/dist/cu113/torch1.11.0/index.html
这里的cuda101与torch1.11.0需要与安装pytorch时匹配
4. 将mmdetection源码下载
- 方法1：git clone下载
git clone https://github.com/open-mmlab/mmdetection.git cd mmdetection pip install -v -e .
- 方法2：本地直接下载再丢到服务器上
5. 安装mmdet
pip install mmdet==2.24.0
这里需要注意，本地下载的mmdet需要与pip install安装的mmdet版本一直，比如我本地下载的源码的2.24.0版本的，那么这里pip install就需要指定安装mmdet==2.24.0版本
6. 推理测试
mim download mmdet --config yolov3_mobilenetv2_320_300e_coco --dest . # 输出： (openmmlab) [fs@localhost mmdetection-2.24.0]$ mim download mmdet --config yolov3_mobilenetv2_320_300e_coco --dest . /home/fs/anaconda3/envs/openmmlab/lib/python3.9/site-packages/_distutils_hack/__init__.py:30: UserWarning: Setuptools is replacing distutils. warnings.warn("Setuptools is replacing distutils.") processing yolov3_mobilenetv2_320_300e_coco... downloading ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 14.5/14.5 MiB 9.9 MB/s eta 0:00:00 Successfully downloaded yolov3_mobilenetv2_320_300e_coco_20210719_215349-d18dff72.pth to /home/fs/LLC/Detection/mmdetection/mmdetection-2.24.0 Successfully dumped yolov3_mobilenetv2_320_300e_coco.py to /home/fs/LLC/Detection/mmdetection/mmdetection-2.24.0
python demo/image_demo.py demo/demo.jpg yolov3_mobilenetv2_320_300e_coco.py yolov3_mobilenetv2_320_300e_coco_20210719_215349-d18dff72.pth --device cpu --out-file result.jpg # 输出： (openmmlab) [fs@localhost mmdetection-2.24.0]$ python demo/image_demo.py demo/demo.jpg yolov3_mobilenetv2_320_300e_coco.py yolov3_mobilenetv2_320_300e_coco_20210719_215349-d18dff72.pth --device cpu --out-file result.jpg load checkpoint from local path: yolov3_mobilenetv2_320_300e_coco_20210719_215349-d18dff72.pth /home/fs/anaconda3/envs/openmmlab/lib/python3.9/site-packages/mmdet/models/dense_heads/yolo_head.py:269: UserWarning: Creating a tensor from a list of numpy.ndarrays is extremely slow. Please consider converting the list to a single numpy.ndarray with numpy.array() before converting to a tensor. (Triggered internally at ../torch/csrc/utils/tensor_new.cpp:210.) flatten_bboxes /= flatten_bboxes.new_tensor(
此时，就可以在目录下看见result.jpg图像，至此mmdet安装成功。如果推理过程中出现任何形如 undefined symbol 或者 ImportError 的错误，都是因为没有正确安装mmdet。
需要注意的就是mmcv-full需要与torch相匹配，然后pip install mmdet需要与下载的源码相匹配。
更多测试例子（具体的参数配置可以查看相应的demo代码）：
单个图像执行推理测试
python demo/image_demo.py demo/demo.jpg configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py checkpoints/faster_rcnn/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth --device cpu
单个视频执行推理测试
python demo/video_demo.py demo/demo.mp4 configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py checkpoints/faster_rcnn/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth --device cpu --out result.mp4
对网络摄像头现场演示
python demo/webcam_demo.py configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py checkpoints/faster_rcnn/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth --device cpu
7. 测试模型数据集
export CUDA_VISIBLE_DEVICES='-1' python tools/test.py configs/pascal_voc/faster_rcnn_r50_fpn_1x_voc0712.py checkpoints/faster_rcnn/faster_rcnn_r50_fpn_1x_voc0712_20220320_192712-54bef0f3.pth --eval mAP # 输出： load checkpoint from local path: checkpoints/faster_rcnn/faster_rcnn_r50_fpn_1x_voc0712_20220320_192712-54bef0f3.pth [>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>] 4952/4952, 0.9 task/s, elapsed: 5806s, ETA: 0s ---------------iou_thr: 0.5--------------- +-------------+------+-------+--------+-------+ | class | gts | dets | recall | ap | +-------------+------+-------+--------+-------+ | aeroplane | 285 | 806 | 0.944 | 0.857 | | bicycle | 337 | 1050 | 0.938 | 0.872 | | bird | 459 | 1054 | 0.891 | 0.797 | | boat | 263 | 1315 | 0.875 | 0.709 | | bottle | 469 | 1421 | 0.802 | 0.700 | | bus | 213 | 694 | 0.939 | 0.859 | | car | 1201 | 3284 | 0.957 | 0.878 | | cat | 358 | 881 | 0.969 | 0.892 | | chair | 756 | 4236 | 0.868 | 0.649 | | cow | 244 | 883 | 0.963 | 0.852 | | diningtable | 206 | 2077 | 0.942 | 0.743 | | dog | 489 | 1294 | 0.982 | 0.878 | | horse | 348 | 1034 | 0.945 | 0.860 | | motorbike | 325 | 990 | 0.932 | 0.849 | | person | 4528 | 11747 | 0.940 | 0.862 | | pottedplant | 480 | 1984 | 0.788 | 0.554 | | sheep | 242 | 716 | 0.938 | 0.822 | | sofa | 239 | 1344 | 0.950 | 0.783 | | train | 282 | 994 | 0.936 | 0.848 | | tvmonitor | 308 | 975 | 0.909 | 0.808 | +-------------+------+-------+--------+-------+ | mAP | | | | 0.804 | +-------------+------+-------+--------+-------+ OrderedDict([('mAP', 0.8035303354263306), ('AP50', 0.804)])
但是，可以现在只验证了VOC2007的数据集部分，没有验证VOC2012的数据集部分，要想两个部分都进行验证，需要重新配置数据集的脚本文件。不过，在VOC2012原本的数据集中本身的Main目录下就没有test.txt，这里就以val.txt来替换，更改如下：
- voc0712_val.py
… data = dict( samples_per_gpu=2, workers_per_gpu=2, train=dict( type='RepeatDataset', times=3, dataset=dict( type=dataset_type, ann_file=[ data_root + 'VOC2007/ImageSets/Main/trainval.txt', data_root + 'VOC2012/ImageSets/Main/trainval.txt' ], img_prefix=[data_root + 'VOC2007/', data_root + 'VOC2012/'], pipeline=train_pipeline)), val=dict( type=dataset_type, ann_file=[ data_root + 'VOC2007/ImageSets/Main/test.txt', data_root + 'VOC2012/ImageSets/Main/test.txt', ], img_prefix=[data_root + 'VOC2007/', data_root + 'VOC2012/'], pipeline=test_pipeline), test=dict( type=dataset_type, ann_file=[ data_root + 'VOC2007/ImageSets/Main/test.txt', data_root + 'VOC2012/ImageSets/Main/val.txt', ], img_prefix=[data_root + 'VOC2007/', data_root + 'VOC2012/'], pipeline=test_pipeline)) evaluation = dict(interval=1, metric='mAP')
- config配置文件：
_base_ = [ '../_base_/models/faster_rcnn_r50_fpn.py', '../_base_/datasets/voc0712_val.py', # 只更改了这个voc数据集的配置文件 '../_base_/default_runtime.py' ] model = dict(roi_head=dict(bbox_head=dict(num_classes=20))) # optimizer optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0001) optimizer_config = dict(grad_clip=None) # learning policy # actual epoch = 3 * 3 = 9 lr_config = dict(policy='step', step=[3]) # runtime settings runner = dict( type='EpochBasedRunner', max_epochs=4) # actual epoch = 4 * 3 = 12
输出：
load checkpoint from local path: checkpoints/faster_rcnn/faster_rcnn_r50_fpn_1x_voc0712_20220320_192712-54bef0f3.pth [>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>] 10775/10775, 38.5 task/s, elapsed: 280s, ETA: 0s Evaluateing data/VOCdevkit/VOC2007/ImageSets/Main/test.txt with 4952 images now ---------------iou_thr: 0.5--------------- +-------------+------+-------+--------+-------+ | class | gts | dets | recall | ap | +-------------+------+-------+--------+-------+ | aeroplane | 285 | 805 | 0.944 | 0.857 | | bicycle | 337 | 1049 | 0.938 | 0.872 | | bird | 459 | 1054 | 0.891 | 0.797 | | boat | 263 | 1315 | 0.875 | 0.709 | | bottle | 469 | 1419 | 0.804 | 0.701 | | bus | 213 | 694 | 0.939 | 0.857 | | car | 1201 | 3284 | 0.957 | 0.878 | | cat | 358 | 882 | 0.969 | 0.892 | | chair | 756 | 4232 | 0.868 | 0.648 | | cow | 244 | 885 | 0.963 | 0.852 | | diningtable | 206 | 2078 | 0.942 | 0.743 | | dog | 489 | 1293 | 0.982 | 0.878 | | horse | 348 | 1033 | 0.945 | 0.860 | | motorbike | 325 | 991 | 0.932 | 0.849 | | person | 4528 | 11745 | 0.940 | 0.862 | | pottedplant | 480 | 1983 | 0.788 | 0.554 | | sheep | 242 | 715 | 0.938 | 0.822 | | sofa | 239 | 1343 | 0.950 | 0.783 | | train | 282 | 997 | 0.936 | 0.848 | | tvmonitor | 308 | 975 | 0.909 | 0.808 | +-------------+------+-------+--------+-------+ | mAP | | | | 0.804 | +-------------+------+-------+--------+-------+ Evaluateing data/VOCdevkit/VOC2012/ImageSets/Main/val.txt with 5823 images now ---------------iou_thr: 0.5--------------- +-------------+------+-------+--------+-------+ | class | gts | dets | recall | ap | +-------------+------+-------+--------+-------+ | aeroplane | 433 | 1150 | 0.970 | 0.950 | | bicycle | 358 | 1087 | 0.958 | 0.929 | | bird | 559 | 1099 | 0.987 | 0.975 | | boat | 424 | 1750 | 0.972 | 0.905 | | bottle | 630 | 1759 | 0.957 | 0.913 | | bus | 301 | 901 | 0.990 | 0.968 | | car | 1004 | 2780 | 0.976 | 0.953 | | cat | 612 | 1246 | 0.993 | 0.985 | | chair | 1176 | 5215 | 0.946 | 0.861 | | cow | 298 | 999 | 0.987 | 0.957 | | diningtable | 305 | 2261 | 0.885 | 0.734 | | dog | 759 | 1649 | 0.997 | 0.989 | | horse | 360 | 963 | 0.981 | 0.965 | | motorbike | 356 | 1094 | 0.986 | 0.959 | | person | 4372 | 11152 | 0.978 | 0.955 | | pottedplant | 489 | 1977 | 0.957 | 0.869 | | sheep | 413 | 1010 | 0.964 | 0.950 | | sofa | 285 | 1353 | 0.951 | 0.812 | | train | 315 | 997 | 0.987 | 0.967 | | tvmonitor | 392 | 1123 | 0.954 | 0.926 | +-------------+------+-------+--------+-------+ | mAP | | | | 0.926 | +-------------+------+-------+--------+-------+ {'0_mAP': 0.8035119771957397, '0_AP50': 0.804, '1_mAP': 0.9260713458061218, '1_AP50': 0.926}
8. 训练现有数据集
- 设置epoch：
# runtime settings = 训练重复次数times x 最大批次次数max_epochs runner = dict( type='EpochBasedRunner', max_epochs=10) # actual epoch = 4 * 3 = 12
- 设置打印次数：
log_config = dict( interval=500, # 每50个迭代就打印一次训练信息 hooks=[ dict(type='TextLoggerHook'), # dict(type='TensorboardLoggerHook') ])
- 设置batch_size：
data = dict( samples_per_gpu=16, # 每个gpu分配16张图像 workers_per_gpu=2, # 分配两个gpu train=dict( type='RepeatDataset', times=3, # 重复训练3次 ... ) )
设置完成后，直接就可以进行python tools/train.py ，进行训练。可以自行进行相关的配置，比如--gpu-id等等。
参考资料：
官方的文档手册：https://mmdetection.readthedocs.io/en/latest/
