mmdetection训练自己的数据集
一、准备数据集
准备自己的数据
mmdetection支持coco格式和voc格式的数据集，下面将分别介绍这两种数据集的使用方式
coco数据集
官方推荐coco数据集按照以下的目录形式存储，以coco2017数据集为例
mmdetection
├── mmdet
├── tools
├── configs
├── data
│ ├── coco
│ │ ├── annotations
│ │ ├── train2017
│ │ ├── val2017
│ │ ├── test2017推荐以软连接的方式创建data文件夹，下面是创建软连接的步骤
cd mmdetection
mkdir data
ln -s $COCO_ROOT data其中，$COCO_ROOT需改为你的coco数据集根目录
voc数据集
与coco数据集类似，将voc数据集按照以下的目录形式存储，以VOC2007为例
mmdetection
├── mmdet
├── tools
├── configs
├── data
│ ├── VOCdevkit
│ │ ├── VOC2007
│ │ │ ├── Annotations
│ │ │ ├── JPEGImages
│ │ │ ├── ImageSets
│ │ │ │ ├── Main
│ │ │ │ │ ├── test.txt
│ │ │ │ │ ├── trainval.txt同样推荐以软连接的方式创建
cd mmdetection
mkdir data
ln -s $VOC2007_ROOT data/VOCdevkit其中，$VOC2007_ROOT需改为你的VOC2007数据集根目录
二、修改一些配置文件和代码文件
修改配置文件，配置文件在configs文件夹下面，根据自己的情况进行选择，
本人选择的是configs/mask_rcnn_r101_fpn_1x.py根据自己情况修改说明，如果选择faster rcnn请根据自己情况进行修改：
# model settings
model = dict(
type='MaskRCNN',
pretrained='torchvision://resnet101',
backbone=dict(
type='ResNet',
depth=101,
num_stages=4,
out_indices=(0, 1, 2, 3),
frozen_stages=1,
style='pytorch'),
neck=dict(
type='FPN',
in_channels=[256, 512, 1024, 2048],
out_channels=256,
num_outs=5),
rpn_head=dict(
type='RPNHead',
in_channels=256,
feat_channels=256,
anchor_scales=[8],
anchor_ratios=[0.5, 1.0, 2.0],
anchor_strides=[4, 8, 16, 32, 64],
target_means=[.0, .0, .0, .0],
target_stds=[1.0, 1.0, 1.0, 1.0],
loss_cls=dict(
type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
loss_bbox=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=1.0)),
bbox_roi_extractor=dict(
type='SingleRoIExtractor',
roi_layer=dict(type='RoIAlign', out_size=7, sample_num=2),
out_channels=256,
featmap_strides=[4, 8, 16, 32]),
bbox_head=dict(
type='SharedFCBBoxHead',
num_fcs=2,
in_channels=256,
fc_out_channels=1024,
roi_feat_size=7,
num_classes=6,#数据集类别数,默认是81,因为coco数据集为80+1(背景),我的数据集只有5个类别,加上背景也就是6个类别
target_means=[0., 0., 0., 0.],
target_stds=[0.1, 0.1, 0.2, 0.2],
reg_class_agnostic=False,
loss_cls=dict(
type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0)),
mask_roi_extractor=dict(
type='SingleRoIExtractor',
roi_layer=dict(type='RoIAlign', out_size=14, sample_num=2),
out_channels=256,
featmap_strides=[4, 8, 16, 32]),
mask_head=dict(
type='FCNMaskHead',
num_convs=4,
in_channels=256,
conv_out_channels=256,
num_classes=6,#数据集类别数,默认是81,因为coco数据集为80+1(背景),我的数据集只有5个类别,加上背景也就是6个类别
loss_mask=dict(
type='CrossEntropyLoss', use_mask=True, loss_weight=1.0)))
# model training and testing settings
train_cfg = dict(
rpn=dict(
assigner=dict(
type='MaxIoUAssigner',
pos_iou_thr=0.7,
neg_iou_thr=0.3,
min_pos_iou=0.3,
ignore_iof_thr=-1),
sampler=dict(
type='RandomSampler',
num=256,
pos_fraction=0.5,
neg_pos_ub=-1,
add_gt_as_proposals=False),
allowed_border=0,
pos_weight=-1,
debug=False),
rpn_proposal=dict(
nms_across_levels=False,
nms_pre=2000,
nms_post=2000,
max_num=2000,
nms_thr=0.7,
min_bbox_size=0),
rcnn=dict(
assigner=dict(
type='MaxIoUAssigner',
pos_iou_thr=0.5,
neg_iou_thr=0.5,
min_pos_iou=0.5,
ignore_iof_thr=-1),
sampler=dict(
type='RandomSampler',
num=512,
pos_fraction=0.25,
neg_pos_ub=-1,
add_gt_as_proposals=True),
mask_size=28,
pos_weight=-1,
debug=False))
test_cfg = dict(
rpn=dict(
nms_across_levels=False,
nms_pre=1000,
nms_post=1000,
max_num=1000,
nms_thr=0.7,
min_bbox_size=0),
rcnn=dict(
score_thr=0.05,
nms=dict(type='nms', iou_thr=0.5),
max_per_img=100,
mask_thr_binary=0.5))
# dataset settings
dataset_type = 'CocoDataset'
data_root = 'data/coco/'
img_norm_cfg = dict(
mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
train_pipeline = [
dict(type='LoadImageFromFile'),
dict(type='LoadAnnotations', with_bbox=True, with_mask=True),
dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
dict(type='RandomFlip', flip_ratio=0.5),
dict(type='Normalize', **img_norm_cfg),
dict(type='Pad', size_divisor=32),
dict(type='DefaultFormatBundle'),
dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_masks']),
]
test_pipeline = [
dict(type='LoadImageFromFile'),
dict(
type='MultiScaleFlipAug',
img_scale=(1333, 800),
flip=False,
transforms=[
dict(type='Resize', keep_ratio=True),
dict(type='RandomFlip'),
dict(type='Normalize', **img_norm_cfg),
dict(type='Pad', size_divisor=32),
dict(type='ImageToTensor', keys=['img']),
dict(type='Collect', keys=['img']),
])
]
data = dict(
imgs_per_gpu=2,#每张gpu训练多少张图片 batch_size = gpu_num(训练使用gpu数量) * imgs_per_gpu
workers_per_gpu=1,
train=dict(
type=dataset_type,
ann_file=data_root + 'annotations/instances_train2014.json',
img_prefix=data_root + 'train2014/',
pipeline=train_pipeline),
val=dict(
type=dataset_type,
ann_file=data_root + 'annotations/instances_val2014.json',
img_prefix=data_root + 'val2014/',
pipeline=test_pipeline),
test=dict(
type=dataset_type,
ann_file=data_root + 'annotations/instances_val2014.json',
img_prefix=data_root + 'val2014/',
pipeline=test_pipeline))
# optimizer
optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0001)#lr = 0.00125*batch_siz
optimizer_config = dict(grad_clip=dict(max_norm=35, norm_type=2))
# learning policy
lr_config = dict(
policy='step',
warmup='linear',
warmup_iters=500,
warmup_ratio=1.0 / 3,
step=[8, 11])
checkpoint_config = dict(interval=1)
# yapf:disable
log_config = dict(
interval=50,
hooks=[
dict(type='TextLoggerHook'),
# dict(type='TensorboardLoggerHook')#如果需要开启tensorboard将其注释取消
])
# yapf:enable
# runtime settings
total_epochs = 12#总的循环次数
dist_params = dict(backend='nccl')
log_level = 'INFO'
work_dir = './work_dirs/mask_rcnn_r101_fpn_1x'#在训练过程中会将训练日志和权重保存在这个文件夹路径下面，详细看下面
load_from = None
resume_from = None
workflow = [('train', 1)]注意：配置文件中的默认学习率是8个gpu和2个img/gpu(batch size= 8*2 = 16)。根据线性缩放规则，如果您使用不同的GPU数目或img/gpu，您需要设置与batch size成比例的学习率。例如，如果4GPUs * 2 img/gpu的lr=0.01，那么16GPUs * 4 img/gpu的lr=0.08。
每张gpu训练多少张图片 batch_size = gpu_num(训练使用gpu数量) * imgs_per_gpu
lr = 0.00125*batch_siz
进一步修改一些地方:
定义数据种类，需要修改的地方在mmdetection/mmdet/datasets/coco.py。把CLASSES的那个tuple改为自己数据集对应的种类tuple即可。例如：
CLASSES = ('bicycle', 'car', 'bus', 'person','tvmonitor')接着在mmdetection/mmdet/core/evaluation/class_names.py修改coco_classes数据集类别，这个关系到后面test的时候结果图中显示的类别名称。例如：
def coco_classes():
return [
'bicycle', 'car', 'bus', 'person','tvmonitor'
]三、训练：
单个GPU训练：
第一次训练会下载预训练模型，如下
(open-mmlab) bubble@XPS-8930:~/mmdetection/0827/mmdetection$ python tools/train.py configs/mask_rcnn_r101_fpn_1x.py
2019-09-16 22:14:50,684 - INFO - Distributed training: False
2019-09-16 22:14:51,399 - INFO - load model from: torchvision://resnet101
Downloading: "https://download.pytorch.org/models/resnet101-5d3b4d8f.pth" to /home/bubble/.cache/torch/checkpoints/resnet101-5d3b4d8f.pth
100.0%训练过程log输出
(open-mmlab) bubble@XPS-8930:~/mmdetection/0827/mmdetection$ python tools/train.py configs/mask_rcnn_r101_fpn_1x.py
2019-09-18 21:31:10,284 - INFO - Distributed training: False
2019-09-18 21:31:11,067 - INFO - load model from: torchvision://resnet101
2019-09-18 21:31:13,031 - WARNING - The model and loaded state dict do not match exactly
unexpected key in source state_dict: fc.weight, fc.bias
loading annotations into memory...
Done (t=0.17s)
creating index...
index created!
2019-09-18 21:31:23,956 - INFO - Start running, host: bubble@XPS-8930, work_dir: /home/bubble/mmdetection/0827/mmdetection/work_dirs/mask_rcnn_r101_fpn_1x
2019-09-18 21:31:23,956 - INFO - workflow: [('train', 1)], max: 12 epochs
2019-09-18 21:31:46,277 - INFO - Epoch [1][50/8001] lr: 0.00797, eta: 11:53:25, time: 0.446, data_time: 0.012, memory: 3241, loss_rpn_cls: 0.2262, loss_rpn_bbox: 0.0602, loss_cls: 0.7983, acc: 94.5195, loss_bbox: 0.0827, loss_mask: 0.6030, loss: 1.7704
2019-09-18 21:32:05,929 - INFO - Epoch [1][100/8001] lr: 0.00931, eta: 11:10:49, time: 0.393, data_time: 0.004, memory: 3241, loss_rpn_cls: 0.1831, loss_rpn_bbox: 0.0712, loss_cls: 0.5860, acc: 94.6523, loss_bbox: 0.1161, loss_mask: 0.4851, loss: 1.4414
......
......
......
2019-09-19 08:42:04,113 - INFO - Epoch [12][7950/8001] lr: 0.00020, eta: 0:00:21, time: 0.427, data_time: 0.005, memory: 3290, loss_rpn_cls: 0.0112, loss_rpn_bbox: 0.0188, loss_cls: 0.1462, acc: 95.1758, loss_bbox: 0.0890, loss_mask: 0.1974, loss: 0.4626
2019-09-19 08:42:25,439 - INFO - Epoch [12][8000/8001] lr: 0.00020, eta: 0:00:00, time: 0.427, data_time: 0.004, memory: 3290, loss_rpn_cls: 0.0106, loss_rpn_bbox: 0.0215, loss_cls: 0.1168, acc: 95.9844, loss_bbox: 0.0715, loss_mask: 0.1791, loss: 0.3995在训练过程
mmdetection整个文件夹情况：
.
├── build
├── checkpoints
├── configs
├── data
├── demo
├── docker
├── docs
├── LICENSE
├── mmdet
├── mmdet.egg-info
├── README.md
├── requirements.txt
├── setup.py
├── tests
├── tools
└── work_dirs训练过程中log和权重保存路径
work_dirs/
└── mask_rcnn_r101_fpn_1x
├── 20190918_213123.log
├── 20190918_213123.log.json
├── epoch_10.pth
├── epoch_11.pth
├── epoch_12.pth
├── epoch_1.pth
├── epoch_2.pth
├── epoch_3.pth
├── epoch_4.pth
├── epoch_5.pth
├── epoch_6.pth
├── epoch_7.pth
├── epoch_8.pth
├── epoch_9.pth
└── latest.pth -> epoch_12.pth四、训练过程各种损失值和准确率可视化效果
4.1、tensorboard
开启tensorboard，记得在config配置文件里将dict(type='TensorboardLoggerHook')注释取消掉
在新的终端中执行如下命令：
tensorboard --logdir=path --port=8090#port=8090可以自己指定的端口，默认不需要--port其端口是60064.1.1、本地跑mmdetection的话直接在PC的浏览器上输入如下链接
http://127.0.0.1:160064.1.2、远程访问服务器上面的tensorboard
bubble@XPS-8930:~$ ssh -p 10005 -L 16006:127.0.0.1:8090 [email protected]
#10005是自己的docker系统镜像的端口号，-L 16006:127.0.0.18090意思是将自己PC的16006端口映射成docker里tensorboard的8090
#接下来，在PC的浏览器上输入如下链接
http://127.0.0.1:16006过程如下图所示：
tensorboard效果如下所示：
4.2、mmdetection自带的log分析工具
mmdetection会自动收集log信息，存储在work_dirs/目录下，官方提供了tools/analyze_logs.py工具可以轻松的可视化日志信息，如可视化损失值并保存为pdf文件，执行如下命令：
python tools/analyze_logs.py plot_curve log.json --keys loss_cls loss_reg --out losses.pdf效果如图所示：
mmdetection使用tensorboard可视化训练集与验证集指标参数
- https://blog.csdn.net/weicao1990/article/details/103324218
mmdetection使用其自带工具可视化训练集与验证集指标参数
- https://github.com/open-mmlab/mmdetection/blob/master/docs/GETTING_STARTED.md
具体可以查看官网：
- https://github.com/open-mmlab/mmdetection
- https://github.com/open-mmlab/mmdetection/blob/master/docs/GETTING_STARTED.md
五、Testing
有两个方法可以进行测试。
5.1、如果只是想看一下效果而不要进行定量指标分析的话，可以运行之前那个demo.py文件，但是要改一下checkpoint_file的地址路径，使用我们上一步跑出来的work_dirs下的pth文件。例如：
checkpoint_file = 'work_dirs/epoch_100.pth'5.2、使用test命令来进行测试评估一些参数
5.2.1 coco数据集，例如：
python tools/test.py configs/your_confige.py work_dirs/your_model_.pth --out ./result/result_100.pkl --eval bbox类似于下面：
root@64e7169a4f30:/mmdetection# python tools/test.py configs/mask_rcnn_r101_fpn_1x.py work_dirs/mask_rcnn_r101_fpn_1x/latest.pth --eval bbox --out results1.pkl
loading annotations into memory...
Done (t=0.00s)
creating index...
index created!
[>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>] 67/67, 17.7 task/s, elapsed: 4s, ETA: 0s
writing results to results1.pkl
Starting evaluate bbox
Loading and preparing results...
DONE (t=0.00s)
creating index...
index created!
Running per image evaluation...
Evaluate annotation type *bbox*
DONE (t=0.04s).
Accumulating evaluation results...
DONE (t=0.01s).
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.698
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.917
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.826
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.742
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = -1.000
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = -1.000
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.734
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.744
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.744
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.744
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = -1.000
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = -1.0005.2.2 voc数据集，例如
修改tools/voc_eval.py文件中的voc_eval函数
注释如下代码：
if hasattr(dataset, 'year') and dataset.year == 2007:
dataset_name = 'voc07'
else:
dataset_name = dataset.CLASSES在eval_map上面添加：
dataset_name = dataset.CLASSES修改后的voc_eval函数代码如下：
def voc_eval(result_file, dataset, iou_thr=0.5):
det_results = mmcv.load(result_file)
gt_bboxes = []
gt_labels = []
gt_ignore = []
for i in range(len(dataset)):
ann = dataset.get_ann_info(i)
bboxes = ann['bboxes']
labels = ann['labels']
if 'bboxes_ignore' in ann:
ignore = np.concatenate([
np.zeros(bboxes.shape[0], dtype=np.bool),
np.ones(ann['bboxes_ignore'].shape[0], dtype=np.bool)
])
gt_ignore.append(ignore)
bboxes = np.vstack([bboxes, ann['bboxes_ignore']])
labels = np.concatenate([labels, ann['labels_ignore']])
gt_bboxes.append(bboxes)
gt_labels.append(labels)
if not gt_ignore:
gt_ignore = None
# if hasattr(dataset, 'year') and dataset.year == 2007:
# dataset_name = 'voc07'
# else:
# dataset_name = dataset.CLASSES
dataset_name = dataset.CLASSES
eval_map(
det_results,
gt_bboxes,
gt_labels,
gt_ignore=gt_ignore,
scale_ranges=None,
iou_thr=iou_thr,
dataset=dataset_name,
print_summary=True)执行如下命令：
python tools/test.py configs/your_confige.py work_dirs/your_model_.pth --out results.pkl测试结束后生成results.pkl文件
采用voc标准计算mAP
执行如下命令：
python tools/voc_eval.py results.pkl configs/your_confige.py类似下面这样子：
参考链接：
- https://blog.csdn.net/hajlyx/article/details/83542167
- https://blog.csdn.net/gaoyi135/article/details/90613895
- https://heary.cn/posts/mmdetection-%E5%9F%BA%E4%BA%8EPyTorch%E7%9A%84%E5%BC%80%E6%BA%90%E7%9B%AE%E6%A0%87%E6%A3%80%E6%B5%8B%E7%B3%BB%E7%BB%9F/
- http://lanck.lzu.edu.cn/?p=490
- https://github.com/open-mmlab/mmdetection/blob/master/docs/GETTING_STARTED.md
- https://blog.csdn.net/weicao1990/article/details/93484603
- https://tianchi.aliyun.com/notebook-ai/detail?postId=74264