基于DSDL数据的训练推理（OpenMMLab）
1. 安装环境
这里使用mmdetection框架进行模型的训练和推理，用户可以参考mmdetection的官网进行安装，目前mmlab2.0(对应mmdet 3.x版本)已经支持DSDL数据集，这里介绍如何安装3.x版本的mmdet，更多信息以及相关的依赖安装请参考MMDetection安装文档.
git clone https://gitlab.shlab.tech/research/dataset_standard/openmmlab-dsdl/mmdetection-dsdl -b dev-3.x
cd mmdetection
pip install -v -e .
安装完mmdetection后，我们可以将数据集下载到mmdetection/data路径下，这里以VOC数据集为例，下载好后的数据集路径如下所示：
mmdetection
└── data
└── PASCAL_VOC2007
├── dsdl
│ ├── config.py
│ ├── defs
│ ├── README.md
│ ├── set-test
│ ├── set-train
│ └── set-val
└── original
├── Annotations
├── ImageSets
├── JPEGImages
├── SegmentationClass
└── SegmentationObject
2. 修改配置文件
mmdet 框架目前支持dsdl数据集的训练，一般只需要对数据集的根目录进行修改即可进行训练。如果是voc，coco等主流数据集，且数据集保存在mmdetection/data路径下，则可以直接使用官网的配置进行训练；如果并非主流数据集，或者数据集没有保存在data路径下，也只需要对配置文件中的路径进行修改即可：
# dataset settings
dataset_type = 'DSDLDetDataset'
data_root = '{path to your dataset}' # 存放数据集的根目录
train_ann = "dsdl/set-train/train.yaml" # 训练的yaml文件
val_ann = "dsdl/set-val/val.yaml" # 验证集的yaml文件
完整的配置文件如下所示，用户也可以根据自身的需求对其它参数进行修改。
dsdl-voc完整训练配置
_base_ = [
'../_base_/models/faster-rcnn_r50_fpn.py',
'../_base_/default_runtime.py'
]
# model setting
model = dict(roi_head=dict(bbox_head=dict(num_classes=20)))
# dataset settings
dataset_type = "DSDLDetDataset"
data_root = "/nvme/wufan/.dsdl/datasets/PascalVOC2007-detection"
train_ann = "yml/set-train/train.yaml"
val_ann = "yml/set-test/test.yaml"
attribute_cfg = dict(
ignore_train= {
"difficult": (True, 1)
}
)
file_client_args = dict(backend='disk')
train_pipeline = [
dict(type='LoadImageFromFile', file_client_args=file_client_args),
dict(type='LoadAnnotations', with_bbox=True),
dict(type='Resize', scale=(1000, 600), keep_ratio=True),
dict(type='RandomFlip', prob=0.5),
dict(type='PackDetInputs')
]
test_pipeline = [
dict(type='LoadImageFromFile', file_client_args=file_client_args),
dict(type='Resize', scale=(1000, 600), keep_ratio=True),
# avoid bboxes being resized
dict(type='LoadAnnotations', with_bbox=True),
dict(
type='PackDetInputs',
meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape', 'scale_factor', 'instances')
)
]
train_dataloader = dict(
batch_size=2,
num_workers=2,
persistent_workers=True,
sampler=dict(type='DefaultSampler', shuffle=True),
batch_sampler=dict(type='AspectRatioBatchSampler'),
dataset=dict(
type=dataset_type,
attribute_cfg=attribute_cfg,
data_root=data_root,
ann_file=train_ann,
filter_cfg=dict(filter_empty_gt=True, min_size=32, bbox_min_size=32),
pipeline=train_pipeline)
)
val_dataloader = dict(
batch_size=1,
num_workers=2,
persistent_workers=True,
drop_last=False,
sampler=dict(type='DefaultSampler', shuffle=False),
dataset=dict(
type=dataset_type,
attribute_cfg=attribute_cfg,
data_root=data_root,
ann_file=val_ann,
test_mode=True,
pipeline=test_pipeline))
test_dataloader = val_dataloader
# Pascal VOC2007 uses `11points` as default evaluate mode, while PASCAL
# VOC2012 defaults to use 'area'.
# val_evaluator = dict(type='VOCMetric', metric='mAP', eval_mode='11points')
val_evaluator = dict(type='CocoMetric', metric='bbox')
test_evaluator = val_evaluator
# training schedule, voc dataset is repeated 3 times, in
# `_base_/datasets/voc0712.py`, so the actual epoch = 4 * 3 = 12
max_epochs = 4
train_cfg = dict(
type='EpochBasedTrainLoop', max_epochs=max_epochs, val_interval=1)
val_cfg = dict(type='ValLoop')
test_cfg = dict(type='TestLoop')
# learning rate
param_scheduler = [
dict(
type='MultiStepLR',
begin=0,
end=max_epochs,
by_epoch=True,
milestones=[3],
gamma=0.1)
]
# optimizer
optim_wrapper = dict(
type='OptimWrapper',
optimizer=dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0001))
# Default setting for scaling LR automatically
# - `enable` means enable scaling LR automatically
# or not by default.
# - `base_batch_size` = (8 GPUs) x (2 samples per GPU).
auto_scale_lr = dict(enable=False, base_batch_size=16)
gpu_ids = range(0, 8)
3 模型训练
- 单卡训练
比如：
- 集群训练
当出现如下日志时，表示训练正在进行中：
2022/12/13 19:00:05 - mmengine - INFO - Checkpoints will be saved to /nvme/wufan/project/mmlab2.x/work_dirs/voc_dsdl.
2022/12/13 19:00:12 - mmengine - INFO - Epoch(train) [1][ 50/3104] lr: 1.0000e-02 eta: 0:26:28 time: 0.1285 data_time: 0.0040 memory: 2581 loss: 0.7176 loss_rpn_cls: 0.1582 loss_rpn_bbox: 0.0301 loss_cls: 0.3771 acc: 92.1875 loss_bbox: 0.1522
2022/12/13 19:00:18 - mmengine - INFO - Epoch(train) [1][ 100/3104] lr: 1.0000e-02 eta: 0:25:32 time: 0.1204 data_time: 0.0024 memory: 2581 loss: 0.5537 loss_rpn_cls: 0.0577 loss_rpn_bbox: 0.0220 loss_cls: 0.2768 acc: 87.8906 loss_bbox: 0.1972
2022/12/13 19:00:24 - mmengine - INFO - Epoch(train) [1][ 150/3104] lr: 1.0000e-02 eta: 0:25:12 time: 0.1210 data_time: 0.0025 memory: 2581 loss: 0.6777 loss_rpn_cls: 0.0885 loss_rpn_bbox: 0.0340 loss_cls: 0.3201 acc: 92.6758 loss_bbox: 0.2350
...
4 模型测试
比如：
推理结果如下：
2022/12/21 11:07:13 - mmengine - INFO - Load checkpoint from work_dirs/voc2007/epoch_4.pth
2022/12/21 11:07:15 - mmengine - INFO - Epoch(test) [ 50/4952] eta: 0:02:17 time: 0.0280 data_time: 0.0028 memory: 348
...
2022/12/21 11:09:15 - mmengine - INFO - Epoch(test) [4950/4952] eta: 0:00:00 time: 0.0246 data_time: 0.0009 memory: 360
2022/12/21 11:09:15 - mmengine - INFO - Converting ground truth to coco format...
2022/12/21 11:09:16 - mmengine - INFO - Evaluating bbox...
2022/12/21 11:09:25 - mmengine - INFO - bbox_mAP_copypaste: 0.454 0.757 0.479 0.123 0.340 0.536
2022/12/21 11:09:25 - mmengine - INFO - Epoch(test) [4952/4952] coco/bbox_mAP: 0.4540 coco/bbox_mAP_50: 0.7570 coco/bbox_mAP_75: 0.4790 coco/bbox_mAP_s: 0.1230 coco/bbox_mAP_m: 0.3400 coco/bbox_mAP_l: 0.5360