detectron2使用教程20200824_2

训练COCO数据(liuhao_coco_kouzhao)

1.准备数据集

#1.创建软链接
ln -s ../../../datasets/kouzhao/liuhao_coco_kouzhao/ ./

/detectron2-master/detectron2/data/datasets/builtin.py


    "liuhao_coco_kouzhao_train2017": ("liuhao_coco_kouzhao/train/train2017", "liuhao_coco_kouzhao/train/annotations/instances_train2017.json"), #自定义数据集
    "liuhao_coco_kouzhao_val2017": ("liuhao_coco_kouzhao/val/val2017", "liuhao_coco_kouzhao/val/annotations/instances_val2017.json"),       #自定义数据集

/detectron2-master/detectron2/data/datasets/builtin_meta.py第7#注意COCO_CATEGORIES要与coco的json的categories相同

  "categories": [
    {
      "supercategory": "none",
      "id": 0,
      "name": "no_mask"
    },
    {
      "supercategory": "none",
      "id": 1,
      "name": "have_mask"
    }
  ]

COCO_CATEGORIES = [
    {"color": [220, 20, 60], "isthing": 1, "id": 0, "name": "no_mask"},
    {"color": [119, 11, 32], "isthing": 1, "id": 1, "name": "have_mask"}
]


COCO_CATEGORIES = [
    {"color": [220, 20, 60], "isthing": 1, "id": 0, "name": "0"},
    {"color": [119, 11, 32], "isthing": 1, "id": 1, "name": "1"},
    {"color": [0, 0, 142], "isthing": 1, "id": 2, "name": "2"},
    {"color": [0, 0, 230], "isthing": 1, "id": 3, "name": "3"},
    {"color": [106, 0, 228], "isthing": 1, "id": 4, "name": "4"},
    {"color": [0, 60, 100], "isthing": 1, "id": 5, "name": "5"},
    {"color": [0, 80, 100], "isthing": 1, "id": 6, "name": "6"},
    {"color": [0, 0, 70], "isthing": 1, "id": 7, "name": "7"},
    {"color": [0, 0, 192], "isthing": 1, "id": 8, "name": "8"},
    {"color": [250, 170, 30], "isthing": 1, "id": 9, "name": "9"},
    {"color": [100, 170, 30], "isthing": 1, "id": 10, "name": "A"},
    {"color": [220, 220, 0], "isthing": 1, "id": 11, "name": "B"},
    {"color": [175, 116, 175], "isthing": 1, "id": 12, "name": "C"},
    {"color": [250, 0, 30], "isthing": 1, "id": 13, "name": "D"},
    {"color": [165, 42, 42], "isthing": 1, "id": 14, "name": "E"},
    {"color": [255, 77, 255], "isthing": 1, "id": 15, "name": "F"},
    {"color": [0, 226, 252], "isthing": 1, "id": 16, "name": "U"},
    {"color": [182, 182, 255], "isthing": 1, "id": 17, "name": "V"},
    {"color": [0, 82, 0], "isthing": 1, "id": 18, "name": "W"},
    {"color": [120, 166, 157], "isthing": 1, "id": 19, "name": "X"},
    {"color": [110, 76, 0], "isthing": 1, "id": 20, "name": "Y"},
    {"color": [174, 57, 255], "isthing": 1, "id": 21, "name": "Z"},
]


/detectron2-master/detectron2/data/datasets/builtin_meta.py第194#assert len(thing_ids) == 80, len(thing_ids)
    assert len(thing_ids) == 2, len(thing_ids) #########################################################



/detectron2-master/detectron2/data/datasets/builtin_meta.py第211#assert len(stuff_ids) == 53, len(stuff_ids)
    assert len(stuff_ids) == 0, len(stuff_ids) #########################################################


这个还不确定要不要修改,不修改也能训练

./detectron2/config/defaults.py:
_C.MODEL.ROI_HEADS.NUM_CLASSES = 80

2.修改配置文件./configs/COCO-Detection/faster_rcnn_R_50_FPN_1x_liuhao_coco_kouzhao

_BASE_: "../Base-RCNN-FPN_liuhao_coco_kouzhao.yaml"
MODEL:
  WEIGHTS: "detectron2://ImageNetPretrained/MSRA/R-50.pkl"
  MASK_ON: False
  RESNETS:
    DEPTH: 50

Base-RCNN-FPN_liuhao_coco_kouzhao.yaml

DATASETS:
  TRAIN: ("liuhao_coco_kouzhao_train2017",) ########### coco_2017_train
  TEST: ("liuhao_coco_kouzhao_val2017",) ############## coco_2017_val
SOLVER:
  IMS_PER_BATCH: 2 #######################
  BASE_LR: 0.02
  STEPS: (600, 700) #######################
  MAX_ITER: 800 ############# 自定义最大迭代次数=(训练集400张图片/batch_size2) * 4epochs
INPUT:
  MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800)
VERSION: 2

3.训练

#在/detectron2-master文件夹下面打开终端

source ~/anaconda3/bin/activate mydetection_python3.7.3


python3 ./tools/train_net.py -h


python3 ./tools/train_net.py --config-file=./configs/COCO-Detection/faster_rcnn_R_50_FPN_1x_liuhao_coco_kouzhao.yaml --num-gpus=1 SOLVER.IMS_PER_BATCH 2 SOLVER.BASE_LR 0.01
#查看/detectron2-master/output/


报错,估计是生成的coco数据的categories没有去重

AssertionError: Attribute 'thing_classes' in the metadata of 'liuhao_coco_kouzhao_train2017' cannot be set to a different value!

4训练成功,部分日志

[04/28 22:11:05 d2.data.datasets.coco]: Loaded 400 images in COCO format from datasets/liuhao_coco_kouzhao/train/annotations/instances_train2017.json
[04/28 22:11:05 d2.data.build]: Removed 68 images with no usable annotations. 332 images left.
[04/28 22:11:05 d2.data.build]: Distribution of instances among all 2 categories:
|  category  | #instances   |  category  | #instances   |
|:----------:|:-------------|:----------:|:-------------|
|  no_mask   | 390          | have_mask  | 521          |
|            |              |            |              |
|   total    | 911          |            |              |




[04/28 22:20:40 fvcore.common.checkpoint]: Saving checkpoint to ./output/model_final.pth
[04/28 22:20:41 d2.data.datasets.coco]: Loaded 177 images in COCO format from datasets/liuhao_coco_kouzhao/val/annotations/instances_val2017.json
[04/28 22:20:41 d2.data.build]: Distribution of instances among all 2 categories:
|  category  | #instances   |  category  | #instances   |
|:----------:|:-------------|:----------:|:-------------|
|  no_mask   | 139          | have_mask  | 209          |
|            |              |            |              |
|   total    | 348          |            |              |


修改./output/为./output_liuhao_coco_kouzhao/

5可视化

/home/liuhao/anaconda3/envs/mybase_python3.7.3/bin/tensorboard --logdir ./output_liuhao_coco_kouzhao/

6预测


python3 ./demo/demo.py --config-file=./configs/COCO-Detection/faster_rcnn_R_50_FPN_1x_liuhao_coco_kouzhao.yaml --input=./datasets/liuhao_coco_kouzhao/val/val2017/402.jpg --opts MODEL.WEIGHTS ./output_liuhao_coco_kouzhao/model_final.pth


python3 ./demo/demo.py --config-file=./configs/COCO-Detection/faster_rcnn_R_50_FPN_1x_liuhao_coco_kouzhao.yaml --input=./datasets/liuhao_coco_kouzhao/val/val2017/569.jpg --opts MODEL.WEIGHTS ./output_liuhao_coco_kouzhao/model_final.pth


#可以预测,效果不错

  • 9
    点赞
  • 19
    收藏
    觉得还不错? 一键收藏
  • 11
    评论

“相关推荐”对你有帮助么?

  • 非常没帮助
  • 没帮助
  • 一般
  • 有帮助
  • 非常有帮助
提交
评论 11
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值