AdelaiDet训练自己的数据集（BlendMask训练COCO格式数据集）

AdelaiDet训练自己的数据集（COCO格式数据集）

不了解coco数据集形式的可以先去查找了解一下

环境配置为 Ubuntu 20.04.3/cuda10.1/python3.7/pythorch1.8

下面的detecton2就是你conda环境中的安装的detectron2，照着下面找一下：
/home/anaconda3/envs/你的conda环境名/lib/python3.7/site-packages/detectron2

注册数据集

从detectron2/data/datasets中修改builtin.py

# builtin中添加的内容
# 在_PREDEFINED_SPLITS_COCO["coco"]中添加自己数据集的路径
_PREDEFINED_SPLITS_COCO["coco"] = {# 自己的数据集：# "coco_handwritten_train"为自定义名称，后面会用到；后面的两个路径第一个是图片目录，第二个是对应json文件路径# 训练集"coco_handwritten_train": ("/mnt/big_disk/big_disk_1/zhuzhibo/download/AdelaiDet/datasets/handwritten_chinese_stroke_2021/train2021","/mnt/big_disk/big_disk_1/zhuzhibo/download/AdelaiDet/datasets/handwritten_chinese_stroke_2021/annotations/instances_train2021.json"),# 验证集"coco_handwritten_val": ("/mnt/big_disk/big_disk_1/zhuzhibo/download/AdelaiDet/datasets/handwritten_chinese_stroke_2021/val2021","/mnt/big_disk/big_disk_1/zhuzhibo/download/AdelaiDet/datasets/handwritten_chinese_stroke_2021/annotations/instances_val2021.json"),
}
...
# 更改if __name__.endswith(".builtin"):下的_root
if __name__.endswith(".builtin"):# Assume pre-defined datasets live in `./datasets`.# 这里后面更改为你自己的AdelaiDet的根目录_root = os.getenv("DETECTRON2_DATASETS", "/mnt/big_disk/big_disk_1/zhuzhibo/download/AdelaiDet")#_root = os.getenv("/mnt/big_disk/big_disk_1/zhuzhibo/download/AdelaiDet", "datasets")

从detectron2/data/datasets中修改builtin_meta.py
这里注意类别的定义顺序要与你转coco数据集时的一样（我没试不一样会怎么样），颜色填不重复的就行，isthing为1，id从1开始，name为这个类别的名称

# 将原来的COCO_CATEGORIES注释掉，换为新的
COCO_CATEGORIES = [{"color": [220, 20, 60], "isthing": 1, "id": 1, "name": "wangou"},{"color": [119, 11, 32], "isthing": 1, "id": 2, "name": "na"},{"color": [0, 0, 142], "isthing": 1, "id": 3, "name": "ti"},{"color": [0, 0, 230], "isthing": 1, "id": 4, "name": "pie"},{"color": [0, 0, 230], "isthing": 1, "id": 5, "name": "piezhe"},{"color": [106, 0, 228], "isthing": 1, "id": 6, "name": "piedian"},{"color": [0, 60, 100], "isthing": 1, "id": 7, "name": "xiegouhuowogou"},{"color": [0, 80, 100], "isthing": 1, "id": 8, "name": "heng"},{"color": [0, 0, 70], "isthing": 1, "id": 9, "name": "hengzhe"},{"color": [0, 0, 192], "isthing": 1, "id": 10, "name": "hengzhezhehuohengzhewan"},{"color": [100, 170, 30], "isthing": 1, "id": 11, "name": "hengzhezhezhe"},{"color": [220, 220, 0], "isthing": 1, "id": 12, "name": "hengzhezhezhegouhuohengpiewangou"},{"color": [175, 116, 175], "isthing": 1, "id": 13, "name": "hengzhezhepie"},{"color": [250, 0, 30], "isthing": 1, "id": 14, "name": "hengzheti"},{"color": [165, 42, 42], "isthing": 1, "id": 15, "name": "hengzhegou"},{"color": [255, 77, 255], "isthing": 1, "id": 16, "name": "hengpiehuohenggou"},{"color": [0, 226, 252], "isthing": 1, "id": 17, "name": "hengxiegou"},{"color": [182, 182, 255], "isthing": 1, "id": 18, "name": "dian"},{"color": [0, 82, 0], "isthing": 1, "id": 19, "name": "shu"},{"color": [120, 166, 157], "isthing": 1, "id": 20, "name": "shuwan"},{"color": [110, 76, 0], "isthing": 1, "id": 21, "name": "shuwangou"},{"color": [174, 57, 255], "isthing": 1, "id": 22, "name": "shuzhezhegou"},{"color": [199, 100, 0], "isthing": 1, "id": 23, "name": "shuzhepiehuoshuzhezhe"},{"color": [72, 0, 118], "isthing": 1, "id": 24, "name": "shuti"},{"color": [169, 164, 131], "isthing": 1, "id": 25, "name": "shugou"},
]# 更改_getcoco_instances_mata()如下
# assert len(thing_ids) == 25, len(thing_ids) 我这里是25类，改成你自己的类别数（不算背景类）
def _get_coco_instances_meta():thing_ids = [k["id"] for k in COCO_CATEGORIES if k["isthing"] == 1]thing_colors = [k["color"] for k in COCO_CATEGORIES if k["isthing"] == 1]assert len(thing_ids) == 25, len(thing_ids)# Mapping from the incontiguous COCO category id to an id in [0, 24]thing_dataset_id_to_contiguous_id = {k: i for i, k in enumerate(thing_ids)}thing_classes = [k["name"] for k in COCO_CATEGORIES if k["isthing"] == 1]ret = {"thing_dataset_id_to_contiguous_id": thing_dataset_id_to_contiguous_id,"thing_classes": thing_classes,"thing_colors": thing_colors,}return ret# 更改_get_coco_panoptic_separated_meat()中以下内容assert len(stuff_ids) == 0, len(stuff_ids)#assert len(stuff_ids) == 53, len(stuff_ids) 原来的语句

更改AdelaiDet下的datasets/prepare_thins_sem_from_instacne.py(这步可能不需要，可以先不改这里)

def get_parser():parser = argparse.ArgumentParser(description="Keep only model in ckpt")parser.add_argument("--dataset-name",default="handwritten_chinese_stroke_2021",help="dataset to generate",)return parserif __name__ == "__main__":args = get_parser().parse_args()dataset_dir = os.path.join(os.path.dirname(__file__), args.dataset_name)if args.dataset_name == "handwritten_chinese_stroke_2021":thing_id_to_contiguous_id = _get_coco_instances_meta()["thing_dataset_id_to_contiguous_id"]split_name = 'train2021'annotation_name = "annotations/instances_{}.json"...

更改AdelaiDet/configs/BlendMask中的Base-BlendMask.yaml文件

MODEL:META_ARCHITECTURE: "BlendMask"MASK_ON: TrueBACKBONE:NAME: "build_fcos_resnet_fpn_backbone"RESNETS:OUT_FEATURES: ["res3", "res4", "res5"]FPN:IN_FEATURES: ["res3", "res4", "res5"]PROPOSAL_GENERATOR:NAME: "FCOS"BASIS_MODULE:LOSS_ON: TruePANOPTIC_FPN:COMBINE:ENABLED: FalseFCOS:THRESH_WITH_CTR: TrueUSE_SCALE: False
DATALOADER: 8
DATASETS: # 这里写自己之前第一步注册的名字TRAIN: ("coco_handwritten_train",)  # 训练TEST: ("coco_handwritten_val",)     # 验证
SOLVER:IMS_PER_BATCH: 32BASE_LR: 0.001  # Note that RetinaNet uses a different default learning rateSTEPS: (200, 400)MAX_ITER: 600
INPUT:MIN_SIZE_TRAIN: (120, 120)  # 训练最小图片尺寸，按照你的需求修改MIN_SIZE_TEST: 120          # 验证最小图片尺寸，按照你的需求修改

更改你需要使用配置的yaml，这里以R_101_3x.yaml为例

_BASE_: "Base-BlendMask.yaml"
MODEL:WEIGHTS: "detectron2://ImageNetPretrained/MSRA/R-101.pkl"BACKBONE:NAME: "build_fcos_resnet_fpn_backbone"RESNETS:DEPTH: 101FCOS:NUM_CLASSES: 25  # 更改为自定义数据集类别
# INPUT的内容根据自己需要修改
INPUT:FORMAT: BGRMASK_FORMAT: polygonMAX_SIZE_TEST: 144MAX_SIZE_TRAIN: 144MIN_SIZE_TEST: 112MIN_SIZE_TRAIN: [ 112, 120 ]
SOLVER:STEPS: [2500, 29999]MAX_ITER: 30000 # default 270000WARMUP_ITERS: 100 # default 1000CHECKPOINT_PERIOD: 2500 # default 5000IMS_PER_BATCH: 64 # default 16BASE_LR: 0.025 # default 0.02
TEST:EVAL_PERIOD: 2500 # default 0, means no evaluation will be performed
VIS_PERIOD: 2500 # default 0, means no visualization will be performed
DATASETS:  # 可以覆盖Base-BlendMask.yaml中的配置TRAIN: ("coco_handwritten_train",)TEST: ("coco_handwritten_val",)
OUTPUT_DIR: "training_dir/blendmask/hw_R_101_3x_fpn"  # 训练输出目录

更改AdelaiDet/adet/data/dataset_mapper.py

        if self.basis_loss_on and self.is_train:# load basis supervisions"""if self.ann_set == "coco":basis_sem_path = (dataset_dict["file_name"].replace("train2017", "thing_train2017").replace("image/train", "thing_train"))"""# 更改上面内容为下面if self.ann_set == "coco":basis_sem_path = (dataset_dict["file_name"].replace("train2021", "thing_train2021"))

更改AdelaiDet/tools/train_net.py

def setup(args):...# 更改使用的yaml配置文件args.config_file = '/home/zhuzhibo/AdelaiDet/configs/BlendMask/R_101_3x.yaml'...