d2的自定义训练和增强

1、detectron2中的DatasetMapper类——detectron2如何做数据增强_小麦先生的博客-CSDN博客_detectron2数据增强

2、 目标检测tricks（基于detectron2）_任小喵r的博客-CSDN博客

3、 detectron2/train_net.py at main · facebookresearch/detectron2 · GitHub

注：这个跟我们自己定义的很像，我们自定义的训练脚本如下：

import os
import cv2
import logging
from collections import OrderedDict

import detectron2.utils.comm as comm
from detectron2.utils.visualizer import Visualizer
from detectron2.checkpoint import DetectionCheckpointer
from detectron2.config import get_cfg
from detectron2.data import DatasetCatalog, MetadataCatalog
from detectron2.data.datasets.coco import load_coco_json
from detectron2.engine import DefaultTrainer, default_argument_parser, default_setup, launch
from detectron2.evaluation import COCOEvaluator, verify_results
from detectron2.modeling import GeneralizedRCNNWithTTA

#使用指定的GPU
# os.environ["CUDA_VISIBLE_DEVICES"] = '1'

# 数据集路径
DATASET_ROOT = './datasets/coco'
ANN_ROOT = os.path.join(DATASET_ROOT, 'annotations')
TRAIN_PATH = os.path.join(DATASET_ROOT, 'images/train2017')
VAL_PATH = os.path.join(DATASET_ROOT, 'images/val2017')
TRAIN_JSON = os.path.join(ANN_ROOT, 'instances_train2017.json')
VAL_JSON = os.path.join(ANN_ROOT, 'instances_val2017.json')

# 数据集类别元数据 
# 需修改！！
DATASET_CATEGORIES = [
    {"name": "DECP2", "id": 1, "isthing": 1, "color": [110, 20, 60]},
    {"name": "DECP3", "id": 2, "isthing": 1, "color": [220, 120, 60]},
    {"name": "PACO", "id": 3, "isthing": 1, "color": [220, 20, 120]},
    {"name": "PDCR", "id": 4, "isthing": 1, "color": [120, 20, 160]},
    {"name": "TARE", "id": 5, "isthing": 1, "color": [220, 120, 160]},
    {"name": "PAMS", "id": 6, "isthing": 1, "color": [15, 20, 60]},
    {"name": "TFDE", "id": 7, "isthing": 1, "color": [110, 190, 60]},
    {"name": "TMMS", "id": 8, "isthing": 1, "color": [15, 120, 160]},
    {"name": "EPDF", "id": 9, "isthing": 1, "color": [2, 120, 160]},
    {"name": "PASC", "id": 10, "isthing": 1, "color": [15, 1, 160]},
]

# 数据集的子集
PREDEFINED_SPLITS_DATASET = {
    "train_2019": (TRAIN_PATH, TRAIN_JSON),
    "val_2019": (VAL_PATH, VAL_JSON),
}

def register_dataset():
    """
    purpose: register all splits of dataset with PREDEFINED_SPLITS_DATASET
    """
    for key, (image_root, json_file) in PREDEFINED_SPLITS_DATASET.items():
        register_dataset_instances(name=key,
                                   metadate=get_dataset_instances_meta(),
                                   json_file=json_file,
                                   image_root=image_root)

def get_dataset_instances_meta():
    """
    purpose: get metadata of dataset from DATASET_CATEGORIES
    return: dict[metadata]
    """
    thing_ids = [k["id"] for k in DATASET_CATEGORIES if k["isthing"] == 1]
    thing_colors = [k["color"] for k in DATASET_CATEGORIES if k["isthing"] == 1]
    # assert len(thing_ids) == 2, len(thing_ids)
    thing_dataset_id_to_contiguous_id = {k: i for i, k in enumerate(thing_ids)}
    thing_classes = [k["name"] for k in DATASET_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


def register_dataset_instances(name, metadate, json_file, image_root):
    """
    purpose: register dataset to DatasetCatalog,
             register metadata to MetadataCatalog and set attribute
    """
    DatasetCatalog.register(name, lambda: load_coco_json(json_file, image_root, name))
    MetadataCatalog.get(name).set(json_file=json_file,
                                  image_root=image_root,
                                  evaluator_type="coco",
                                  **metadate)



# 查看数据集标注
def checkout_dataset_annotation(name="val_2019"):
    dataset_dicts = load_coco_json(TRAIN_JSON, TRAIN_PATH, name)
    for d in dataset_dicts:
        img = cv2.imread(d["file_name"])
        visualizer = Visualizer(img[:, :, ::-1], metadata=MetadataCatalog.get(name), scale=1.5)
        vis = visualizer.draw_dataset_dict(d)
        cv2.imshow('show', vis.get_image()[:, :, ::-1])
        cv2.waitKey(0)


class Trainer(DefaultTrainer):
    @classmethod
    def build_evaluator(cls, cfg, dataset_name, output_folder=None):
        if output_folder is None:
            output_folder = os.path.join(cfg.OUTPUT_DIR, "inference")
        return COCOEvaluator(dataset_name, cfg, distributed=False, output_dir=output_folder)

    @classmethod
    def test_with_TTA(cls, cfg, model):
        logger = logging.getLogger("detectron2.trainer")
        # In the end of training, run an evaluation with TTA
        # Only support some R-CNN models.
        logger.info("Running inference with test-time augmentation ...")
        model = GeneralizedRCNNWithTTA(cfg, model)
        evaluators = [
            cls.build_evaluator(
                cfg, name, output_folder=os.path.join(cfg.OUTPUT_DIR, "inference_TTA")
            )
            for name in cfg.DATASETS.TEST
        ]
        res = cls.test(cfg, model, evaluators)
        res = OrderedDict({k + "_TTA": v for k, v in res.items()})
        return res


def setup(args):
    """
    Create configs and perform basic setups.
    """
    cfg = get_cfg()  # 拷贝default config副本
    # args.config_file = "configs/COCO-InstanceSegmentation/mask_rcnn_X_101_32x8d_FPN_3x.yaml"
    args.config_file = "configs/COCO-Detection/retinanet_R_101_FPN_3x.yaml"
    # args.config_file = "configs/COCO-Detection/faster_rcnn_R_101_FPN_3x.yaml"
    # args.config_file = "configs/Misc/mask_rcnn_R_50_FPN_3x_dconv_c3-c5.yaml"
    # args.config_file = "configs/Misc/cascade_mask_rcnn_R_50_FPN_3x.yaml"
    # args.config_file = "configs/Misc/cascade_mask_rcnn_X_152_32x8d_FPN_IN5k_gn_dconv.yaml"
    
    # args.config_file = "configs/Misc/cascade_mask_rcnn_X_152_32x8d_FPN_IN5k_gn_dconv.yaml"
    cfg.merge_from_file(args.config_file)  # 从config file 覆盖配置
    cfg.merge_from_list(args.opts)  # 从CLI参数 覆盖配置

    # 更改配置参数
    cfg.DATASETS.TRAIN = ("train_2019",)
    cfg.DATASETS.TEST = ("val_2019",)
    cfg.DATALOADER.NUM_WORKERS = 16  #
    cfg.INPUT.MAX_SIZE_TRAIN = 1300
    cfg.INPUT.MAX_SIZE_TEST = 1300       #test
    cfg.INPUT.MIN_SIZE_TRAIN = (1000,)   # （160，）
    cfg.INPUT.MIN_SIZE_TEST = 1000     #test
    cfg.MODEL.ROI_HEADS.NUM_CLASSES = 10  # 类别数
    cfg.MODEL.MASK_ON = False
    cfg.MODEL.RETINANET.NUM_CLASSES = 10             #80  retrainet
    # cfg.MODEL.WEIGHTS = "./models/cascade_mask_rcnn_R_50_FPN_3x/model_final_480dd8.pkl"  # 预训练模型权重
    # cfg.MODEL.WEIGHTS = "./models/R-50.pkl"  # 预训练模型权重
    # cfg.MODEL.WEIGHTS = "./models/R-101.pkl"    # 预训练模型权重
    # cfg.MODEL.WEIGHTS = "./models/X-152-32x8d-IN5k.pkl"
    # cfg.MODEL.WEIGHTS = "./models/mask_rcnn_X_101_32x8d_FPN_3x/model_final_2d9806.pkl"
    # cfg.MODEL.WEIGHTS = "./output/model_0003999.pth"    # 预训练模型权重
    # cfg.MODEL.WEIGHTS = "./output/model_final.pth"    # 预训练模型权重

    cfg.MODEL.WEIGHTS = "./models/retinanet_R_101_FPN_3x/model_final_971ab9.pkl"
    # cfg.MODEL.WEIGHTS = "./models/cascade_mask_rcnn_x152/model_0039999_e76410.pkl"    # cat uesd it
    # cfg.MODEL.WEIGHTS = "./models/mask_rcnn_R_50_FPN_3x_dconv_c3-c5/model_final_821d0b.pkl"  # 预训练模型权重
    cfg.SOLVER.IMS_PER_BATCH = 8 # batch_size=2; iters_in_one_epoch = dataset_imgs/batch_size
    ITERS_IN_ONE_EPOCH = int(960 / cfg.SOLVER.IMS_PER_BATCH)
    cfg.SOLVER.MAX_ITER = (ITERS_IN_ONE_EPOCH * 40) - 1  # 4k
    cfg.SOLVER.BASE_LR = 0.002  # LR
    cfg.SOLVER.MOMENTUM = 0.9
    cfg.SOLVER.WEIGHT_DECAY = 0.0001
    cfg.SOLVER.WEIGHT_DECAY_NORM = 0.0
    cfg.MODEL.RETINANET.SCORE_THRESH_TEST = 0.8
    cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5
    cfg.MODEL.PANOPTIC_FPN.COMBINE.INSTANCES_CONFIDENCE_THRESH = 0.8
    #学习率线性
    # cfg.SOLVER.WARMUP_METHOD = "linear" 
    # cfg.SOLVER.GAMMA = 0.7  # lr down rate
    # cfg.SOLVER.STEPS = (1500, 2000, 2500, 3000, 3500, 4000, 4500, 5000, 5500)  # lr delay  each step
    # cfg.SOLVER.WARMUP_FACTOR = 1.0 / 1000  
    # cfg.SOLVER.WARMUP_ITERS = 1000  # lr up under to base_lr
    
    #学习率cos          
    cfg.SOLVER.LR_SCHEDULER_NAME ="WarmupCosineLR"
    cfg.SOLVER.MAX_ITER = 4600
    cfg.SOLVER.WARMUP_FACTOR = 1.0 / 1000
    cfg.SOLVER.WARMUP_ITERS = 500  # lr up under to base_lr

    cfg.SOLVER.CHECKPOINT_PERIOD = 500   # save model  interval
    #cfg.SOLVER.CHECKPT_PERIOD = 1  # save model  interval??? no work

    cfg.TEST.EVAL_PERIOD = 750        # 迭代到指定次数，进行一次评估
    cfg.freeze()
    default_setup(cfg, args)
    return cfg


def main(args):
    cfg = setup(args)
    print(cfg)
    # 注册数据集
    register_dataset()
    #只测试不训练
    if args.eval_only:
        model = Trainer.build_model(cfg)
        DetectionCheckpointer(model, save_dir=cfg.OUTPUT_DIR).resume_or_load(
            cfg.MODEL.WEIGHTS, resume=args.resume
        )
        res = Trainer.test(cfg, model)
        if comm.is_main_process():
            verify_results(cfg, res)
        return res
    #开始训练
    trainer = Trainer(cfg)
    trainer.resume_or_load(resume=args.resume)        # If `resume==True`, and last checkpoint exists, resume from it, load all checkpointables
    return trainer.train()


if __name__ == "__main__":
    args = default_argument_parser().parse_args()
    print("Command Line Args:", args)
    launch(
        main,
        args.num_gpus,             #使用的gpu数量
        num_machines=args.num_machines,
        machine_rank=args.machine_rank,
        dist_url=args.dist_url,
        args=(args,),
    )

4、 https://github.com/facebookresearch/detectron2/blob/main/tools/plain_train_net.py

5、Detectron2 代码 难 解读 - 连载中_空洞卷积-重新思考底层特征-程序员资料 - 程序员资料

 6、目标检测tricks（基于detectron2）_任小喵r的博客-CSDN博客

7、detectron2和mmdetection对比_小涵涵的博客-程序员秘密_mmdetection和detectron2 - 程序员秘密

 8、[Detectron2] 03-数据加载模块(dataloader) - 知乎

9、使用 Detectron2 中的 mask-RCNN 检测建筑物轮廓_RS2GIS的博客-CSDN博客_武汉大学建筑物数据集

10、[Detectron2] 03-自定义数据增强样例-StrictRandomCrop - 知乎

11、detectron2模型使用、读写、训练及测试_alex1801的博客-CSDN博客_detectron2模型导出

 12、Detectron2训练自己的数据集（较详细）_qq_29750461的博客-CSDN博客_detectron2训练自己的数据集

13、detectron2（目标检测框架）无死角玩转-00：目录_江南才尽，年少无知！的博客-CSDN博客_detectron2框架 

14、detectron2/defaults.py at 45b3fcea6e76bf7a351e54e01c7d6e1a3a0100a5 · facebookresearch/detectron2 · GitHub

（后面自己写推理导出onnx时可以参考）

15、Sartorius-Cell-Instance-Segmentation/x2-customed-mosaic-with-albumentations-detectron2.ipynb at main · aengusng8/Sartorius-Cell-Instance-Segmentation · GitHub

 可以参考推理：

https://github.com/aengusng8/Sartorius-Cell-Instance-Segmentation​​​​​​​）

 

16、Detectron2-数据增强（Data Augmentation）官方文档中文翻译__KATANA_的博客-CSDN博客_detectron2 数据增强

17、copy_paste_aug_detectron2/detectron2_copypaste.ipynb at main · MarkPotanin/copy_paste_aug_detectron2 · GitHub

（copy-paste技术！）

18、'CUDA out of memory' after applying custom augmentations · Issue #3709 · facebookresearch/detectron2 · GitHub