在mmdet框架中使用config配置文件构建网络模型

from   mmdet.models.builder import build_detector
from   mmcv import Config
import torch
import mmcv

from   os.path import join,realpath,relpath
import glob
config_dpath = 'configs'+"/yolo"
config_fpaths = list(glob.glob(join(config_dpath,'*.py')))
config_fpaths = [p for p in config_fpaths if p.find('_base_') == -1]
config_names = [relpath(p, config_dpath) for p in config_fpaths]

config_names

['yolov3 - 副本.py','yolov3.py','yolov3_d53_320_273e_coco.py','yolov3_d53_mstrain-416_273e_coco.py','yolov3_d53_mstrain-608_273e_coco.py']

1. 解析yolov3配置文件

for config_fname in config_names[1:2]:config_fpath = join(config_dpath, config_fname)config_mod = Config.fromfile(config_fpath)
print(config_fname,'\n',config_fpath,'\n',config_mod)

yolov3.py configs/yolo\yolov3.py Config (path: configs/yolo\yolov3.py): {'model': {'type': 'YOLOV3', 'pretrained': 'open-mmlab://darknet53', 'backbone': {'type': 'Darknet', 'depth': 53, 'out_indices': (3, 4, 5)}, 'neck': {'type': 'YOLOV3Neck', 'num_scales': 3, 'in_channels': [1024, 512, 256], 'out_channels': [512, 256, 128]}, 'bbox_head': {'type': 'YOLOV3Head', 'num_classes': 80, 'in_channels': [512, 256, 128], 'out_channels': [1024, 512, 256], 'anchor_generator': {'type': 'YOLOAnchorGenerator', 'base_sizes': [[(116, 90), (156, 198), (373, 326)], [(30, 61), (62, 45), (59, 119)], [(10, 13), (16, 30), (33, 23)]], 'strides': [32, 16, 8]}, 'bbox_coder': {'type': 'YOLOBBoxCoder'}, 'featmap_strides': [32, 16, 8], 'loss_cls': {'type': 'CrossEntropyLoss', 'use_sigmoid': True, 'loss_weight': 1.0, 'reduction': 'sum'}, 'loss_conf': {'type': 'CrossEntropyLoss', 'use_sigmoid': True, 'loss_weight': 1.0, 'reduction': 'sum'}, 'loss_xy': {'type': 'CrossEntropyLoss', 'use_sigmoid': True, 'loss_weight': 2.0, 'reduction': 'sum'}, 'loss_wh': {'type': 'MSELoss', 'loss_weight': 2.0, 'reduction': 'sum'}}, 'train_cfg': {'assigner': {'type': 'GridAssigner', 'pos_iou_thr': 0.5, 'neg_iou_thr': 0.5, 'min_pos_iou': 0}}, 'test_cfg': {'nms_pre': 1000, 'min_bbox_size': 0, 'score_thr': 0.05, 'conf_thr': 0.005, 'nms': {'type': 'nms', 'iou_threshold': 0.45}, 'max_per_img': 100}}, 'dataset_type': 'CocoDataset', 'data_root': 'F:/datasets/COCO2017/', 'img_norm_cfg': {'mean': [0, 0, 0], 'std': [255.0, 255.0, 255.0], 'to_rgb': True}, 'test_pipeline': [{'type': 'LoadImageFromFile', 'to_float32': True}, {'type': 'LoadAnnotations', 'with_bbox': True}, {'type': 'PhotoMetricDistortion'}, {'type': 'Resize', 'img_scale': [(320, 320), (608, 608)], 'keep_ratio': True}], 'train_pipeline': [{'type': 'LoadImageFromFile', 'to_float32': True}, {'type': 'LoadAnnotations', 'with_bbox': True}, {'type': 'PhotoMetricDistortion'}, {'type': 'Resize', 'img_scale': [(320, 320), (608, 608)], 'keep_ratio': True}], 'data': {'samples_per_gpu': 8, 'workers_per_gpu': 4, 'train': {'type': 'CocoDataset', 'ann_file': 'F:/datasets/COCO2017/annotations/instances_train2017.json', 'img_prefix': 'F:/datasets/COCO2017/train2017/', 'pipeline': [{'type': 'LoadImageFromFile', 'to_float32': True}, {'type': 'LoadAnnotations', 'with_bbox': True}, {'type': 'PhotoMetricDistortion'}, {'type': 'Resize', 'img_scale': [(320, 320), (608, 608)], 'keep_ratio': True}]}, 'val': {'type': 'CocoDataset', 'ann_file': 'F:/datasets/COCO2017/annotations/instances_val2017.json', 'img_prefix': 'F:/datasets/COCO2017/val2017/', 'pipeline': [{'type': 'LoadImageFromFile', 'to_float32': True}, {'type': 'LoadAnnotations', 'with_bbox': True}, {'type': 'PhotoMetricDistortion'}, {'type': 'Resize', 'img_scale': [(320, 320), (608, 608)], 'keep_ratio': True}]}, 'test': {'type': 'CocoDataset', 'ann_file': 'F:/datasets/COCO2017/annotations/instances_val2017.json', 'img_prefix': 'F:/datasets/COCO2017/val2017/', 'pipeline': [{'type': 'LoadImageFromFile', 'to_float32': True}, {'type': 'LoadAnnotations', 'with_bbox': True}, {'type': 'PhotoMetricDistortion'}, {'type': 'Resize', 'img_scale': [(320, 320), (608, 608)], 'keep_ratio': True}]}}}

for k,v in  config_mod.items():print(k,":",v)

model : {'type': 'YOLOV3', 'pretrained': 'open-mmlab://darknet53', 'backbone': {'type': 'Darknet', 'depth': 53, 'out_indices': (3, 4, 5)}, 'neck': {'type': 'YOLOV3Neck', 'num_scales': 3, 'in_channels': [1024, 512, 256], 'out_channels': [512, 256, 128]}, 'bbox_head': {'type': 'YOLOV3Head', 'num_classes': 80, 'in_channels': [512, 256, 128], 'out_channels': [1024, 512, 256], 'anchor_generator': {'type': 'YOLOAnchorGenerator', 'base_sizes': [[(116, 90), (156, 198), (373, 326)], [(30, 61), (62, 45), (59, 119)], [(10, 13), (16, 30), (33, 23)]], 'strides': [32, 16, 8]}, 'bbox_coder': {'type': 'YOLOBBoxCoder'}, 'featmap_strides': [32, 16, 8], 'loss_cls': {'type': 'CrossEntropyLoss', 'use_sigmoid': True, 'loss_weight': 1.0, 'reduction': 'sum'}, 'loss_conf': {'type': 'CrossEntropyLoss', 'use_sigmoid': True, 'loss_weight': 1.0, 'reduction': 'sum'}, 'loss_xy': {'type': 'CrossEntropyLoss', 'use_sigmoid': True, 'loss_weight': 2.0, 'reduction': 'sum'}, 'loss_wh': {'type': 'MSELoss', 'loss_weight': 2.0, 'reduction': 'sum'}}, 'train_cfg': {'assigner': {'type': 'GridAssigner', 'pos_iou_thr': 0.5, 'neg_iou_thr': 0.5, 'min_pos_iou': 0}}, 'test_cfg': {'nms_pre': 1000, 'min_bbox_size': 0, 'score_thr': 0.05, 'conf_thr': 0.005, 'nms': {'type': 'nms', 'iou_threshold': 0.45}, 'max_per_img': 100}}
dataset_type : CocoDataset
data_root : data/coco/
img_norm_cfg : {'mean': [0, 0, 0], 'std': [255.0, 255.0, 255.0], 'to_rgb': True}
test_pipeline : []
train_pipeline : []
data : {'samples_per_gpu': 8, 'workers_per_gpu': 4, 'train': {'type': 'CocoDataset', 'ann_file': 'data/coco/annotations/instances_train2017.json', 'img_prefix': 'data/coco/train2017/', 'pipeline': []}, 'val': {'type': 'CocoDataset', 'ann_file': 'data/coco/annotations/instances_val2017.json', 'img_prefix': 'data/coco/val2017/', 'pipeline': []}, 'test': {'type': 'CocoDataset', 'ann_file': 'data/coco/annotations/instances_val2017.json', 'img_prefix': 'data/coco/val2017/', 'pipeline': []}}

config_mod.model

{'type': 'YOLOV3','pretrained': 'open-mmlab://darknet53','backbone': {'type': 'Darknet', 'depth': 53, 'out_indices': (3, 4, 5)},'neck': {'type': 'YOLOV3Neck','num_scales': 3,'in_channels': [1024, 512, 256],'out_channels': [512, 256, 128]},'bbox_head': {'type': 'YOLOV3Head','num_classes': 80,'in_channels': [512, 256, 128],'out_channels': [1024, 512, 256],'anchor_generator': {'type': 'YOLOAnchorGenerator','base_sizes': [[(116, 90), (156, 198), (373, 326)],[(30, 61), (62, 45), (59, 119)],[(10, 13), (16, 30), (33, 23)]],'strides': [32, 16, 8]},'bbox_coder': {'type': 'YOLOBBoxCoder'},'featmap_strides': [32, 16, 8],'loss_cls': {'type': 'CrossEntropyLoss','use_sigmoid': True,'loss_weight': 1.0,'reduction': 'sum'},'loss_conf': {'type': 'CrossEntropyLoss','use_sigmoid': True,'loss_weight': 1.0,'reduction': 'sum'},'loss_xy': {'type': 'CrossEntropyLoss','use_sigmoid': True,'loss_weight': 2.0,'reduction': 'sum'},'loss_wh': {'type': 'MSELoss', 'loss_weight': 2.0, 'reduction': 'sum'}},'train_cfg': {'assigner': {'type': 'GridAssigner','pos_iou_thr': 0.5,'neg_iou_thr': 0.5,'min_pos_iou': 0}},'test_cfg': {'nms_pre': 1000,'min_bbox_size': 0,'score_thr': 0.05,'conf_thr': 0.005,'nms': {'type': 'nms', 'iou_threshold': 0.45},'max_per_img': 100}}

if 'pretrained' in config_mod.model:config_mod.model['pretrained'] = None

2.构建检测器（使用构建函数和hook）

from mmdet.models import build_detector
model=build_detector(config_mod.model)

print(model)

YOLOV3((backbone): Darknet((conv1): ConvModule((conv): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv_res_block1): Sequential((conv): ConvModule((conv): Conv2d(32, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(res0): ResBlock((conv1): ConvModule((conv): Conv2d(64, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))))(conv_res_block2): Sequential((conv): ConvModule((conv): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(res0): ResBlock((conv1): ConvModule((conv): Conv2d(128, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res1): ResBlock((conv1): ConvModule((conv): Conv2d(128, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))))(conv_res_block3): Sequential((conv): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(res0): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res1): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res2): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res3): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res4): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res5): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res6): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res7): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))))(conv_res_block4): Sequential((conv): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(res0): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res1): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res2): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res3): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res4): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res5): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res6): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res7): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))))(conv_res_block5): Sequential((conv): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(res0): ResBlock((conv1): ConvModule((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res1): ResBlock((conv1): ConvModule((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res2): ResBlock((conv1): ConvModule((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res3): ResBlock((conv1): ConvModule((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))))(neck): YOLOV3Neck((detect1): DetectionBlock((conv1): ConvModule((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv3): ConvModule((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv4): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv5): ConvModule((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(detect2): DetectionBlock((conv1): ConvModule((conv): Conv2d(768, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv3): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv4): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv5): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(conv2): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(detect3): DetectionBlock((conv1): ConvModule((conv): Conv2d(384, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv3): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv4): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv5): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))))(bbox_head): YOLOV3Head((loss_cls): CrossEntropyLoss()(loss_conf): CrossEntropyLoss()(loss_xy): CrossEntropyLoss()(loss_wh): MSELoss()(convs_bridge): ModuleList((0): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(1): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(convs_pred): ModuleList((0): Conv2d(1024, 255, kernel_size=(1, 1), stride=(1, 1))(1): Conv2d(512, 255, kernel_size=(1, 1), stride=(1, 1))(2): Conv2d(256, 255, kernel_size=(1, 1), stride=(1, 1))))
)

from mmdet.models import DETECTORS
DETECTORS.module_dict

{'SingleStageDetector': mmdet.models.detectors.single_stage.SingleStageDetector,'ATSS': mmdet.models.detectors.atss.ATSS,'TwoStageDetector': mmdet.models.detectors.two_stage.TwoStageDetector,'CascadeRCNN': mmdet.models.detectors.cascade_rcnn.CascadeRCNN,'CornerNet': mmdet.models.detectors.cornernet.CornerNet,'DETR': mmdet.models.detectors.detr.DETR,'FastRCNN': mmdet.models.detectors.fast_rcnn.FastRCNN,'FasterRCNN': mmdet.models.detectors.faster_rcnn.FasterRCNN,'FCOS': mmdet.models.detectors.fcos.FCOS,'FOVEA': mmdet.models.detectors.fovea.FOVEA,'FSAF': mmdet.models.detectors.fsaf.FSAF,'GFL': mmdet.models.detectors.gfl.GFL,'GridRCNN': mmdet.models.detectors.grid_rcnn.GridRCNN,'HybridTaskCascade': mmdet.models.detectors.htc.HybridTaskCascade,'MaskRCNN': mmdet.models.detectors.mask_rcnn.MaskRCNN,'MaskScoringRCNN': mmdet.models.detectors.mask_scoring_rcnn.MaskScoringRCNN,'NASFCOS': mmdet.models.detectors.nasfcos.NASFCOS,'PAA': mmdet.models.detectors.paa.PAA,'PointRend': mmdet.models.detectors.point_rend.PointRend,'RepPointsDetector': mmdet.models.detectors.reppoints_detector.RepPointsDetector,'RetinaNet': mmdet.models.detectors.retinanet.RetinaNet,'RPN': mmdet.models.detectors.rpn.RPN,'SCNet': mmdet.models.detectors.scnet.SCNet,'SparseRCNN': mmdet.models.detectors.sparse_rcnn.SparseRCNN,'TridentFasterRCNN': mmdet.models.detectors.trident_faster_rcnn.TridentFasterRCNN,'VFNet': mmdet.models.detectors.vfnet.VFNet,'YOLACT': mmdet.models.detectors.yolact.YOLACT,'YOLOV3': mmdet.models.detectors.yolo.YOLOV3}

# model=build_detector({'type':'YOLOV5','backbone':{}, 'neck':{},'bbox_head':{}})

m=DETECTORS.get(config_mod.model['type'])
# print(m(backbone={},neck={},bbox_head={}))

2.1 构建主干网络

from mmdet.models import build_backbone,BACKBONES
BACKBONES.module_dict

{'Darknet': mmdet.models.backbones.darknet.Darknet,'ResNet': mmdet.models.backbones.resnet.ResNet,'ResNetV1d': mmdet.models.backbones.resnet.ResNetV1d,'DetectoRS_ResNet': mmdet.models.backbones.detectors_resnet.DetectoRS_ResNet,'DetectoRS_ResNeXt': mmdet.models.backbones.detectors_resnext.DetectoRS_ResNeXt,'HourglassNet': mmdet.models.backbones.hourglass.HourglassNet,'HRNet': mmdet.models.backbones.hrnet.HRNet,'ResNeXt': mmdet.models.backbones.resnext.ResNeXt,'RegNet': mmdet.models.backbones.regnet.RegNet,'Res2Net': mmdet.models.backbones.res2net.Res2Net,'ResNeSt': mmdet.models.backbones.resnest.ResNeSt,'SSDVGG': mmdet.models.backbones.ssd_vgg.SSDVGG,'TridentResNet': mmdet.models.backbones.trident_resnet.TridentResNet}

2.1.1 Use backbone builder

backbone_config=config_mod.model['backbone']
print(backbone_config)

{'type': 'Darknet', 'depth': 53, 'out_indices': (3, 4, 5)}

backone=build_backbone(config_mod.model['backbone'])
print(backone)

Darknet((conv1): ConvModule((conv): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv_res_block1): Sequential((conv): ConvModule((conv): Conv2d(32, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(res0): ResBlock((conv1): ConvModule((conv): Conv2d(64, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))))(conv_res_block2): Sequential((conv): ConvModule((conv): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(res0): ResBlock((conv1): ConvModule((conv): Conv2d(128, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res1): ResBlock((conv1): ConvModule((conv): Conv2d(128, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))))(conv_res_block3): Sequential((conv): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(res0): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res1): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res2): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res3): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res4): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res5): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res6): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res7): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))))(conv_res_block4): Sequential((conv): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(res0): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res1): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res2): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res3): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res4): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res5): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res6): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res7): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))))(conv_res_block5): Sequential((conv): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(res0): ResBlock((conv1): ConvModule((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res1): ResBlock((conv1): ConvModule((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res2): ResBlock((conv1): ConvModule((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res3): ResBlock((conv1): ConvModule((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))))
)

2.1.2 Use code hook

bo=BACKBONES.get(backbone_config['type'])
bc=backbone_config.pop('type')
print(bc)

Darknet

backbone_config

{'depth': 53, 'out_indices': (3, 4, 5)}

bo(**backbone_config)

Darknet((conv1): ConvModule((conv): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv_res_block1): Sequential((conv): ConvModule((conv): Conv2d(32, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(res0): ResBlock((conv1): ConvModule((conv): Conv2d(64, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))))(conv_res_block2): Sequential((conv): ConvModule((conv): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(res0): ResBlock((conv1): ConvModule((conv): Conv2d(128, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res1): ResBlock((conv1): ConvModule((conv): Conv2d(128, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))))(conv_res_block3): Sequential((conv): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(res0): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res1): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res2): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res3): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res4): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res5): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res6): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res7): ResBlock((conv1): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))))(conv_res_block4): Sequential((conv): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(res0): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res1): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res2): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res3): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res4): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res5): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res6): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res7): ResBlock((conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))))(conv_res_block5): Sequential((conv): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(res0): ResBlock((conv1): ConvModule((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res1): ResBlock((conv1): ConvModule((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res2): ResBlock((conv1): ConvModule((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(res3): ResBlock((conv1): ConvModule((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))))
)

2.2 Builder necks

from mmdet.models import build_neck,NECKS
NECKS.module_dict

{'BFP': mmdet.models.necks.bfp.BFP,'ChannelMapper': mmdet.models.necks.channel_mapper.ChannelMapper,'FPG': mmdet.models.necks.fpg.FPG,'FPN': mmdet.models.necks.fpn.FPN,'FPN_CARAFE': mmdet.models.necks.fpn_carafe.FPN_CARAFE,'HRFPN': mmdet.models.necks.hrfpn.HRFPN,'NASFPN': mmdet.models.necks.nas_fpn.NASFPN,'NASFCOS_FPN': mmdet.models.necks.nasfcos_fpn.NASFCOS_FPN,'PAFPN': mmdet.models.necks.pafpn.PAFPN,'RFP': mmdet.models.necks.rfp.RFP,'YOLOV3Neck': mmdet.models.necks.yolo_neck.YOLOV3Neck}

neck = build_neck(config_mod.model['neck'])

neck

YOLOV3Neck((detect1): DetectionBlock((conv1): ConvModule((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv3): ConvModule((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv4): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv5): ConvModule((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(detect2): DetectionBlock((conv1): ConvModule((conv): Conv2d(768, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv3): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv4): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv5): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(conv2): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(detect3): DetectionBlock((conv1): ConvModule((conv): Conv2d(384, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv3): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv4): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv5): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))
)

neck_config=config_mod.model['neck']
neck_config.pop('type')
print(neck_config)

{'num_scales': 3, 'in_channels': [1024, 512, 256], 'out_channels': [512, 256, 128]}

NECKS.get('YOLOV3Neck')(**neck_config)

YOLOV3Neck((detect1): DetectionBlock((conv1): ConvModule((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv3): ConvModule((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv4): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv5): ConvModule((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(conv1): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(detect2): DetectionBlock((conv1): ConvModule((conv): Conv2d(768, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv3): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv4): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv5): ConvModule((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(conv2): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(detect3): DetectionBlock((conv1): ConvModule((conv): Conv2d(384, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv3): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv4): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(conv5): ConvModule((conv): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))
)

2.3 构建检测头 head

from  mmdet.models import build_head,HEADS
HEADS.module_dict

{'AnchorFreeHead': mmdet.models.dense_heads.anchor_free_head.AnchorFreeHead,'AnchorHead': mmdet.models.dense_heads.anchor_head.AnchorHead,'ATSSHead': mmdet.models.dense_heads.atss_head.ATSSHead,'RPNHead': mmdet.models.dense_heads.rpn_head.RPNHead,'StageCascadeRPNHead': mmdet.models.dense_heads.cascade_rpn_head.StageCascadeRPNHead,'CascadeRPNHead': mmdet.models.dense_heads.cascade_rpn_head.CascadeRPNHead,'CornerHead': mmdet.models.dense_heads.corner_head.CornerHead,'CentripetalHead': mmdet.models.dense_heads.centripetal_head.CentripetalHead,'EmbeddingRPNHead': mmdet.models.dense_heads.embedding_rpn_head.EmbeddingRPNHead,'FCOSHead': mmdet.models.dense_heads.fcos_head.FCOSHead,'FoveaHead': mmdet.models.dense_heads.fovea_head.FoveaHead,'RetinaHead': mmdet.models.dense_heads.retina_head.RetinaHead,'FreeAnchorRetinaHead': mmdet.models.dense_heads.free_anchor_retina_head.FreeAnchorRetinaHead,'FSAFHead': mmdet.models.dense_heads.fsaf_head.FSAFHead,'GuidedAnchorHead': mmdet.models.dense_heads.guided_anchor_head.GuidedAnchorHead,'GARetinaHead': mmdet.models.dense_heads.ga_retina_head.GARetinaHead,'GARPNHead': mmdet.models.dense_heads.ga_rpn_head.GARPNHead,'GFLHead': mmdet.models.dense_heads.gfl_head.GFLHead,'NASFCOSHead': mmdet.models.dense_heads.nasfcos_head.NASFCOSHead,'PAAHead': mmdet.models.dense_heads.paa_head.PAAHead,'PISARetinaHead': mmdet.models.dense_heads.pisa_retinanet_head.PISARetinaHead,'SSDHead': mmdet.models.dense_heads.ssd_head.SSDHead,'PISASSDHead': mmdet.models.dense_heads.pisa_ssd_head.PISASSDHead,'RepPointsHead': mmdet.models.dense_heads.reppoints_head.RepPointsHead,'RetinaSepBNHead': mmdet.models.dense_heads.retina_sepbn_head.RetinaSepBNHead,'SABLRetinaHead': mmdet.models.dense_heads.sabl_retina_head.SABLRetinaHead,'TransformerHead': mmdet.models.dense_heads.transformer_head.TransformerHead,'VFNetHead': mmdet.models.dense_heads.vfnet_head.VFNetHead,'YOLACTHead': mmdet.models.dense_heads.yolact_head.YOLACTHead,'YOLACTSegmHead': mmdet.models.dense_heads.yolact_head.YOLACTSegmHead,'YOLACTProtonet': mmdet.models.dense_heads.yolact_head.YOLACTProtonet,'YOLOV3Head': mmdet.models.dense_heads.yolo_head.YOLOV3Head,'BBoxHead': mmdet.models.roi_heads.bbox_heads.bbox_head.BBoxHead,'ConvFCBBoxHead': mmdet.models.roi_heads.bbox_heads.convfc_bbox_head.ConvFCBBoxHead,'Shared2FCBBoxHead': mmdet.models.roi_heads.bbox_heads.convfc_bbox_head.Shared2FCBBoxHead,'Shared4Conv1FCBBoxHead': mmdet.models.roi_heads.bbox_heads.convfc_bbox_head.Shared4Conv1FCBBoxHead,'DIIHead': mmdet.models.roi_heads.bbox_heads.dii_head.DIIHead,'DoubleConvFCBBoxHead': mmdet.models.roi_heads.bbox_heads.double_bbox_head.DoubleConvFCBBoxHead,'SABLHead': mmdet.models.roi_heads.bbox_heads.sabl_head.SABLHead,'SCNetBBoxHead': mmdet.models.roi_heads.bbox_heads.scnet_bbox_head.SCNetBBoxHead,'CascadeRoIHead': mmdet.models.roi_heads.cascade_roi_head.CascadeRoIHead,'StandardRoIHead': mmdet.models.roi_heads.standard_roi_head.StandardRoIHead,'DoubleHeadRoIHead': mmdet.models.roi_heads.double_roi_head.DoubleHeadRoIHead,'DynamicRoIHead': mmdet.models.roi_heads.dynamic_roi_head.DynamicRoIHead,'GridRoIHead': mmdet.models.roi_heads.grid_roi_head.GridRoIHead,'HybridTaskCascadeRoIHead': mmdet.models.roi_heads.htc_roi_head.HybridTaskCascadeRoIHead,'FCNMaskHead': mmdet.models.roi_heads.mask_heads.fcn_mask_head.FCNMaskHead,'CoarseMaskHead': mmdet.models.roi_heads.mask_heads.coarse_mask_head.CoarseMaskHead,'FeatureRelayHead': mmdet.models.roi_heads.mask_heads.feature_relay_head.FeatureRelayHead,'FusedSemanticHead': mmdet.models.roi_heads.mask_heads.fused_semantic_head.FusedSemanticHead,'GlobalContextHead': mmdet.models.roi_heads.mask_heads.global_context_head.GlobalContextHead,'GridHead': mmdet.models.roi_heads.mask_heads.grid_head.GridHead,'HTCMaskHead': mmdet.models.roi_heads.mask_heads.htc_mask_head.HTCMaskHead,'MaskPointHead': mmdet.models.roi_heads.mask_heads.mask_point_head.MaskPointHead,'MaskIoUHead': mmdet.models.roi_heads.mask_heads.maskiou_head.MaskIoUHead,'SCNetMaskHead': mmdet.models.roi_heads.mask_heads.scnet_mask_head.SCNetMaskHead,'SCNetSemanticHead': mmdet.models.roi_heads.mask_heads.scnet_semantic_head.SCNetSemanticHead,'MaskScoringRoIHead': mmdet.models.roi_heads.mask_scoring_roi_head.MaskScoringRoIHead,'PISARoIHead': mmdet.models.roi_heads.pisa_roi_head.PISARoIHead,'PointRendRoIHead': mmdet.models.roi_heads.point_rend_roi_head.PointRendRoIHead,'SCNetRoIHead': mmdet.models.roi_heads.scnet_roi_head.SCNetRoIHead,'SparseRoIHead': mmdet.models.roi_heads.sparse_roi_head.SparseRoIHead,'TridentRoIHead': mmdet.models.roi_heads.trident_roi_head.TridentRoIHead}

config_mod.model['bbox_head']

{'type': 'YOLOV3Head','num_classes': 80,'in_channels': [512, 256, 128],'out_channels': [1024, 512, 256],'anchor_generator': {'type': 'YOLOAnchorGenerator','base_sizes': [[(116, 90), (156, 198), (373, 326)],[(30, 61), (62, 45), (59, 119)],[(10, 13), (16, 30), (33, 23)]],'strides': [32, 16, 8]},'bbox_coder': {'type': 'YOLOBBoxCoder'},'featmap_strides': [32, 16, 8],'loss_cls': {'type': 'CrossEntropyLoss','use_sigmoid': True,'loss_weight': 1.0,'reduction': 'sum'},'loss_conf': {'type': 'CrossEntropyLoss','use_sigmoid': True,'loss_weight': 1.0,'reduction': 'sum'},'loss_xy': {'type': 'CrossEntropyLoss','use_sigmoid': True,'loss_weight': 2.0,'reduction': 'sum'},'loss_wh': {'type': 'MSELoss', 'loss_weight': 2.0, 'reduction': 'sum'},'train_cfg': {'assigner': {'type': 'GridAssigner','pos_iou_thr': 0.5,'neg_iou_thr': 0.5,'min_pos_iou': 0}},'test_cfg': {'nms_pre': 1000,'min_bbox_size': 0,'score_thr': 0.05,'conf_thr': 0.005,'nms': {'type': 'nms', 'iou_threshold': 0.45},'max_per_img': 100}}

bbox_head=build_head(config_mod.model['bbox_head'])
print(bbox_head)

YOLOV3Head((loss_cls): CrossEntropyLoss()(loss_conf): CrossEntropyLoss()(loss_xy): CrossEntropyLoss()(loss_wh): MSELoss()(convs_bridge): ModuleList((0): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(1): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(convs_pred): ModuleList((0): Conv2d(1024, 255, kernel_size=(1, 1), stride=(1, 1))(1): Conv2d(512, 255, kernel_size=(1, 1), stride=(1, 1))(2): Conv2d(256, 255, kernel_size=(1, 1), stride=(1, 1)))
)

head=HEADS.get('YOLOV3Head')

head_config = config_mod.model['bbox_head']
print(head_config)

{'type': 'YOLOV3Head', 'num_classes': 80, 'in_channels': [512, 256, 128], 'out_channels': [1024, 512, 256], 'anchor_generator': {'type': 'YOLOAnchorGenerator', 'base_sizes': [[(116, 90), (156, 198), (373, 326)], [(30, 61), (62, 45), (59, 119)], [(10, 13), (16, 30), (33, 23)]], 'strides': [32, 16, 8]}, 'bbox_coder': {'type': 'YOLOBBoxCoder'}, 'featmap_strides': [32, 16, 8], 'loss_cls': {'type': 'CrossEntropyLoss', 'use_sigmoid': True, 'loss_weight': 1.0, 'reduction': 'sum'}, 'loss_conf': {'type': 'CrossEntropyLoss', 'use_sigmoid': True, 'loss_weight': 1.0, 'reduction': 'sum'}, 'loss_xy': {'type': 'CrossEntropyLoss', 'use_sigmoid': True, 'loss_weight': 2.0, 'reduction': 'sum'}, 'loss_wh': {'type': 'MSELoss', 'loss_weight': 2.0, 'reduction': 'sum'}, 'train_cfg': {'assigner': {'type': 'GridAssigner', 'pos_iou_thr': 0.5, 'neg_iou_thr': 0.5, 'min_pos_iou': 0}}, 'test_cfg': {'nms_pre': 1000, 'min_bbox_size': 0, 'score_thr': 0.05, 'conf_thr': 0.005, 'nms': {'type': 'nms', 'iou_threshold': 0.45}, 'max_per_img': 100}}

head_config.pop('type')

'YOLOV3Head'

head(**head_config)

YOLOV3Head((loss_cls): CrossEntropyLoss()(loss_conf): CrossEntropyLoss()(loss_xy): CrossEntropyLoss()(loss_wh): MSELoss()(convs_bridge): ModuleList((0): ConvModule((conv): Conv2d(512, 1024, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(1): ConvModule((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True))(2): ConvModule((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)(bn): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)(activate): LeakyReLU(negative_slope=0.1, inplace=True)))(convs_pred): ModuleList((0): Conv2d(1024, 255, kernel_size=(1, 1), stride=(1, 1))(1): Conv2d(512, 255, kernel_size=(1, 1), stride=(1, 1))(2): Conv2d(256, 255, kernel_size=(1, 1), stride=(1, 1)))
)

3. 构建数据集

from mmdet.datasets import build_dataset,build_dataloader,DATASETS,PIPELINES
DATASETS.module_dict

{'CustomDataset': mmdet.datasets.custom.CustomDataset,'CocoDataset': mmdet.datasets.coco.CocoDataset,'CityscapesDataset': mmdet.datasets.cityscapes.CityscapesDataset,'ConcatDataset': mmdet.datasets.dataset_wrappers.ConcatDataset,'RepeatDataset': mmdet.datasets.dataset_wrappers.RepeatDataset,'ClassBalancedDataset': mmdet.datasets.dataset_wrappers.ClassBalancedDataset,'DeepFashionDataset': mmdet.datasets.deepfashion.DeepFashionDataset,'LVISV05Dataset': mmdet.datasets.lvis.LVISV05Dataset,'LVISDataset': mmdet.datasets.lvis.LVISV05Dataset,'LVISV1Dataset': mmdet.datasets.lvis.LVISV1Dataset,'XMLDataset': mmdet.datasets.xml_style.XMLDataset,'VOCDataset': mmdet.datasets.voc.VOCDataset,'WIDERFaceDataset': mmdet.datasets.wider_face.WIDERFaceDataset}

PIPELINES.module_dict

{'Compose': mmdet.datasets.pipelines.compose.Compose,'AutoAugment': mmdet.datasets.pipelines.auto_augment.AutoAugment,'Shear': mmdet.datasets.pipelines.auto_augment.Shear,'Rotate': mmdet.datasets.pipelines.auto_augment.Rotate,'Translate': mmdet.datasets.pipelines.auto_augment.Translate,'ColorTransform': mmdet.datasets.pipelines.auto_augment.ColorTransform,'EqualizeTransform': mmdet.datasets.pipelines.auto_augment.EqualizeTransform,'BrightnessTransform': mmdet.datasets.pipelines.auto_augment.BrightnessTransform,'ContrastTransform': mmdet.datasets.pipelines.auto_augment.ContrastTransform,'ToTensor': mmdet.datasets.pipelines.formating.ToTensor,'ImageToTensor': mmdet.datasets.pipelines.formating.ImageToTensor,'Transpose': mmdet.datasets.pipelines.formating.Transpose,'ToDataContainer': mmdet.datasets.pipelines.formating.ToDataContainer,'DefaultFormatBundle': mmdet.datasets.pipelines.formating.DefaultFormatBundle,'Collect': mmdet.datasets.pipelines.formating.Collect,'WrapFieldsToLists': mmdet.datasets.pipelines.formating.WrapFieldsToLists,'InstaBoost': mmdet.datasets.pipelines.instaboost.InstaBoost,'LoadImageFromFile': mmdet.datasets.pipelines.loading.LoadImageFromFile,'LoadImageFromWebcam': mmdet.datasets.pipelines.loading.LoadImageFromWebcam,'LoadMultiChannelImageFromFiles': mmdet.datasets.pipelines.loading.LoadMultiChannelImageFromFiles,'LoadAnnotations': mmdet.datasets.pipelines.loading.LoadAnnotations,'LoadProposals': mmdet.datasets.pipelines.loading.LoadProposals,'FilterAnnotations': mmdet.datasets.pipelines.loading.FilterAnnotations,'MultiScaleFlipAug': mmdet.datasets.pipelines.test_time_aug.MultiScaleFlipAug,'Resize': mmdet.datasets.pipelines.transforms.Resize,'RandomFlip': mmdet.datasets.pipelines.transforms.RandomFlip,'Pad': mmdet.datasets.pipelines.transforms.Pad,'Normalize': mmdet.datasets.pipelines.transforms.Normalize,'RandomCrop': mmdet.datasets.pipelines.transforms.RandomCrop,'SegRescale': mmdet.datasets.pipelines.transforms.SegRescale,'PhotoMetricDistortion': mmdet.datasets.pipelines.transforms.PhotoMetricDistortion,'Expand': mmdet.datasets.pipelines.transforms.Expand,'MinIoURandomCrop': mmdet.datasets.pipelines.transforms.MinIoURandomCrop,'Corrupt': mmdet.datasets.pipelines.transforms.Corrupt,'Albu': mmdet.datasets.pipelines.transforms.Albu,'RandomCenterCropPad': mmdet.datasets.pipelines.transforms.RandomCenterCropPad,'CutOut': mmdet.datasets.pipelines.transforms.CutOut,'GtBoxBasedCrop': mmdet.datasets.pipelines.transforms.GtBoxBasedCrop}

config_mod.data['train']

{'type': 'CocoDataset','ann_file': 'F:/datasets/COCO2017/annotations/instances_train2017.json','img_prefix': 'F:/datasets/COCO2017/train2017/','pipeline': [{'type': 'LoadImageFromFile', 'to_float32': True},{'type': 'LoadAnnotations', 'with_bbox': True},{'type': 'PhotoMetricDistortion'},{'type': 'Resize','img_scale': [(320, 320), (608, 608)],'keep_ratio': True}]}

cocodataset=build_dataset(config_mod.data['train'])

loading annotations into memory...
Done (t=42.56s)
creating index...
index created!

dataloader=build_dataloader(cocodataset,1,1)

for batch in dataloader:print(batch)break

mmdet_config_builder_win

在mmdet框架中使用config配置文件构建网络模型

1. 解析yolov3配置文件

2.构建检测器（使用构建函数和hook）

2.1 构建主干网络

2.2 Builder necks

2.3 构建检测头 head

3. 构建数据集

mmdet_config_builder_win相关推荐

最新文章

热门文章