1. 环境搭建2. 修改配置文件3. 下载权重4. 自定义数据集转COCO格式4. 训练&测试5.实验结果6. 遇到的Bug
1. 环境搭建论文链接：https://arxiv.org/abs/2003.10152
代码链接：https://github.com/WXinlong/SOLO
除了mmcv单独安装，其他用:
pip3 install requirements.txt
安装mmcv：solo的代码用mmdet v1.0.0,搭配的是mmcv 0.2.16（有人用将mmdet替换成openlab-mmdetection官网最新mmcv，希望这样可以更换mmcv为最新的，是不可行的！所以这样就必须用mmcv 0.2.16了）
用下面的方式会报错：
pip install mmcv-full==0.2.16 -f https://download.openmmlab.com/mmcv/dist/cu111/torch1.9.0/index.html
所以采用先下载mmcv 0.2.16代码，再编译：
git clone -b v0.2.16 https://github.com/open-mmlab/mmcv.gitcd mmcvpip install -e . 把代码加入环境变量：(**是自己的文件路径)
export PYTHonPATH=$PYTHONPATH:**/SOLO-master 2. 修改配置文件
SOLO-master/mmdet/core/evaluation/class_names.py
def coco_classes():
return [
'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train',
'truck', 'boat', 'traffic_light', 'fire_hydrant', 'stop_sign',
'parking_meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep',
'cow', 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella',
'handbag', 'tie', 'suitcase', 'frisbee', 'skis', 'snowboard',
'sports_ball', 'kite', 'baseball_bat', 'baseball_glove', 'skateboard',
'surfboard', 'tennis_racket', 'bottle', 'wine_glass', 'cup', 'fork',
'knife', 'spoon', 'bowl', 'banana', 'apple', 'sandwich', 'orange',
'broccoli', 'carrot', 'hot_dog', 'pizza', 'donut', 'cake', 'chair',
'couch', 'potted_plant', 'bed', 'dining_table', 'toilet', 'tv',
'laptop', 'mouse', 'remote', 'keyboard', 'cell_phone', 'microwave',
'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase',
'scissors', 'teddy_bear', 'hair_drier', 'toothbrush'
]
修改为自己的类别：
def coco_classes():
return [
'park',
]
SOLO-master/mmdet/datasets/coco.py
@DATASETS.register_module
class CocoDataset(CustomDataset):
CLASSES = ('person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus',
'train', 'truck', 'boat', 'traffic_light', 'fire_hydrant',
'stop_sign', 'parking_meter', 'bench', 'bird', 'cat', 'dog',
'horse', 'sheep', 'cow', 'elephant', 'bear', 'zebra', 'giraffe',
'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee',
'skis', 'snowboard', 'sports_ball', 'kite', 'baseball_bat',
'baseball_glove', 'skateboard', 'surfboard', 'tennis_racket',
'bottle', 'wine_glass', 'cup', 'fork', 'knife', 'spoon', 'bowl',
'banana', 'apple', 'sandwich', 'orange', 'broccoli', 'carrot',
'hot_dog', 'pizza', 'donut', 'cake', 'chair', 'couch',
'potted_plant', 'bed', 'dining_table', 'toilet', 'tv', 'laptop',
'mouse', 'remote', 'keyboard', 'cell_phone', 'microwave',
'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock',
'vase', 'scissors', 'teddy_bear', 'hair_drier', 'toothbrush')
修改为：
@DATASETS.register_module
class CocoDataset(CustomDataset):
CLASSES = ('park',)
SOLO-master/configs/solov2/solov2_x101_dcn_fpn_8gpu_3x.py：
修改权重resnext101_64x4d-ee2c6f71.pth 和SOLOv2_X101_DCN_3x.pth路径model中bbox_head的num_classes改为自己分类数+1（背景）model中mask_feat_head= w*h*s^2修改train、val、test的ann_file和img_prefix路径添加evaluation = dict(interval=1)：每1个epoch对val数据测试（但是在验证阶段，显存会增加，a100，40G在训练时用的32G，验证时直接爆了，我加了with torch.no_grad():，在验证阶段不反向传播，还是爆了，有知道怎么弄了评论区告诉我一下！），代码中**是自己的路径修改文件保存路径work_dir
# model settings
model = dict(
type='SOLOv2',
# pretrained='open-mmlab://resnext101_64x4d',
pretrained='pre_model/resnext101_64x4d-ee2c6f71.pth', # 改为自己的权重路径
backbone=dict(
type='ResNeXt',
depth=101,
groups=64,
base_width=4,
num_stages=4,
out_indices=(0, 1, 2, 3),
frozen_stages=1,
style='pytorch',
dcn=dict(
type='DCNv2',
deformable_groups=1,
fallback_on_stride=False),
stage_with_dcn=(False, True, True, True)),
neck=dict(
type='FPN',
in_channels=[256, 512, 1024, 2048],
out_channels=256,
start_level=0,
num_outs=5),
bbox_head=dict(
type='SOLOv2Head',
# num_classes=81,
num_classes=2,
in_channels=256,
stacked_convs=4,
use_dcn_in_tower=True,
type_dcn='DCNv2',
seg_feat_channels=512,
strides=[8, 8, 16, 32, 32],
scale_ranges=((1, 96), (48, 192), (96, 384), (192, 768), (384, 2048)),
sigma=0.2,
num_grids=[40, 36, 24, 16, 12],
ins_out_channels=256,
loss_ins=dict(
type='DiceLoss',
use_sigmoid=True,
loss_weight=3.0),
loss_cate=dict(
type='FocalLoss',
use_sigmoid=True,
gamma=2.0,
alpha=0.25,
loss_weight=1.0)),
mask_feat_head=dict(
type='MaskFeatHead',
in_channels=256,
out_channels=128,
start_level=0,
end_level=3,
num_classes=256,
# num_classes=1,
conv_cfg=dict(type='DCNv2'),
norm_cfg=dict(type='GN', num_groups=32, requires_grad=True)),
)
# training and testing settings
train_cfg = dict()
test_cfg = dict(
nms_pre=500,
score_thr=0.1,
mask_thr=0.5,
update_thr=0.05,
kernel='gaussian', # gaussian/linear
sigma=2.0,
max_per_img=100)
# dataset settings
dataset_type = 'CocoDataset'
# data_root = '/home/wenjie.yuan/Cow/cow_baseline/cow_datasets/train_dataset/200/'
data_root = '/home/wenjie.yuan/Ins/Zero-shot-Instance-Segmentation-main/data/coco/'
img_norm_cfg = dict(
mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
train_pipeline = [
dict(type='LoadImageFromFile'),
dict(type='LoadAnnotations', with_bbox=True, with_mask=True),
dict(type='Resize',
img_scale=[(1333, 800), (1333, 768), (1333, 736),
(1333, 704), (1333, 672), (1333, 640)],
# img_scale=[(4096, 800), (4096, 1000)],
multiscale_mode='value',
keep_ratio=True),
dict(type='RandomFlip', flip_ratio=0.5),
dict(type='Normalize', **img_norm_cfg),
dict(type='Pad', size_divisor=32),
dict(type='DefaultFormatBundle'),
dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_masks']),
]
test_pipeline = [
dict(type='LoadImageFromFile'),
dict(
type='MultiScaleFlipAug',
img_scale=(1333, 800),
# img_scale=(4096, 800),
# img_scale=[(4096, 800), (4096, 900),(4096, 1000)],
flip=False,
transforms=[
dict(type='Resize', keep_ratio=True),
dict(type='RandomFlip'),
dict(type='Normalize', **img_norm_cfg),
dict(type='Pad', size_divisor=32),
dict(type='ImageToTensor', keys=['img']),
dict(type='Collect', keys=['img']),
])
]
data = dict(
imgs_per_gpu=8,
workers_per_gpu=0,
train=dict(
type=dataset_type,
ann_file='**/SOLO-master/train_flfr_200000_0318.json',
img_prefix='**/dataset/flfr/images',
pipeline=train_pipeline),
val=dict(
type=dataset_type,
ann_file='**/SOLO-master/val_flfr_2500_0318.json',
img_prefix='**/dataset/flfr/images',
pipeline=test_pipeline),
test=dict(
type=dataset_type,
ann_file='**/SOLO-master/val_flfr_2500_0318.json',
img_prefix='**/dataset/flfr/images',
pipeline=test_pipeline))
# evaluation = dict(metric=['bbox','segm'], interval=1)
evaluation = dict(interval=1)
# optimizer
optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0001)
optimizer_config = dict(grad_clip=dict(max_norm=35, norm_type=2))
# learning policy
lr_config = dict(
policy='step',
warmup='linear',
warmup_iters=500,
warmup_ratio=0.01,
step=[27, 33])
checkpoint_config = dict(interval=12)
# yapf:disable
log_config = dict(
# interval=50,
interval=10,
hooks=[
dict(type='TextLoggerHook'),
# dict(type='TensorboardLoggerHook')
])
# yapf:enable
# runtime settings
total_epochs = 36
device_ids = range(8)
dist_params = dict(backend='nccl')
log_level = 'INFO'
work_dir = './work_dirs/solov2_release_x101_dcn_fpn_8gpu_3x_batch8_train200000'
load_from = './pre_model/SOLOv2_X101_DCN_3x.pth'
resume_from = None
workflow = [('train', 1)]
3. 下载权重
resnext101_64x4d-ee2c6f71.pth
官网没找到，从一篇博客中找到的：resnext101_64x4d-ee2c6f71.pthSOLOv2_X101_DCN_3x
去官网https://github.com/WXinlong/SOLO下载 4. 自定义数据集转COCO格式
参考链接：https://zhuanlan.zhihu.com/p/56881826
我的数据格式：
array类型，里面每个元素是一张图片每个图片里面所有segmentation放在一个array里面，bbox也一样，按照coco格式，是要
coco格式注意点：
image的id和annotation的image_id号一致annotation的id在不同图片里面也要不一样，即每个bbox的ground true必须不一样，我用i*100+j把每张图片的id区分开（即每个图片里bbox不超过100即可，我试了20是还会重复的），特别重要！因为一样的情况代码mmdet v1.0.0代码没有提示错误，会继续运行，但是运行时间很短，一定要注意！
coco的image和annotation每个元素代表什么如下：
image{
"id" : int, # 图像id，可从0开始
"width" : int, # 图像的宽
"height" : int, # 图像的高
"file_name" : str, # 文件名
"license" : int, # 遵循哪个协议
"flickr_url" : str, # flickr图片链接url
"coco_url" : str, # COCO图片链接url
"date_captured" : datetime, # 获取数据的日期
}
annotation{
"id" : int, # 注释id编号
"image_id" : int, # 图像id编号
"category_id" : int, # 类别id编号
"segmentation" : RLE or [polygon], # 分割具体数据
"area" : float, # 目标检测的区域大小
"bbox" : [x,y,width,height], # 目标检测框的坐标详细位置信息
"iscrowd" : 0 or 1, # 目标是否被遮盖，默认为0
}
categories[{
"id" : int, # 类别id编号
"name" : str, # 类别名字
"supercategory" : str, # 类别所属的大类，如哈巴狗和狐狸犬都属于犬科这个大类
}]
改写后的代码如下：
# *_* : coding: utf-8 *_*
'''
datasets process for object detection project.
for convert customer dataset format to coco data format,
'''
import traceback
import argparse
import datetime
import json
import cv2
import os
import numpy as np
__CLASS__ = ['__background__', 'park'] # class dictionary, background must be in first index.
def argparser():
parser = argparse.ArgumentParser("define argument parser for pycococreator!")
parser.add_argument("-r", "--root_path", default="dataset/flfr", help="path of root directory")
parser.add_argument("-p", "--json_path", default='SOLO-master/train_flfr_0105.json', help="datasets path ")
parser.add_argument("-s", "--save_path", default='SOLO-master/train_flfr_10000_0317.json', help="save path ")
return parser.parse_args()
def MainProcessing(args):
'''main process source code.'''
annotations = {} # annotations dictionary, which will dump to json format file.
root_path = args.root_path
json_path = args.json_path
save_path = args.save_path
# coco annotations info.
annotations["info"] = {
"description": "customer dataset format convert to COCO format",
"url": "/share/parking/pony.pan/dataset/flfr",
"version": "1.0",
"year": 2022,
"contributor": "wenjie.yuan",
"date_created": "2022/03/17"
}
# coco annotations licenses.
annotations["licenses"] = [{
"url": "https://www.apache.org/licenses/LICENSE-2.0.html",
"id": 1,
"name": "Apache License 2.0"
}]
# coco annotations categories.
annotations["categories"] = []
for cls, clsname in enumerate(__CLASS__):
if clsname == '__background__':
continue
annotations["categories"].append(
{
"supercategory": "object",
"id": cls,
"name": clsname
}
)
annotations["images"] = []
annotations["annotations"] = []
with open(json_path, 'r') as f:
json_ = json.load(f)
# for i in range(len(json_)):
# for i in range(10000):
for i in range(200000):
tmp = json_[i]
id = i
annotations["images"].append(
{
"license": 1,
"file_name":tmp['filename'].split('/')[-1],
"coco_url": "",
"height": tmp['height'],
"width": tmp['width'],
"date_captured": datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S'),
"flickr_url": "",
"id": id
}
)
for j in range(len(tmp['ann']['kpts'])):
annotations["annotations"].append(
{
"id": i*100+j,
"image_id": id,
# "category_id": cls+1,
"category_id": 1,
# "segmentation": np.array(tmp['ann']['kpts'][j])[[True,True, False,True,True, False,True,True, False,True,True, False]].tolist(),
"segmentation": [[tmp['ann']['kpts'][j][0], tmp['ann']['kpts'][j][1], tmp['ann']['kpts'][j][3], tmp['ann']['kpts'][j][4], tmp['ann']['kpts'][j][6], tmp['ann']['kpts'][j][7], tmp['ann']['kpts'][j][9], tmp['ann']['kpts'][j][10]]],
"area": tmp['ann']['areas'][j],
"bbox": tmp['ann']['bboxes'][j],
"iscrowd": 0,
}
)
with open(save_path, "w") as f:
json.dump(annotations, f)
if __name__ == "__main__":
print("begining to convert customer format to coco format!")
args = argparser()
try:
MainProcessing(args)
except Exception as e:
traceback.print_exc()
print("successful to convert customer format to coco format")
4. 训练&测试
8卡训练：
bash tools/dist_train.sh solov2_x101_dcn_fpn_8gpu_3x.py 8 训练过程中验证：
bash tools/dist_train.sh solov2_x101_dcn_fpn_8gpu_3x.py 8 --validate 预测并可视化python tools/test_ins_vis.py solov2_x101_dcn_fpn_8gpu_3x.py work_dirs/solov2_release_x101_dcn_fpn_8gpu_3x_batch8_train10000_job/epoch_36.pth --show --save_dir ./out_file8卡评估（单卡太慢了）：
./tools/dist_test.sh solov2_x101_dcn_fpn_8gpu_3x.py work_dirs/solov2_release_x101_dcn_fpn_8gpu_3x/epoch_36.pth 8 --show --out results_solo.pkl --eval segm 5.实验结果
AssertionError: MMCV==0.2.16 is used but incompatible. Please install mmcv>=1.3.8, <=1.4.0.
原因：不用安装mmcv 1.3.8,本代码就是用的MMCV==0.2.16，是因为环境变量没加进去的原因办法：export PYTHonPATH=$PYTHONPATH:**/SOLO-master
