首先要准备一份xml配置的数据集，数据集打完标注文件如下：
ImageSets中暂为空，再执行【划分训练集、验证集和测试集】 ，划分后如下图：
再将xml格式数据集转为txt文件格式【数据集格式转换xml2txt】：
整合成最终数据集：
images是数据集的所有图片；labels是数据集的所有txt标签数据；train.txt是训练集的所有文件名；
数据集放在mm/datasets下， datasets与mmdetection-master代码同级：
这里是将数据集重新组织为 COCO 格式（JSON格式）。官网：在自定义数据集上进行训练.
2.1、COCO数据集格式介绍整体结构
{
"images": [image],
"annotations": [annotation],
"categories": [category]
}
images
images是包含多个image实例的数组 下面是一个image实例：
{
"file_name": "文件名或文件路径",
"height": 360,
"width": 640,
"id": 1 # image id
}
annotations
annotations是包含多个annotation实例的数组 下面是一个annotation实例：
annotation{
"id": int, # 标注id
"image_id": int, # 图片文件id
"category_id": int, # 类别id
# "segmentation": RLE or [polygon], # iscrowded=0 polygon格式 iscrowded=1 RLE格式
"area": float, # 标注区域的面积
"bbox": [x, y, width, height], # bbox标注 xywh
# "iscrowd": 0 or 1, # iscrowd=0 单个的对象 iscrowd=1 一群对象（比如一群人）
}
categories
categories是包含多个categorie实例的数组 下面是一个categorie实例：
{
"id": int, # 类别id
"name": str, # 类别名
# "supercategory": str, # 类别父类 选填
}
2.2、txt2json
【数据集格式转换txt2json】.用这个脚本将数据集格式从txt转为json格式（COCO格式）
转换后的数据集
imagesets到这一步往后其实没什么用了，可以删去也可以保留。
下面修改相关配置文件，这里以faster-rcnn为例：
faster_rcnn_r50_fpn_1x_pest.py
# Config
# 1 model config
model = dict(
type='FasterRCNN',
backbone=dict(
type='ResNet',
depth=50,
num_stages=4,
out_indices=(0, 1, 2, 3),
frozen_stages=1,
norm_cfg=dict(type='BN', requires_grad=True),
norm_eval=True,
style='pytorch'),
neck=dict(
type='FPN',
in_channels=[256, 512, 1024, 2048],
out_channels=256,
num_outs=5),
rpn_head=dict(
type='RPNHead',
in_channels=256,
feat_channels=256,
anchor_generator=dict(
type='AnchorGenerator',
scales=[8],
ratios=[0.5, 1.0, 2.0],
strides=[4, 8, 16, 32, 64]),
bbox_coder=dict(
type='DeltaXYWHBBoxCoder',
target_means=[0.0, 0.0, 0.0, 0.0],
target_stds=[1.0, 1.0, 1.0, 1.0]),
loss_cls=dict(
type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
loss_bbox=dict(type='L1Loss', loss_weight=1.0)),
roi_head=dict(
type='StandardRoIHead',
bbox_roi_extractor=dict(
type='SingleRoIExtractor',
roi_layer=dict(type='RoIAlign', output_size=7, sampling_ratio=0),
out_channels=256,
featmap_strides=[4, 8, 16, 32]),
bbox_head=dict(
type='Shared2FCBBoxHead',
in_channels=256,
fc_out_channels=1024,
roi_feat_size=7,
num_classes=3, # modify 1: dataset class num
bbox_coder=dict(
type='DeltaXYWHBBoxCoder',
target_means=[0.0, 0.0, 0.0, 0.0],
target_stds=[0.1, 0.1, 0.2, 0.2]),
reg_class_agnostic=False,
loss_cls=dict(
type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
loss_bbox=dict(type='L1Loss', loss_weight=1.0))),
train_cfg=dict(
rpn=dict(
assigner=dict(
type='MaxIoUAssigner',
pos_iou_thr=0.7,
neg_iou_thr=0.3,
min_pos_iou=0.3,
match_low_quality=True,
ignore_iof_thr=-1),
sampler=dict(
type='RandomSampler',
num=256,
pos_fraction=0.5,
neg_pos_ub=-1,
add_gt_as_proposals=False),
allowed_border=-1,
pos_weight=-1,
debug=False),
rpn_proposal=dict(
nms_pre=2000,
max_per_img=1000,
nms=dict(type='nms', iou_threshold=0.7),
min_bbox_size=0),
rcnn=dict(
assigner=dict(
type='MaxIoUAssigner',
pos_iou_thr=0.5,
neg_iou_thr=0.5,
min_pos_iou=0.5,
match_low_quality=False,
ignore_iof_thr=-1),
sampler=dict(
type='RandomSampler',
num=512,
pos_fraction=0.25,
neg_pos_ub=-1,
add_gt_as_proposals=True),
pos_weight=-1,
debug=False)),
test_cfg=dict(
rpn=dict(
nms_pre=1000,
max_per_img=1000,
nms=dict(type='nms', iou_threshold=0.7),
min_bbox_size=0),
rcnn=dict(
score_thr=0.05,
nms=dict(type='nms', iou_threshold=0.5),
max_per_img=100)))
# 2 pipeline config
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),
dict(type='Resize', img_scale=(1280, 640), keep_ratio=True), # modify img max size and min size
dict(type='RandomFlip', flip_ratio=0.5),
dict(
type='Normalize',
mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
to_rgb=True),
dict(type='Pad', size_divisor=32),
dict(type='DefaultFormatBundle'),
dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
]
test_pipeline = [
dict(type='LoadImageFromFile'),
dict(
type='MultiScaleFlipAug',
img_scale=(1280, 640), # modify by yourself img max size and min size
flip=False,
transforms=[
dict(type='Resize', keep_ratio=True),
dict(type='RandomFlip'),
dict(
type='Normalize',
mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
to_rgb=True),
dict(type='Pad', size_divisor=32),
dict(type='ImageToTensor', keys=['img']),
dict(type='Collect', keys=['img'])
])
]
# 3 dataset config
classes = ('powdery_mildew', 'leaf_miner', 'anthracnose') # modify 2: dataset classes
data = dict(
samples_per_gpu=2, # batch_size = samples_per_gpu*gpu_num
workers_per_gpu=2, # numworks = workers_per_gpu*gpu_num
train=dict(
type='CocoDataset', # modify 3: dataset type
classes=classes, # modify 2: dataset classes
data_root='../datasets/pest/', # modify 4: dataset root
ann_file='train.json', # modify 5: dataset json annotation
img_prefix='images', # modify 6 dataset image prefix
pipeline=[
dict(type='LoadImageFromFile'),
dict(type='LoadAnnotations', with_bbox=True),
dict(type='Resize', img_scale=(1280, 640), keep_ratio=True), # modify 7: img max size and min size
dict(type='RandomFlip', flip_ratio=0.5),
dict(
type='Normalize',
mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
to_rgb=True),
dict(type='Pad', size_divisor=32),
dict(type='DefaultFormatBundle'),
dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
]),
val=dict(
type='CocoDataset', # same with train
classes=classes, # same with train
data_root='../datasets/pest/', # same with train
ann_file='val.json', # same with train
img_prefix='images/', # same with train
pipeline=[
dict(type='LoadImageFromFile'),
dict(
type='MultiScaleFlipAug',
img_scale=(1280, 1280), # same with train
flip=False,
transforms=[
dict(type='Resize', keep_ratio=True),
dict(type='RandomFlip'),
dict(
type='Normalize',
mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
to_rgb=True),
dict(type='Pad', size_divisor=32),
dict(type='ImageToTensor', keys=['img']),
dict(type='Collect', keys=['img'])
])
]),
test=dict(
type='CocoDataset', # same with train
classes=classes, # same with train
data_root='../datasets/pest/', # same with train
ann_file='test.json', # same with train
img_prefix='images/', # same with train
pipeline=[
dict(type='LoadImageFromFile'),
dict(
type='MultiScaleFlipAug',
img_scale=(1280, 1280), # same with train
flip=False,
transforms=[
dict(type='Resize', keep_ratio=True),
dict(type='RandomFlip'),
dict(
type='Normalize',
mean=[123.675, 116.28, 103.53],
std=[58.395, 57.12, 57.375],
to_rgb=True),
dict(type='Pad', size_divisor=32),
dict(type='ImageToTensor', keys=['img']),
dict(type='Collect', keys=['img'])
])
]))
# 4 other config
# lr default 8 GPU = 0.02 1 GPU = 0.02/8
optimizer = dict(type='SGD', lr=0.02/8, momentum=0.9, weight_decay=0.0001) # modify 8: lr change
optimizer_config = dict(grad_clip=None)
lr_config = dict(
policy='step',
warmup='linear',
warmup_iters=500,
warmup_ratio=0.001,
step=[8, 11])
runner = dict(type='EpochbasedRunner', max_epochs=5)
checkpoint_config = dict(interval=1) # save checkpoint per interval
log_config = dict(interval=5, hooks=[dict(type='TextLoggerHook')]) # print log per 5 interval
custom_hooks = [dict(type='NumClassCheckHook')]
dist_params = dict(backend='nccl')
log_level = 'INFO'
load_from = "checkpoints/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth" # 9: modify checkpoints root
resume_from = None
workflow = [('train_example', 1)]
四、开始训练
文件目录：
单GPU训练指令：
python tools/train.py
--gpus --work_dir
example:
python tools/train.py mycode/train_example/pest/faster_rcnn_r50_fpn_1x_pest.py --work-dir mycode/train_example/work_dir
多GPU训练指令：
tools/dist_train.sh
gpu_num --validate
example:
tools/dist_train.sh mycode/train_example/pest/faster_rcnn_r50_fpn_1x_pest.py 4 --validate
–validate: perform evaluation every k (default=1) epochs during the training.
Referencecsdn：mmdetection训练自己的数据.
github Docs > 2: 在自定义数据集上进行训练.
