pytorch 华为云杯垃圾分类总结（目标检测）
摘要
我们队伍名称冲冲冲，A榜第八，B榜第11，决赛榜第16，决赛服务器评分出现问题和恰巧我们的模型速度在增加了200多张照片的情况下又超时了，直接就出局了。接下来就分享一些常用的提分点。数据下载链接链接：https://pan.baidu.com/s/1aCh_fIVsRBjKXQkFIpdvRQ
提取码：1234
baseline
前二十基本全是mmdetection。竞赛必备，里面包含了很多算法，目前2.0版本主要使用cascade——rcnn系列，
第一名使用的是res2net前置网络，当时我们并没有使用这个网络，res2net的基础多分就有70分，我们使用的cascade——rfp的resnet50作为前置网络，基础得分只有67左右，基础分就低了很多。使用senet154基础分有75.不同的前置网络分数差距很大，需要在训练前都测试，找出最好的baseline，使用senet154的前置得分很高但是在进行提分就困难很多了，第四名选手说多尺度加上都没涨点。很多策略使用之后都不会涨点，上分变的困难了一些。
所以学会使用mmdetection，和能够灵活运用前置网络是重点。找到适合的网络就在起步领先其他人。后面再加上trick也不会那么费力了。
trick
目前常用的mixup和mosic。和mmdet里面使用albu自带的各种增强需要自己去组合使用。不同的trick可能会有不同的影响，比如我最进训练mixup和cutout同时使用效果大幅度降低。单独测试都可以提升2个点。还有类别平衡，标签平滑。其中类别平衡在mmdet里面自带一种和自己进行线下增强，将类别少的进行随机变换扩充。第一名的代码里面的使用的有atss特征提取。是结合了atss下面的代码自己构造的新网络。需要对mmdet的代码编写十分熟悉才行。mmdet本身就将各个网络写的组件形式，所以需要花很大功夫去学习源码。最终要的一点是机器配置。有些trick加上去很难收敛。所以有时候需要你去预训练权重在coco数据集上（玩不起）。或者使用学习率退火一般来说也会更好一点，或者mmdet自带的分布式训练。
部分代码
模型修剪，上传文件只需要参数信息
import torch
yuan=torch.load('./rs_cut_mix_pafpn_box.pth')
new = {'meta': yuan['meta'],'state_dict': yuan['state_dict']}
torch.save(new,'./rs_cu4_4_min.pth')
coco_detection
# dataset settings
dataset_type = 'VOCDataset'
data_root = '/home/jmy/hjc/code/rubbish_classification/datasets/VOCdevkit/'
img_norm_cfg = dict(
mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
# TODO: augmentation from github
albu_train_transforms = [
dict(
type='MotionBlur',
blur_limit=(3, 7),
p=0.2),
#add two albu
dict(
type='ShiftScaleRotate',
shift_limit=0.0625,
scale_limit=0.0,
rotate_limit=[-10, 10],
interpolation=1,
p=0.5),
]
train_pipeline = [
dict(type='LoadImageFromFile'),
dict(type='LoadAnnotations', with_bbox=True),
# TODO: augmentation from github
dict(type='Mixup', prob=0.4, lambd=0.5, mixup=True,
trainval_path='/home/jmy/hjc/code/rubbish_classification/mmdetection/augmentation_zx/ap_05_id_in_all.txt',
#trainval_path='/home/jmy/hjc/code/rubbish_classification/mmdetection/augmentation_zx/ap_05_id_in_trainset.txt',
img_path='/home/jmy/hjc/code/rubbish_classification/datasets/VOCdevkit/VOC2007/JPEGImages',
annotation_path='/home/jmy/hjc/code/rubbish_classification/datasets/VOCdevkit/VOC2007/Annotations'),
#dict(type='Resize', img_scale=(800, 480), keep_ratio=True),
dict(type='Resize', img_scale=[(800, 600), (800, 360)], keep_ratio=True, multiscale_mode='range'),
dict(type='RandomFlip', flip_ratio=0.5),
dict(type='Normalize', **img_norm_cfg),
dict(type='Pad', size_divisor=32),
# TODO: augmentation from github
dict(
type='Albu',
transforms=albu_train_transforms,
bbox_params=dict(
type='BboxParams',
format='pascal_voc',
label_fields=['gt_labels'],
min_visibility=0.0,
filter_lost_elements=True),
keymap={
'img': 'image',
'gt_masks': 'masks',
'gt_bboxes': 'bboxes'
},
update_pad_shape=False,
skip_img_without_anno=True),
dict(type='DefaultFormatBundle'),
dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']),
]
test_pipeline = [
dict(type='LoadImageFromFile'),
dict(
type='MultiScaleFlipAug',
#img_scale=(1000, 600),
img_scale=(800, 480),
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(
#The default learning rate in config files is for 8 GPUs and 2 img/gpu (batch size = 8*2 = 16)
#e.g., lr=0.01 for 4 GPUs * 2 img/gpu and lr=0.08 for 16 GPUs * 4 img/gpu.
# samples_per_gpu=2,
# workers_per_gpu=2,
samples_per_gpu=4,
workers_per_gpu=0,
train=dict(
type='RepeatDataset',
#hjc:The VOC dataset uses 3 times the size of dataset during training
#times=3,
times=1,
dataset=dict(
type=dataset_type,
# ann_file=[
# data_root + 'VOC2007/ImageSets/Main/trainval.txt',
# #data_root + 'VOC2012/ImageSets/Main/trainval.txt'
# ],
ann_file=[
data_root + 'VOC2007/ImageSets/Main/train.txt'
#'/home/jmy/hjc/code/rubbish_classification/mmdetection/data/lowap2000_grid/VOCdevkit/VOC2007/ImageSets/Main/train.txt'
],
img_prefix=[data_root + 'VOC2007/', data_root + 'VOC2012/'],
#img_prefix=[data_root + 'VOC2007/', '/home/jmy/hjc/code/rubbish_classification/mmdetection/data/lowap2000_grid/VOCdevkit/VOC2007'],
pipeline=train_pipeline)),
val=dict(
type=dataset_type,
# ann_file=data_root + 'VOC2007/ImageSets/Main/test.txt',
ann_file=data_root + 'VOC2007/ImageSets/Main/val.txt',
img_prefix=data_root + 'VOC2007/',
pipeline=test_pipeline),
test=dict(
type=dataset_type,
#ann_file=data_root + 'VOC2007/ImageSets/Main/trainval.txt',
ann_file=data_root + 'VOC2007/ImageSets/Main/val.txt',
img_prefix=data_root + 'VOC2007/',
pipeline=test_pipeline))
evaluation = dict(interval=1, metric='mAP')
cascade_rcnn_r50_fpn
# model settings
model = dict(
type='CascadeRCNN',
pretrained='torchvision://resnet50',
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],
scales=[7],
ratios=[0.5, 1.0, 2.0],
#ratios=[0.2, 0.4, 0.5, 0.6, 0.75, 17/20, 1.0, 20/17, 4/3, 5/3, 2.0, 2.5, 5.0],
strides=[4, 8, 16, 32, 64]),
bbox_coder=dict(
type='DeltaXYWHBBoxCoder',
target_means=[.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='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=1.0)),
roi_head=dict(
type='CascadeRoIHead',
num_stages=3,
stage_loss_weights=[1, 0.5, 0.25],
bbox_roi_extractor=dict(
type='SingleRoIExtractor',
roi_layer=dict(type='RoIAlign', out_size=7, sample_num=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=80,
#V2.0 donot need N + 1 classes count. just N
num_classes=44,
bbox_coder=dict(
type='DeltaXYWHBBoxCoder',
target_means=[0., 0., 0., 0.],
target_stds=[0.1, 0.1, 0.2, 0.2]),
reg_class_agnostic=True,
loss_cls=dict(
type='CrossEntropyLoss',
use_sigmoid=False,
loss_weight=1.0),
loss_bbox=dict(type='SmoothL1Loss', beta=1.0,
loss_weight=1.0)),
dict(
type='Shared2FCBBoxHead',
in_channels=256,
fc_out_channels=1024,
roi_feat_size=7,
#num_classes=80,
num_classes=44,
bbox_coder=dict(
type='DeltaXYWHBBoxCoder',
target_means=[0., 0., 0., 0.],
target_stds=[0.05, 0.05, 0.1, 0.1]),
reg_class_agnostic=True,
loss_cls=dict(
type='CrossEntropyLoss',
use_sigmoid=False,
loss_weight=1.0),
loss_bbox=dict(type='SmoothL1Loss', beta=1.0,
loss_weight=1.0)),
dict(
type='Shared2FCBBoxHead',
in_channels=256,
fc_out_channels=1024,
roi_feat_size=7,
#num_classes=80,
num_classes=44,
bbox_coder=dict(
type='DeltaXYWHBBoxCoder',
target_means=[0., 0., 0., 0.],
target_stds=[0.033, 0.033, 0.067, 0.067]),
reg_class_agnostic=True,
loss_cls=dict(
type='CrossEntropyLoss',
use_sigmoid=False,
loss_weight=1.0),
loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0))
]))
# model training and testing settings
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=0,
pos_weight=-1,
debug=False),
rpn_proposal=dict(
nms_across_levels=False,
nms_pre=2000,
nms_post=2000,
max_num=2000,
nms_thr=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),
dict(
assigner=dict(
type='MaxIoUAssigner',
pos_iou_thr=0.6,
neg_iou_thr=0.6,
min_pos_iou=0.6,
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),
dict(
assigner=dict(
type='MaxIoUAssigner',
pos_iou_thr=0.7,
neg_iou_thr=0.7,
min_pos_iou=0.7,
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_across_levels=False,
nms_pre=1000,
nms_post=1000,
max_num=1000,
nms_thr=0.7,
min_bbox_size=0),
rcnn=dict(
score_thr=0.05, nms=dict(type='nms', iou_thr=0.5), max_per_img=100))
#score_thr=0.001, nms=dict(type='nms', iou_thr=0.5), max_per_img=100))
schedule_lr
# optimizer
#optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0001)
##note!:lr = 0.00125*batch_size
optimizer = dict(type='SGD', lr=0.005, momentum=0.9, weight_decay=0.0001)
# optimizer_config = dict(grad_clip=None)
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_iters=4000,
warmup_ratio=0.001,
#step=[8, 11])
step=[9, 12])
total_epochs = 14
default_runtime
checkpoint_config = dict(interval=1)
# yapf:disable
log_config = dict(
#interval=50,
interval=50,
hooks=[
dict(type='TextLoggerHook'),
# dict(type='TensorboardLoggerHook')
])
# yapf:enable
dist_params = dict(backend='nccl')
log_level = 'INFO'
#load_from = None
#load_from = '/home/jmy/hjc/code/rubbish_classification/mmdetection/checkpoints/cascade_rcnn/cascade_rcnn_r50_fpn_1x_coco_20200316-3dc56deb.pth'
#load_from = '/home/jmy/hjc/code/rubbish_classification/mmdetection/checkpoints/cascade_rcnn/cascade_rcnn_r50_coco_pretrained_weights_classes_45.pth'
#load_from = '/home/jmy/hjc/code/rubbish_classification/mmdetection/checkpoints/cascade_rcnn/cascade_rcnn_r101_coco_pretrained_weights_classes_45.pth'
load_from = '/home/jmy/hjc/code/rubbish_classification/mmdetection/checkpoints/cascade_rcnn/cascade_rcnn_x101_32x4d_coco_pretrained_weights_classes_45.pth'
resume_from = None
workflow = [('train', 1)]
总结
这是前期部分代码，主要是作为一个新手学习点。更多的策略合数据增强都是自己去mmdet下自己书写代码。比赛最中要的是有足够好的设备。和掌握大部分的mmdet的源码即可，就我们队伍而言。就有10张卡差不多。排名前几的，有的设备是64张卡。或者8张32显存的v100，足够支撑训练coco预训练权重的，一个模型三天即可。更多的还是要自己学会改代码。利用mmdet提供的组件进行修改。可以自己观看configs下的全部文件进行学习。