Mmdetection: gn result is worse than bn on pascal voc
I train Faster-rcnn:R-50-FPN use bn and gn on voc 07+12, but the gn result (AP 0.5:0.95) is worse than bn, both batch size=2, 1 gpu
| Backbone | AP | AP50 | AP75 |
|:---------------------------------:|:------:|:-------:|:-------:|
| Faster-rcnn:R-50-FPN(bn) | 0.501 | 0.785 | 0.542 |
| Faster-rcnn:R-50-FPN(gn) | 0.474 | 0.790 | 0.498 |
hust-kevin
All 18 comments
Which pretrained weights for backbone are you using?
thangvubk
on 29 Mar 2019
@thangvubk
gn: 'open-mmlab://detectron/resnet50_gn'
bn: 'modelzoo://resnet50'
hust-kevin
on 29 Mar 2019
Can you share the whole config file for GN?
thangvubk
on 29 Mar 2019
# model settings
normalize=dict(type='GN', num_groups=32, frozen=False)
model = dict(
type='FasterRCNN',
pretrained='open-mmlab://detectron/resnet50_gn',
backbone=dict(
type='ResNet',
depth=50,
num_stages=4,
out_indices=(0, 1, 2, 3),
frozen_stages=1,
style='pytorch',
normalize=normalize),
neck=dict(
type='FPN',
in_channels=[256, 512, 1024, 2048],
out_channels=256,
num_outs=5,
normalize=normalize),
rpn_head=dict(
type='RPNHead',
in_channels=256,
feat_channels=256,
anchor_scales=[8],
anchor_ratios=[0.5, 1.0, 2.0],
anchor_strides=[4, 8, 16, 32, 64],
target_means=[.0, .0, .0, .0],
target_stds=[1.0, 1.0, 1.0, 1.0],
use_sigmoid_cls=True),
bbox_roi_extractor=dict(
type='SingleRoIExtractor',
roi_layer=dict(type='RoIAlign', out_size=7, sample_num=2),
out_channels=256,
featmap_strides=[4, 8, 16, 32]),
bbox_head=dict(
type='SharedFCBBoxHead',
num_fcs=2,
in_channels=256,
fc_out_channels=1024,
roi_feat_size=7,
num_classes=21,
target_means=[0., 0., 0., 0.],
target_stds=[0.1, 0.1, 0.2, 0.2],
reg_class_agnostic=False,
normalize=normalize))
# 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,
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,
smoothl1_beta=1 / 9.0,
debug=False),
rcnn=dict(
assigner=dict(
type='MaxIoUAssigner',
pos_iou_thr=0.5,
neg_iou_thr=0.5,
min_pos_iou=0.5,
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=2000,
nms_post=2000,
max_num=2000,
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)
# soft-nms is also supported for rcnn testing
# e.g., nms=dict(type='soft_nms', iou_thr=0.5, min_score=0.05)
)
# dataset settings 和coco不一样
dataset_type = 'VOCDataset'
data_root = '/home/xx/share/dataset/Pascal VOC/VOCdevkit/'
img_norm_cfg = dict(
mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
data = dict(
imgs_per_gpu=2,
workers_per_gpu=2,
train=dict(
type='RepeatDataset', # to avoid reloading datasets frequently
times=3,
dataset=dict(
type=dataset_type,
ann_file=[
data_root + 'VOC2007/ImageSets/Main/trainval.txt',
data_root + 'VOC2012/ImageSets/Main/trainval.txt'
],
img_prefix=[data_root + 'VOC2007/', data_root + 'VOC2012/'],
img_scale=(1000, 600),
img_norm_cfg=img_norm_cfg,
size_divisor=32,
flip_ratio=0.5,
with_mask=False,
with_crowd=True,
with_label=True)),
val=dict(
type=dataset_type,
ann_file=data_root + 'VOC2007/ImageSets/Main/test.txt',
img_prefix=data_root + 'VOC2007/',
img_scale=(1000, 600),
img_norm_cfg=img_norm_cfg,
size_divisor=32,
flip_ratio=0,
with_mask=False,
with_crowd=True,
with_label=True),
test=dict(
type=dataset_type,
ann_file=data_root + 'VOC2007/ImageSets/Main/test.txt',
img_prefix=data_root + 'VOC2007/',
img_scale=(1000, 600),
img_norm_cfg=img_norm_cfg,
size_divisor=32,
flip_ratio=0,
with_mask=False,
with_label=False,
test_mode=True))
# optimizer
optimizer = dict(type='SGD', lr=0.01/8, 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', step=[3]) # actual epoch = 3 * 3 = 9
checkpoint_config = dict(interval=1)
# yapf:disable
log_config = dict(
interval=50,
hooks=[
dict(type='TextLoggerHook'),
# dict(type='TensorboardLoggerHook')
])
# yapf:enable
# runtime settings
total_epochs = 4 # actual epoch = 4 * 3 = 12
dist_params = dict(backend='nccl')
log_level = 'INFO'
work_dir = './work_dirs/faster_rcnn_r50_fpn_gn_1x_voc'
load_from = None
resume_from = None
workflow = [('train', 1)]
Detectron normalize image in a different way from pytorch model zoo.
You may have to change mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True) to mean=[102.9801, 115.9465, 122.7717], std=[1.0, 1.0, 1.0], to_rgb=False)
thangvubk
on 29 Mar 2019
thanks , I will try it again, then I will tell you the result
hust-kevin
on 29 Mar 2019
thanks , I will try it again, then I will tell you the result
hi, How is your result with gn now? I tried and there is no obvious improvement.
donglee-afar
on 1 Apr 2019
@thangvubk @hellock @yhcao6 @donglee-afar
unfortunately，the result is also bad, I think bn/gn is the normalization on hidden layer, so i think the input normalization is not so important ，so the bad result may caused by other
what's more,I try the gn+ws, the result is more bad.
PS:when I train gn+ws, the input normalization I have't changed it, I used
img_norm_cfg = dict(
mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
| Backbone | AP | AP50 | AP75 |
|:---------------------------------:|:------:|:-------:|:-------:|
| Faster-rcnn:R-50-FPN(bn) | 0.501 | 0.785 | 0.542 |
| Faster-rcnn:R-50-FPN(gn) | 0.479 | 0.796 | 0.507 |
| Faster-rcnn:R-50-FPN(gn+ws) | 0.464 | 0.779 | 0.482 |
hust-kevin
on 2 Apr 2019
@hust-kevin I will try to train VOC and let you know the results.
thangvubk
on 3 Apr 2019
ok, Look forward to your reply
hust-kevin
on 3 Apr 2019
I test with provided script on 8GPUs. Here is the results:
VOC_BN: 79.9
VOC_GN: 79.1
They are comparable. In this case you can simply use BN instead of GN.
thangvubk
on 3 Apr 2019
@thangvubk what's the result of ap75 and ap(0.5:0.95) , In my test, AP50 is comparable,but ap75 and ap (0.5:0.95) is worse
hust-kevin
on 3 Apr 2019
I still got comparable AP75:
VOC_BN: 51.4
VOC_GN: 51.1
I notice that you didnt use 4Conv+1FC for boxhead. It might be the reason.
thangvubk
on 3 Apr 2019
what's your config file for GN?
I just use the default config, so the same network for bn and gn
hust-kevin
on 3 Apr 2019
# model settings
normalize = dict(type='GN', num_groups=32, frozen=False)
model = dict(
type='FasterRCNN',
pretrained='open-mmlab://detectron/resnet50_gn',
backbone=dict(
type='ResNet',
depth=50,
num_stages=4,
out_indices=(0, 1, 2, 3),
frozen_stages=1,
style='pytorch',
normalize=normalize),
neck=dict(
type='FPN',
in_channels=[256, 512, 1024, 2048],
out_channels=256,
num_outs=5,
normalize=normalize),
rpn_head=dict(
type='RPNHead',
in_channels=256,
feat_channels=256,
anchor_scales=[8],
anchor_ratios=[0.5, 1.0, 2.0],
anchor_strides=[4, 8, 16, 32, 64],
target_means=[.0, .0, .0, .0],
target_stds=[1.0, 1.0, 1.0, 1.0],
use_sigmoid_cls=True),
bbox_roi_extractor=dict(
type='SingleRoIExtractor',
roi_layer=dict(type='RoIAlign', out_size=7, sample_num=2),
out_channels=256,
featmap_strides=[4, 8, 16, 32]),
bbox_head=dict(
type='ConvFCBBoxHead',
num_shared_convs=4,
num_shared_fcs=1,
in_channels=256,
conv_out_channels=256,
fc_out_channels=1024,
roi_feat_size=7,
num_classes=21,
target_means=[0., 0., 0., 0.],
target_stds=[0.1, 0.1, 0.2, 0.2],
reg_class_agnostic=False,
normalize=normalize))
# 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,
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,
smoothl1_beta=1 / 9.0,
debug=False),
rcnn=dict(
assigner=dict(
type='MaxIoUAssigner',
pos_iou_thr=0.5,
neg_iou_thr=0.5,
min_pos_iou=0.5,
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=2000,
nms_post=2000,
max_num=2000,
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)
# soft-nms is also supported for rcnn testing
# e.g., nms=dict(type='soft_nms', iou_thr=0.5, min_score=0.05)
)
# dataset settings
dataset_type = 'VOCDataset'
data_root = 'data/VOCdevkit/'
img_norm_cfg = dict(
mean=[102.9801, 115.9465, 122.7717], std=[1.0, 1.0, 1.0], to_rgb=False)
data = dict(
imgs_per_gpu=2,
workers_per_gpu=2,
train=dict(
type='RepeatDataset', # to avoid reloading datasets frequently
times=3,
dataset=dict(
type=dataset_type,
ann_file=[
data_root + 'VOC2007/ImageSets/Main/trainval.txt',
data_root + 'VOC2012/ImageSets/Main/trainval.txt'
],
img_prefix=[data_root + 'VOC2007/', data_root + 'VOC2012/'],
img_scale=(1000, 600),
img_norm_cfg=img_norm_cfg,
size_divisor=32,
flip_ratio=0.5,
with_mask=False,
with_crowd=True,
with_label=True)),
val=dict(
type=dataset_type,
ann_file=data_root + 'VOC2007/ImageSets/Main/test.txt',
img_prefix=data_root + 'VOC2007/',
img_scale=(1000, 600),
img_norm_cfg=img_norm_cfg,
size_divisor=32,
flip_ratio=0,
with_mask=False,
with_crowd=True,
with_label=True),
test=dict(
type=dataset_type,
ann_file=data_root + 'VOC2007/ImageSets/Main/test.txt',
img_prefix=data_root + 'VOC2007/',
img_scale=(1000, 600),
img_norm_cfg=img_norm_cfg,
size_divisor=32,
flip_ratio=0,
with_mask=False,
with_label=False,
test_mode=True))
# 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', step=[3]) # actual epoch = 3 * 3 = 9
checkpoint_config = dict(interval=1)
# yapf:disable
log_config = dict(
interval=50,
hooks=[
dict(type='TextLoggerHook'),
# dict(type='TensorboardLoggerHook')
])
# yapf:enable
# runtime settings
total_epochs = 4 # actual epoch = 4 * 3 = 12
dist_params = dict(backend='nccl')
log_level = 'INFO'
work_dir = './work_dirs/faster_rcnn_r50_fpn_1x_voc0712_gn'
load_from = None
resume_from = None
workflow = [('train', 1)]
yes, I read the gn paper again, maybe you are right
thanks for your patient reply !!
hust-kevin
on 3 Apr 2019
yes, I read the gn paper again, maybe you are right
thanks for your patient reply !!
我也打算实现一下gn+ws，能加个qq交流一下吗，都是关山口的校友，qq：1955662900，向大佬学习！
donglee-afar
on 6 Apr 2019
I still got comparable AP75:
VOC_BN: 51.4
VOC_GN: 51.1
I notice that you didnt use 4Conv+1FC for boxhead. It might be the reason.
why must use the 4Conv+1FC for boxhead in GN? but using SharedFCBBoxHead in BN, thank your @thangvubk
breezelj
on 4 Apr 2020
Related issues
yangcong955
·
3Comments
fatLime
·
3Comments
michaelisc
·
3Comments
Hemantr05
·
3Comments
liugaolian
·
3Comments
Most helpful comment