Skip to content
Snippets Groups Projects
Commit 27c917ab authored by wanggh's avatar wanggh
Browse files

Merge branch 'master' into gpu3

parents f9e77805 29efbf63
No related branches found
No related tags found
1 merge request!8add RPN FM
Hide whitespace changes
Inline Side-by-side
Showing
with 832 additions and 14 deletions
configs/distillers/DeFeat/DeFeat_deneck_decls_faster_rcnn_r152_fpn_1x_distill_faster_rcnn_r50_fpn_1x_coco.py 0 → 100644
+ 99
− 0
View file @ 27c917ab
_base_ = [
'../../_base_/models/faster_rcnn_r50_fpn.py',
'../../_base_/datasets/coco_detection.py',
'../../_base_/schedules/schedule_1x.py', '../../_base_/default_runtime.py'
]
# model settings
# model settings
find_unused_parameters=True
pos_w=0.5
neg_w=2
roi_pos_w=0.05
roi_pos_T=3
roi_neg_w=2
roi_neg_T=1
distiller = dict(
type='DeFeat',
teacher_pretrained = '/mnt/data3/wangguohua/model/mmdet/faster_rcnn/faster_rcnn_r152_fpn_1x_coco.pth',
load_teacher_part = 'neck_head',
distill_cfg = [ dict(feature_level = 0,
methods=[dict(type='MSELoss',
name='loss_p_0',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_0',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
dict(feature_level = 1,
methods=[dict(type='MSELoss',
name='loss_p_1',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_1',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
dict(feature_level = 2,
methods=[dict(type='MSELoss',
name='loss_p_2',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_2',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
dict(feature_level = 3,
methods=[dict(type='MSELoss',
name='loss_p_3',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_3',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
dict(student_module = 'roi_head.bbox_head.fc_cls',
teacher_module = 'roi_head.bbox_head.fc_cls',
register_hook = True,
methods=[dict(type='KLwithT',
name='loss_roi_p',
T=roi_pos_T,
weight = roi_pos_w,
),
dict(type='KLwithT',
name='loss_roi_n',
T=roi_neg_T,
weight = roi_neg_w,
),
]
),
]
)
student_cfg = 'configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py'
teacher_cfg = 'configs/faster_rcnn/faster_rcnn_r152_fpn_1x_coco.py'
configs/distillers/DeFeat/DeFeat_deneck_faster_rcnn_r152_fpn_1x_distill_faster_rcnn_r50_fpn_1x_coco.py 0 → 100644
+ 80
− 0
View file @ 27c917ab
_base_ = [
'../../_base_/models/faster_rcnn_r50_fpn.py',
'../../_base_/datasets/coco_detection.py',
'../../_base_/schedules/schedule_1x.py', '../../_base_/default_runtime.py'
]
# model settings
# model settings
find_unused_parameters=True
pos_w=0.5
neg_w=2
distiller = dict(
type='DeFeat',
teacher_pretrained = '/mnt/data3/wangguohua/model/mmdet/faster_rcnn/faster_rcnn_r152_fpn_1x_coco.pth',
load_teacher_part = '',
distill_cfg = [ dict(feature_level = 0,
methods=[dict(type='MSELoss',
name='loss_p_0',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_0',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
dict(feature_level = 1,
methods=[dict(type='MSELoss',
name='loss_p_1',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_1',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
dict(feature_level = 2,
methods=[dict(type='MSELoss',
name='loss_p_2',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_2',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
dict(feature_level = 3,
methods=[dict(type='MSELoss',
name='loss_p_3',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_3',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
]
)
student_cfg = 'configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py'
teacher_cfg = 'configs/faster_rcnn/faster_rcnn_r152_fpn_1x_coco.py'
configs/distillers/DeFeat/DeFeat_deneck_mask_rcnn_swinT_fpn_3x_distill_faster_rcnn_r50_fpn_1x_coco.py 0 → 100644
+ 79
− 0
View file @ 27c917ab
_base_ = [
'../../_base_/models/faster_rcnn_r50_fpn.py',
'../../_base_/datasets/coco_detection.py',
'../../_base_/schedules/schedule_1x.py', '../../_base_/default_runtime.py'
]
# model settings
find_unused_parameters=True
pos_w=0.5
neg_w=2
distiller = dict(
type='DeFeat',
teacher_pretrained = '/mnt/data3/wangguohua/model/mmdet/swin/mask_rcnn_swin_tiny_patch4_window7.pth',
load_teacher_part = 'neck_head',
distill_cfg = [ dict(feature_level = 0,
methods=[dict(type='MSELoss',
name='loss_p_0',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_0',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
dict(feature_level = 1,
methods=[dict(type='MSELoss',
name='loss_p_1',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_1',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
dict(feature_level = 2,
methods=[dict(type='MSELoss',
name='loss_p_2',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_2',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
dict(feature_level = 3,
methods=[dict(type='MSELoss',
name='loss_p_3',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_3',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
]
)
student_cfg = 'configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py'
teacher_cfg = 'configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py'
configs/distillers/DeFeat/DeFeat_faster_rcnn_r152_fpn_1x_distill_faster_rcnn_r50_fpn_1x_coco.py 0 → 100644
+ 80
− 0
View file @ 27c917ab
_base_ = [
'../../_base_/models/faster_rcnn_r50_fpn.py',
'../../_base_/datasets/coco_detection.py',
'../../_base_/schedules/schedule_1x.py', '../../_base_/default_runtime.py'
]
# model settings
# model settings
find_unused_parameters=True
pos_w=1
neg_w=2
distiller = dict(
type='DeFeat',
teacher_pretrained = '/mnt/data3/wangguohua/model/mmdet/faster_rcnn/faster_rcnn_r152_fpn_1x_coco.pth',
load_teacher_part = 'neck_head',
distill_cfg = [ dict(feature_level = 0,
methods=[dict(type='MSELoss',
name='loss_p_0',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_0',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
dict(feature_level = 1,
methods=[dict(type='MSELoss',
name='loss_p_1',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_1',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
dict(feature_level = 2,
methods=[dict(type='MSELoss',
name='loss_p_2',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_2',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
dict(feature_level = 3,
methods=[dict(type='MSELoss',
name='loss_p_3',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_3',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
]
)
student_cfg = 'configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py'
teacher_cfg = 'configs/faster_rcnn/faster_rcnn_r152_fpn_1x_coco.py'
configs/distillers/FGFI/FGFI_faster_rcnn_r152_fpn_1x_distill_faster_rcnn_r50_fpn_1x_coco.py 0 → 100644
+ 80
− 0
View file @ 27c917ab
_base_ = [
'../../_base_/models/faster_rcnn_r50_fpn.py',
'../../_base_/datasets/coco_detection.py',
'../../_base_/schedules/schedule_1x.py', '../../_base_/default_runtime.py'
]
# model settings
# model settings
find_unused_parameters=True
pos_w=1
neg_w=2
distiller = dict(
type='FGFI',
teacher_pretrained = '/mnt/data3/wangguohua/model/mmdet/faster_rcnn/faster_rcnn_r152_fpn_1x_coco.pth',
load_teacher_part = 'neck_head',
distill_cfg = [ dict(feature_level = 0,
methods=[dict(type='MSELoss',
name='loss_p_0',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_0',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
dict(feature_level = 1,
methods=[dict(type='MSELoss',
name='loss_p_1',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_1',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
dict(feature_level = 2,
methods=[dict(type='MSELoss',
name='loss_p_2',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_2',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
dict(feature_level = 3,
methods=[dict(type='MSELoss',
name='loss_p_3',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_3',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
]
)
student_cfg = 'configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py'
teacher_cfg = 'configs/faster_rcnn/faster_rcnn_r152_fpn_1x_coco.py'
configs/distillers/FGFI/FGFI_mask_rcnn_swinT_fpn_3x_distill_faster_rcnn_r50_fpn_1x_coco.py 0 → 100644
+ 79
− 0
View file @ 27c917ab
_base_ = [
'../../_base_/models/faster_rcnn_r50_fpn.py',
'../../_base_/datasets/coco_detection.py',
'../../_base_/schedules/schedule_1x.py', '../../_base_/default_runtime.py'
]
# model settings
find_unused_parameters=True
pos_w=1
neg_w=2
distiller = dict(
type='FGFI',
teacher_pretrained = '/mnt/data3/wangguohua/model/mmdet/swin/mask_rcnn_swin_tiny_patch4_window7.pth',
load_teacher_part = 'neck_head',
distill_cfg = [ dict(feature_level = 0,
methods=[dict(type='MSELoss',
name='loss_p_0',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_0',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
dict(feature_level = 1,
methods=[dict(type='MSELoss',
name='loss_p_1',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_1',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
dict(feature_level = 2,
methods=[dict(type='MSELoss',
name='loss_p_2',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_2',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
dict(feature_level = 3,
methods=[dict(type='MSELoss',
name='loss_p_3',
student_channels = 256,
teacher_channels = 256,
weight = pos_w,
),
dict(type='MSELoss',
name='loss_n_3',
student_channels = 256,
teacher_channels = 256,
weight = neg_w,
)
]
),
]
)
student_cfg = 'configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py'
teacher_cfg = 'configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py'
configs/distillers/ROIfeature_mimicking/roi_fm_faster_rcnn_r152_fpn_1x_distill_faster_rcnn_r50_fpn_1x_coco.py
+ 1
− 1
View file @ 27c917ab
......@@ -6,7 +6,7 @@ _base_ = [
# model settings
find_unused_parameters=True
weight=1
weight=10
distiller = dict(
type='ROIMimicking',
teacher_pretrained = '/mnt/data3/wangguohua/model/mmdet/faster_rcnn/faster_rcnn_r152_fpn_1x_coco.pth',
......
configs/distillers/ROIfeature_mimicking/roi_fm_mask_rcnn_swinT_fpn_3x_distill_faster_rcnn_r50_fpn_1x_coco.py 0 → 100644
+ 30
− 0
View file @ 27c917ab
_base_ = [
'../../_base_/models/faster_rcnn_r50_fpn.py',
'../../_base_/datasets/coco_detection.py',
'../../_base_/schedules/schedule_1x.py', '../../_base_/default_runtime.py'
]
# model settings
find_unused_parameters=True
weight=10
distiller = dict(
type='ROIMimicking',
teacher_pretrained = '/mnt/data3/wangguohua/model/mmdet/swin/mask_rcnn_swin_tiny_patch4_window7.pth',
load_teacher_part = 'neck_head',
distill_cfg = [ dict(student_module = 'roi_head.bbox_head.shared_fcs.1',
teacher_module = 'roi_head.bbox_head.shared_fcs.1',
register_hook = True,
methods=[dict(type='MSELoss',
name='loss_fm',
student_channels = 1024,
teacher_channels = 1024,
weight = weight,
)
]
),
]
)
student_cfg = 'configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py'
teacher_cfg = 'configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py'
configs/distillers/fgd/fgd_faster_rcnn_r101_fpn_1x_distill_faster_rcnn_r50_fpn_1x_voc0712.py
+ 1
− 1
View file @ 27c917ab
......@@ -10,7 +10,7 @@ beta_fgd=0.000025
gamma_fgd=0.00005
lambda_fgd=0.0000005
distiller = dict(
type='DetectionDistiller',
type='FGD',
teacher_pretrained = '/mnt/data3/wangguohua/model/mmdet/faster_rcnn_r101_fpn_1x_voc0712.pth',
init_student = True,
distill_cfg = [ dict(student_module = 'neck.fpn_convs.3.conv',
......
configs/distillers/fgd/fgd_faster_rcnn_r101_fpn_2x_distill_faster_rcnn_r50_fpn_2x_coco.py
+ 1
− 1
View file @ 27c917ab
......@@ -10,7 +10,7 @@ beta_fgd=0.000025
gamma_fgd=0.00005
lambda_fgd=0.0000005
distiller = dict(
type='DetectionDistiller',
type='FGD',
teacher_pretrained = 'https://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r101_fpn_2x_coco/faster_rcnn_r101_fpn_2x_coco_bbox_mAP-0.398_20200504_210455-1d2dac9c.pth',
init_student = True,
distill_cfg = [ dict(student_module = 'neck.fpn_convs.3.conv',
......
configs/distillers/fgd/fgd_faster_rcnn_r152_fpn_1x_distill_faster_rcnn_r50_fpn_1x_coco.py 0 → 100644
+ 87
− 0
View file @ 27c917ab
_base_ = [
'../../_base_/models/faster_rcnn_r50_fpn.py',
'../../_base_/datasets/coco_detection.py',
'../../_base_/schedules/schedule_1x.py', '../../_base_/default_runtime.py'
]
# model settings
find_unused_parameters=True
temp=0.5
alpha_fgd=0.00005
beta_fgd=0.000025
gamma_fgd=0.00005
lambda_fgd=0.0000005
distiller = dict(
type='FGD',
teacher_pretrained = '/mnt/data3/wangguohua/model/mmdet/faster_rcnn/faster_rcnn_r152_fpn_1x_coco.pth',
init_student = True,
distill_cfg = [ dict(student_module = 'neck.fpn_convs.3.conv',
teacher_module = 'neck.fpn_convs.3.conv',
output_hook = True,
methods=[dict(type='FeatureLoss',
name='loss_fgd_fpn_3',
student_channels = 256,
teacher_channels = 256,
temp = temp,
alpha_fgd=alpha_fgd,
beta_fgd=beta_fgd,
gamma_fgd=gamma_fgd,
lambda_fgd=lambda_fgd,
)
]
),
dict(student_module = 'neck.fpn_convs.2.conv',
teacher_module = 'neck.fpn_convs.2.conv',
output_hook = True,
methods=[dict(type='FeatureLoss',
name='loss_fgd_fpn_2',
student_channels = 256,
teacher_channels = 256,
temp = temp,
alpha_fgd=alpha_fgd,
beta_fgd=beta_fgd,
gamma_fgd=gamma_fgd,
lambda_fgd=lambda_fgd,
)
]
),
dict(student_module = 'neck.fpn_convs.1.conv',
teacher_module = 'neck.fpn_convs.1.conv',
output_hook = True,
methods=[dict(type='FeatureLoss',
name='loss_fgd_fpn_1',
student_channels = 256,
teacher_channels = 256,
temp = temp,
alpha_fgd=alpha_fgd,
beta_fgd=beta_fgd,
gamma_fgd=gamma_fgd,
lambda_fgd=lambda_fgd,
)
]
),
dict(student_module = 'neck.fpn_convs.0.conv',
teacher_module = 'neck.fpn_convs.0.conv',
output_hook = True,
methods=[dict(type='FeatureLoss',
name='loss_fgd_fpn_0',
student_channels = 256,
teacher_channels = 256,
temp = temp,
alpha_fgd=alpha_fgd,
beta_fgd=beta_fgd,
gamma_fgd=gamma_fgd,
lambda_fgd=lambda_fgd,
)
]
),
]
)
student_cfg = 'configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py'
teacher_cfg = 'configs/faster_rcnn/faster_rcnn_r152_fpn_1x_coco.py'
# optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0001)
# optimizer_config = dict(_delete_=True, grad_clip=dict(max_norm=35, norm_type=2))
# data = dict(
# samples_per_gpu=2,
# workers_per_gpu=2,)
\ No newline at end of file
configs/distillers/fgd/fgd_mask_rcnn_swinS_fpn_3x_distill_mask_rcnn_swinT_fpn_1x_coco.py
+ 1
− 1
View file @ 27c917ab
......@@ -12,7 +12,7 @@ beta_fgd=0.000025
gamma_fgd=0.00005
lambda_fgd=0.0000005
distiller = dict(
type='DetectionDistiller',
type='FGD',
teacher_pretrained = '/mnt/data3/wangguohua/project/swin/checkpoint/mask_rcnn_swin_small_patch4_window7.pth',
init_student = True,
distill_cfg = [ dict(student_module = 'neck.fpn_convs.3.conv',
......
configs/distillers/fgd/fgd_mask_rcnn_swinT_fpn_3x_distill_faster_rcnn_r50_fpn_1x_coco.py 0 → 100644
+ 88
− 0
View file @ 27c917ab
_base_ = [
'../../_base_/models/faster_rcnn_r50_fpn.py',
'../../_base_/datasets/coco_detection.py',
'../../_base_/schedules/schedule_1x.py', '../../_base_/default_runtime.py'
]
# model settings
find_unused_parameters=True
temp=0.5
alpha_fgd=0.00005
beta_fgd=0.000025
gamma_fgd=0.00005
lambda_fgd=0.0000005
distiller = dict(
type='FGD',
teacher_pretrained = '/mnt/data3/wangguohua/model/mmdet/swin/mask_rcnn_swin_tiny_patch4_window7.pth',
init_student = True,
distill_cfg = [ dict(student_module = 'neck.fpn_convs.3.conv',
teacher_module = 'neck.fpn_convs.3.conv',
output_hook = True,
methods=[dict(type='FeatureLoss',
name='loss_fgd_fpn_3',
student_channels = 256,
teacher_channels = 256,
temp = temp,
alpha_fgd=alpha_fgd,
beta_fgd=beta_fgd,
gamma_fgd=gamma_fgd,
lambda_fgd=lambda_fgd,
)
]
),
dict(student_module = 'neck.fpn_convs.2.conv',
teacher_module = 'neck.fpn_convs.2.conv',
output_hook = True,
methods=[dict(type='FeatureLoss',
name='loss_fgd_fpn_2',
student_channels = 256,
teacher_channels = 256,
temp = temp,
alpha_fgd=alpha_fgd,
beta_fgd=beta_fgd,
gamma_fgd=gamma_fgd,
lambda_fgd=lambda_fgd,
)
]
),
dict(student_module = 'neck.fpn_convs.1.conv',
teacher_module = 'neck.fpn_convs.1.conv',
output_hook = True,
methods=[dict(type='FeatureLoss',
name='loss_fgd_fpn_1',
student_channels = 256,
teacher_channels = 256,
temp = temp,
alpha_fgd=alpha_fgd,
beta_fgd=beta_fgd,
gamma_fgd=gamma_fgd,
lambda_fgd=lambda_fgd,
)
]
),
dict(student_module = 'neck.fpn_convs.0.conv',
teacher_module = 'neck.fpn_convs.0.conv',
output_hook = True,
methods=[dict(type='FeatureLoss',
name='loss_fgd_fpn_0',
student_channels = 256,
teacher_channels = 256,
temp = temp,
alpha_fgd=alpha_fgd,
beta_fgd=beta_fgd,
gamma_fgd=gamma_fgd,
lambda_fgd=lambda_fgd,
)
]
),
]
)
student_cfg = 'configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py'
teacher_cfg = 'configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py'
optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0001)
optimizer_config = dict(_delete_=True, grad_clip=dict(max_norm=35, norm_type=2))
data = dict(
samples_per_gpu=2,
workers_per_gpu=2,)
\ No newline at end of file
configs/distillers/fgd/fgd_mask_rcnn_swinT_fpn_3x_distill_mask_rcnn_swinT_fpn_1x_coco.py
+ 1
− 1
View file @ 27c917ab
......@@ -12,7 +12,7 @@ beta_fgd=0.000025
gamma_fgd=0.00005
lambda_fgd=0.0000005
distiller = dict(
type='DetectionDistiller',
type='FGD',
teacher_pretrained = '/mnt/data3/wangguohua/model/mmdet/swin/mask_rcnn_swin_tiny_patch4_window7.pth',
init_student = True,
distill_cfg = [ dict(student_module = 'neck.fpn_convs.3.conv',
......
configs/distillers/mimic_fpn/mfpn_faster_rcnn_r152_fpn_1x_distill_faster_rcnn_r50_fpn_1x_coco.py 0 → 100644
+ 55
− 0
View file @ 27c917ab
_base_ = [
'../../_base_/models/faster_rcnn_r50_fpn.py',
'../../_base_/datasets/coco_detection.py',
'../../_base_/schedules/schedule_1x.py', '../../_base_/default_runtime.py'
]
# model settings
find_unused_parameters=True
weight=2
distiller = dict(
type='FPNDistiller',
teacher_pretrained = '/mnt/data3/wangguohua/model/mmdet/faster_rcnn/faster_rcnn_r152_fpn_1x_coco.pth',
init_student = 'neck_head',
train_head = False,
distill_cfg = [ dict(feature_level = 0,
methods=[dict(type='MSELoss',
name='loss_mfpn_0',
student_channels = 256,
teacher_channels = 256,
weight = weight,
)
]
),
dict(feature_level = 1,
methods=[dict(type='MSELoss',
name='loss_mfpn_1',
student_channels = 256,
teacher_channels = 256,
weight = weight,
)
]
),
dict(feature_level = 2,
methods=[dict(type='MSELoss',
name='loss_mfpn_2',
student_channels = 256,
teacher_channels = 256,
weight = weight,
)
]
),
dict(feature_level = 3,
methods=[dict(type='MSELoss',
name='loss_mfpn_3',
student_channels = 256,
teacher_channels = 256,
weight = weight,
)
]
),
]
)
student_cfg = 'configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py'
teacher_cfg = 'configs/faster_rcnn/faster_rcnn_r152_fpn_1x_coco.py'
configs/distillers/mimic_fpn/mfpn_mask_rcnn_swinT_fpn_3x_distill_faster_rcnn_r50_fpn_1x_coco.py 0 → 100644
+ 55
− 0
View file @ 27c917ab
_base_ = [
'../../_base_/models/faster_rcnn_r50_fpn.py',
'../../_base_/datasets/coco_detection.py',
'../../_base_/schedules/schedule_1x.py', '../../_base_/default_runtime.py'
]
# model settings
find_unused_parameters=True
weight=2
distiller = dict(
type='FPNDistiller',
teacher_pretrained = '/mnt/data3/wangguohua/model/mmdet/swin/mask_rcnn_swin_tiny_patch4_window7.pth',
init_student = 'neck_head',
train_head = False,
distill_cfg = [ dict(feature_level = 0,
methods=[dict(type='MSELoss',
name='loss_mfpn_0',
student_channels = 256,
teacher_channels = 256,
weight = weight,
)
]
),
dict(feature_level = 1,
methods=[dict(type='MSELoss',
name='loss_mfpn_1',
student_channels = 256,
teacher_channels = 256,
weight = weight,
)
]
),
dict(feature_level = 2,
methods=[dict(type='MSELoss',
name='loss_mfpn_2',
student_channels = 256,
teacher_channels = 256,
weight = weight,
)
]
),
dict(feature_level = 3,
methods=[dict(type='MSELoss',
name='loss_mfpn_3',
student_channels = 256,
teacher_channels = 256,
weight = weight,
)
]
),
]
)
student_cfg = 'configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py'
teacher_cfg = 'configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_3x_coco.py'
mmdet/distillation/distillers/detection_distiller.py → mmdet/distillation/distillers/FGD.py
+ 4
− 3
View file @ 27c917ab
......@@ -11,8 +11,8 @@ from collections import OrderedDict
@DISTILLER.register_module()
class DetectionDistiller(BaseDetector):
"""Base distiller for detectors.
class FGD(BaseDetector):
"""FGD distiller for detectors.
It typically consists of teacher_model and student_model.
"""
def __init__(self,
......@@ -22,7 +22,7 @@ class DetectionDistiller(BaseDetector):
teacher_pretrained=None,
init_student=False):
super(DetectionDistiller, self).__init__()
super(FGD, self).__init__()
self.teacher = build_detector(teacher_cfg.model,
train_cfg=teacher_cfg.get('train_cfg'),
......@@ -34,6 +34,7 @@ class DetectionDistiller(BaseDetector):
self.student= build_detector(student_cfg.model,
train_cfg=student_cfg.get('train_cfg'),
test_cfg=student_cfg.get('test_cfg'))
self.student.init_weights()
if init_student:
t_checkpoint = _load_checkpoint(teacher_pretrained, map_location='cpu')
all_name = []
......
mmdet/distillation/distillers/FGFI.py
+ 6
− 3
View file @ 27c917ab
......@@ -122,7 +122,7 @@ class FGFI(BaseDetector):
for level in range(len(anchor_list[batch])):
gt_level = gt_bboxes[batch][target_lvls==level]
h, w = featmap_sizes[level][0], featmap_sizes[level][1]
mask_per_img = torch.zeros([h, w], dtype=torch.double).cuda()
mask_per_img = torch.zeros([h, w], dtype=torch.float).cuda()
if gt_level.shape[0] > 0:
IoU_map = bbox_overlaps(anchor_list[batch][level], gt_level)
max_iou, _ = torch.max(IoU_map, dim=0)
......@@ -131,7 +131,7 @@ class FGFI(BaseDetector):
max_iou_per_gt = max_iou[ins] * phi
mask_per_gt = torch.sum(IoU_map[:,:,:,ins] > max_iou_per_gt, dim = 2)
mask_per_img += mask_per_gt
mask_per_img = (mask_per_img > 0).double()
mask_per_img = (mask_per_img > 0).float()
mask_level.append(mask_per_img)
mask_batch.append(mask_level)
mask_batch_level = []
......@@ -184,7 +184,8 @@ class FGFI(BaseDetector):
losses.update(rpn_losses)
neck_mask_batch = self.get_roi_mask(rpn_outs[0], img_metas, gt_bboxes, phi=0.5)
with torch.no_grad():
neck_mask_batch = self.get_roi_mask(rpn_outs[0], img_metas, gt_bboxes, phi=0.5)
for item_loc in self.distill_cfg:
......@@ -193,6 +194,7 @@ class FGFI(BaseDetector):
f_t_l = f_t[feature_level]
mask = neck_mask_batch[feature_level]
mask = mask.unsqueeze(1).repeat(1, f_s_l.size(1), 1, 1)
losses['{}_ratio'.format(feature_level)] = mask.sum() / mask.numel()
for item_loss in item_loc.methods:
loss_name = item_loss.name
if 'n' in loss_name:
......@@ -200,6 +202,7 @@ class FGFI(BaseDetector):
losses[loss_name] = self.distill_losses[loss_name](f_s_l, f_t_l, mask)
roi_losses = self.student.roi_head.forward_train(f_s, img_metas, proposal_list,
gt_bboxes, gt_labels,
gt_bboxes_ignore, gt_masks,
......
mmdet/distillation/distillers/__init__.py
+ 2
− 2
View file @ 27c917ab
from .detection_distiller import DetectionDistiller
from .FGD import FGD
from .backbone_distiller import BackboneDistiller
from .fpn_distiller import FPNDistiller
from .head_distiller import HeadDistiller
......@@ -8,7 +8,7 @@ from .DeFeat import DeFeat
from .RPN_FM import RPN_FM
__all__ = [
'DetectionDistiller',
'FGD',
'BackboneDistiller',
'FPNDistiller',
'RPN_FM',
......
submit_work.sh
+ 3
− 1
View file @ 27c917ab
......@@ -13,4 +13,6 @@ sleep 2
# when $pid finished, run these
#PORT=29504 CUDA_VISIBLE_DEVICES=0,1,2,3 tools/dist_train.sh configs/distillers/mimic_fpn/mfpn_trainH3_mask_rcnn_swinS_fpn_3x_distill_mask_rcnn_swinT_fpn_1x_coco.py 4
#PORT=29502 CUDA_VISIBLE_DEVICES=4,5,6,7 tools/dist_train.sh configs/pascal_voc/faster_rcnn_r101_fpn_1x_voc0712.py 4
PORT=29505 tools/dist_train.sh configs/distillers/feature_mimicking/roi_fm_faster_rcnn_r152_fpn_1x_distill_faster_rcnn_r50_fpn_1x_coco.py 8
\ No newline at end of file
#PORT=29505 tools/dist_train.sh configs/distillers/feature_mimicking/roi_fm_faster_rcnn_r152_fpn_1x_distill_faster_rcnn_r50_fpn_1x_coco.py 8
#PORT=29505 tools/dist_train.sh configs/distillers/mimic_fpn/mfpn_tH_mask_rcnn_swinT_fpn_3x_distill_faster_rcnn_r50_fpn_1x_coco.py 8
PORT=29505 tools/dist_train.sh configs/distillers/ROIfeature_mimicking/roi_fm_mask_rcnn_swinT_fpn_3x_distill_faster_rcnn_r50_fpn_1x_coco.py 8
\ No newline at end of file
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment