Skip to content
This repository has been archived by the owner on Nov 21, 2023. It is now read-only.

Commit

Permalink
add group norm
Browse files Browse the repository at this point in the history 
Summary: Add GroupNorm support to master Detectron.

Reviewed By: rbgirshick

Differential Revision: D7611892

fbshipit-source-id: dc4fb84a0e2167b05fd8a94ee0ff1ab1c21369b7
  • Loading branch information
@facebook-github-bot
Kaiming He authored and facebook-github-bot committed Apr 24, 2018
1 parent 0fda5f9 commit c7692eb
Show file tree
Hide file tree
Showing 15 changed files with 787 additions and 43 deletions.
52 changes: 52 additions & 0 deletions configs/04_2018_gn_baselines/e2e_mask_rcnn_R-101-FPN_2x_gn.yaml
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,52 @@
MODEL:
TYPE: generalized_rcnn
CONV_BODY: FPN.add_fpn_ResNet101_conv5_body
NUM_CLASSES: 81
FASTER_RCNN: True
MASK_ON: True
NUM_GPUS: 8
SOLVER:
WEIGHT_DECAY: 0.0001
LR_POLICY: steps_with_decay
BASE_LR: 0.02
GAMMA: 0.1
MAX_ITER: 180000
STEPS: [0, 120000, 160000]
FPN:
FPN_ON: True
MULTILEVEL_ROIS: True
MULTILEVEL_RPN: True
USE_GN: True # Note: use GN on the FPN-specific layers
RESNETS:
STRIDE_1X1: False # default True for MSRA; False for C2 or Torch models
TRANS_FUNC: bottleneck_gn_transformation # Note: this is a GN bottleneck transform
STEM_FUNC: basic_gn_stem # Note: this is a GN stem
SHORTCUT_FUNC: basic_gn_shortcut # Note: this is a GN shortcut
FAST_RCNN:
ROI_BOX_HEAD: fast_rcnn_heads.add_roi_Xconv1fc_gn_head # Note: this is a Conv GN head
ROI_XFORM_METHOD: RoIAlign
ROI_XFORM_RESOLUTION: 7
ROI_XFORM_SAMPLING_RATIO: 2
MRCNN:
ROI_MASK_HEAD: mask_rcnn_heads.mask_rcnn_fcn_head_v1up4convs_gn # Note: this is a GN mask head
RESOLUTION: 28 # (output mask resolution) default 14
ROI_XFORM_METHOD: RoIAlign
ROI_XFORM_RESOLUTION: 14 # default 7
ROI_XFORM_SAMPLING_RATIO: 2 # default 0
DILATION: 1 # default 2
CONV_INIT: MSRAFill # default GaussianFill
TRAIN:
WEIGHTS: https://s3-us-west-2.amazonaws.com/detectron/ImageNetPretrained/47592356/R-101-GN.pkl # Note: a GN pre-trained model
DATASETS: ('coco_2014_train', 'coco_2014_valminusminival')
SCALES: (800,)
MAX_SIZE: 1333
BATCH_SIZE_PER_IM: 512
RPN_PRE_NMS_TOP_N: 2000 # Per FPN level
TEST:
DATASETS: ('coco_2014_minival',)
SCALE: 800
MAX_SIZE: 1333
NMS: 0.5
RPN_PRE_NMS_TOP_N: 1000 # Per FPN level
RPN_POST_NMS_TOP_N: 1000
OUTPUT_DIR: .
52 changes: 52 additions & 0 deletions configs/04_2018_gn_baselines/e2e_mask_rcnn_R-101-FPN_3x_gn.yaml
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,52 @@
MODEL:
TYPE: generalized_rcnn
CONV_BODY: FPN.add_fpn_ResNet101_conv5_body
NUM_CLASSES: 81
FASTER_RCNN: True
MASK_ON: True
NUM_GPUS: 8
SOLVER:
WEIGHT_DECAY: 0.0001
LR_POLICY: steps_with_decay
BASE_LR: 0.02
GAMMA: 0.1
MAX_ITER: 270000
STEPS: [0, 210000, 250000]
FPN:
FPN_ON: True
MULTILEVEL_ROIS: True
MULTILEVEL_RPN: True
USE_GN: True # Note: use GN on the FPN-specific layers
RESNETS:
STRIDE_1X1: False # default True for MSRA; False for C2 or Torch models
TRANS_FUNC: bottleneck_gn_transformation # Note: this is a GN bottleneck transform
STEM_FUNC: basic_gn_stem # Note: this is a GN stem
SHORTCUT_FUNC: basic_gn_shortcut # Note: this is a GN shortcut
FAST_RCNN:
ROI_BOX_HEAD: fast_rcnn_heads.add_roi_Xconv1fc_gn_head # Note: this is a Conv GN head
ROI_XFORM_METHOD: RoIAlign
ROI_XFORM_RESOLUTION: 7
ROI_XFORM_SAMPLING_RATIO: 2
MRCNN:
ROI_MASK_HEAD: mask_rcnn_heads.mask_rcnn_fcn_head_v1up4convs_gn # Note: this is a GN mask head
RESOLUTION: 28 # (output mask resolution) default 14
ROI_XFORM_METHOD: RoIAlign
ROI_XFORM_RESOLUTION: 14 # default 7
ROI_XFORM_SAMPLING_RATIO: 2 # default 0
DILATION: 1 # default 2
CONV_INIT: MSRAFill # default GaussianFill
TRAIN:
WEIGHTS: https://s3-us-west-2.amazonaws.com/detectron/ImageNetPretrained/47592356/R-101-GN.pkl # Note: a GN pre-trained model
DATASETS: ('coco_2014_train', 'coco_2014_valminusminival')
SCALES: (800,)
MAX_SIZE: 1333
BATCH_SIZE_PER_IM: 512
RPN_PRE_NMS_TOP_N: 2000 # Per FPN level
TEST:
DATASETS: ('coco_2014_minival',)
SCALE: 800
MAX_SIZE: 1333
NMS: 0.5
RPN_PRE_NMS_TOP_N: 1000 # Per FPN level
RPN_POST_NMS_TOP_N: 1000
OUTPUT_DIR: .
52 changes: 52 additions & 0 deletions configs/04_2018_gn_baselines/e2e_mask_rcnn_R-50-FPN_2x_gn.yaml
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,52 @@
MODEL:
TYPE: generalized_rcnn
CONV_BODY: FPN.add_fpn_ResNet50_conv5_body
NUM_CLASSES: 81
FASTER_RCNN: True
MASK_ON: True
NUM_GPUS: 8
SOLVER:
WEIGHT_DECAY: 0.0001
LR_POLICY: steps_with_decay
BASE_LR: 0.02
GAMMA: 0.1
MAX_ITER: 180000
STEPS: [0, 120000, 160000]
FPN:
FPN_ON: True
MULTILEVEL_ROIS: True
MULTILEVEL_RPN: True
USE_GN: True # Note: use GN on the FPN-specific layers
RESNETS:
STRIDE_1X1: False # default True for MSRA; False for C2 or Torch models
TRANS_FUNC: bottleneck_gn_transformation # Note: this is a GN bottleneck transform
STEM_FUNC: basic_gn_stem # Note: this is a GN stem
SHORTCUT_FUNC: basic_gn_shortcut # Note: this is a GN shortcut
FAST_RCNN:
ROI_BOX_HEAD: fast_rcnn_heads.add_roi_Xconv1fc_gn_head # Note: this is a Conv GN head
ROI_XFORM_METHOD: RoIAlign
ROI_XFORM_RESOLUTION: 7
ROI_XFORM_SAMPLING_RATIO: 2
MRCNN:
ROI_MASK_HEAD: mask_rcnn_heads.mask_rcnn_fcn_head_v1up4convs_gn # Note: this is a GN mask head
RESOLUTION: 28 # (output mask resolution) default 14
ROI_XFORM_METHOD: RoIAlign
ROI_XFORM_RESOLUTION: 14 # default 7
ROI_XFORM_SAMPLING_RATIO: 2 # default 0
DILATION: 1 # default 2
CONV_INIT: MSRAFill # default GaussianFill
TRAIN:
WEIGHTS: https://s3-us-west-2.amazonaws.com/detectron/ImageNetPretrained/47261647/R-50-GN.pkl # Note: a GN pre-trained model
DATASETS: ('coco_2014_train', 'coco_2014_valminusminival')
SCALES: (800,)
MAX_SIZE: 1333
BATCH_SIZE_PER_IM: 512
RPN_PRE_NMS_TOP_N: 2000 # Per FPN level
TEST:
DATASETS: ('coco_2014_minival',)
SCALE: 800
MAX_SIZE: 1333
NMS: 0.5
RPN_PRE_NMS_TOP_N: 1000 # Per FPN level
RPN_POST_NMS_TOP_N: 1000
OUTPUT_DIR: .
52 changes: 52 additions & 0 deletions configs/04_2018_gn_baselines/e2e_mask_rcnn_R-50-FPN_3x_gn.yaml
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,52 @@
MODEL:
TYPE: generalized_rcnn
CONV_BODY: FPN.add_fpn_ResNet50_conv5_body
NUM_CLASSES: 81
FASTER_RCNN: True
MASK_ON: True
NUM_GPUS: 8
SOLVER:
WEIGHT_DECAY: 0.0001
LR_POLICY: steps_with_decay
BASE_LR: 0.02
GAMMA: 0.1
MAX_ITER: 270000
STEPS: [0, 210000, 250000]
FPN:
FPN_ON: True
MULTILEVEL_ROIS: True
MULTILEVEL_RPN: True
USE_GN: True # Note: use GN on the FPN-specific layers
RESNETS:
STRIDE_1X1: False # default True for MSRA; False for C2 or Torch models
TRANS_FUNC: bottleneck_gn_transformation # Note: this is a GN bottleneck transform
STEM_FUNC: basic_gn_stem # Note: this is a GN stem
SHORTCUT_FUNC: basic_gn_shortcut # Note: this is a GN shortcut
FAST_RCNN:
ROI_BOX_HEAD: fast_rcnn_heads.add_roi_Xconv1fc_gn_head # Note: this is a Conv GN head
ROI_XFORM_METHOD: RoIAlign
ROI_XFORM_RESOLUTION: 7
ROI_XFORM_SAMPLING_RATIO: 2
MRCNN:
ROI_MASK_HEAD: mask_rcnn_heads.mask_rcnn_fcn_head_v1up4convs_gn # Note: this is a GN mask head
RESOLUTION: 28 # (output mask resolution) default 14
ROI_XFORM_METHOD: RoIAlign
ROI_XFORM_RESOLUTION: 14 # default 7
ROI_XFORM_SAMPLING_RATIO: 2 # default 0
DILATION: 1 # default 2
CONV_INIT: MSRAFill # default GaussianFill
TRAIN:
WEIGHTS: https://s3-us-west-2.amazonaws.com/detectron/ImageNetPretrained/47261647/R-50-GN.pkl # Note: a GN pre-trained model
DATASETS: ('coco_2014_train', 'coco_2014_valminusminival')
SCALES: (800,)
MAX_SIZE: 1333
BATCH_SIZE_PER_IM: 512
RPN_PRE_NMS_TOP_N: 2000 # Per FPN level
TEST:
DATASETS: ('coco_2014_minival',)
SCALE: 800
MAX_SIZE: 1333
NMS: 0.5
RPN_PRE_NMS_TOP_N: 1000 # Per FPN level
RPN_POST_NMS_TOP_N: 1000
OUTPUT_DIR: .
52 changes: 52 additions & 0 deletions configs/04_2018_gn_baselines/mask_rcnn_R-50-FPN_1x_gn.yaml
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,52 @@
# WARNING: this script uses **pre-computed** BN-based proposals, and is for quick debugging only.
MODEL:
TYPE: generalized_rcnn
CONV_BODY: FPN.add_fpn_ResNet50_conv5_body
NUM_CLASSES: 81
MASK_ON: True
NUM_GPUS: 8
SOLVER:
WEIGHT_DECAY: 0.0001
LR_POLICY: steps_with_decay
BASE_LR: 0.02
GAMMA: 0.1
MAX_ITER: 90000
STEPS: [0, 60000, 80000]
FPN:
FPN_ON: True
MULTILEVEL_ROIS: True
MULTILEVEL_RPN: True
USE_GN: True # Note: use GN on the FPN-specific layers
RESNETS:
STRIDE_1X1: False # default True for MSRA; False for C2 or Torch models
TRANS_FUNC: bottleneck_gn_transformation # Note: this is a GN bottleneck transform
STEM_FUNC: basic_gn_stem # Note: this is a GN stem
SHORTCUT_FUNC: basic_gn_shortcut # Note: this is a GN shortcut
FAST_RCNN:
ROI_BOX_HEAD: fast_rcnn_heads.add_roi_Xconv1fc_gn_head # Note: this is a Conv GN head
ROI_XFORM_METHOD: RoIAlign
ROI_XFORM_RESOLUTION: 7
ROI_XFORM_SAMPLING_RATIO: 2
MRCNN:
ROI_MASK_HEAD: mask_rcnn_heads.mask_rcnn_fcn_head_v1up4convs_gn # Note: this is a GN mask head
RESOLUTION: 28 # (output mask resolution) default 14
ROI_XFORM_METHOD: RoIAlign
ROI_XFORM_RESOLUTION: 14 # default 7
ROI_XFORM_SAMPLING_RATIO: 2 # default 0
DILATION: 1 # default 2
CONV_INIT: MSRAFill # default GaussianFill
TRAIN:
WEIGHTS: https://s3-us-west-2.amazonaws.com/detectron/ImageNetPretrained/47261647/R-50-GN.pkl # Note: a GN pre-trained model
DATASETS: ('coco_2014_train', 'coco_2014_valminusminival')
PROPOSAL_FILES: ('https://s3-us-west-2.amazonaws.com/detectron/35998814/12_2017_baselines/rpn_R-50-FPN_1x.yaml.08_06_03.Axg0r179/output/test/coco_2014_train/generalized_rcnn/rpn_proposals.pkl', 'https://s3-us-west-2.amazonaws.com/detectron/35998814/12_2017_baselines/rpn_R-50-FPN_1x.yaml.08_06_03.Axg0r179/output/test/coco_2014_valminusminival/generalized_rcnn/rpn_proposals.pkl')
SCALES: (800,)
MAX_SIZE: 1333
BATCH_SIZE_PER_IM: 512
TEST:
DATASETS: ('coco_2014_minival',)
PROPOSAL_FILES: ('https://s3-us-west-2.amazonaws.com/detectron/35998814/12_2017_baselines/rpn_R-50-FPN_1x.yaml.08_06_03.Axg0r179/output/test/coco_2014_minival/generalized_rcnn/rpn_proposals.pkl',)
PROPOSAL_LIMIT: 1000
SCALE: 800
MAX_SIZE: 1333
NMS: 0.5
OUTPUT_DIR: .
52 changes: 52 additions & 0 deletions configs/04_2018_gn_baselines/scratch_e2e_mask_rcnn_R-101-FPN_3x_gn.yaml
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,52 @@
MODEL:
TYPE: generalized_rcnn
CONV_BODY: FPN.add_fpn_ResNet101_conv5_body
NUM_CLASSES: 81
FASTER_RCNN: True
MASK_ON: True
NUM_GPUS: 8
SOLVER:
WEIGHT_DECAY: 0.0001
LR_POLICY: steps_with_decay
BASE_LR: 0.02
GAMMA: 0.1
MAX_ITER: 270000
STEPS: [0, 210000, 250000]
FPN:
FPN_ON: True
MULTILEVEL_ROIS: True
MULTILEVEL_RPN: True
USE_GN: True # Note: use GN on the FPN-specific layers
RESNETS:
STRIDE_1X1: False # default True for MSRA; False for C2 or Torch models
TRANS_FUNC: bottleneck_gn_transformation # Note: this is a GN bottleneck transform
STEM_FUNC: basic_gn_stem # Note: this is a GN stem
SHORTCUT_FUNC: basic_gn_shortcut # Note: this is a GN shortcut
FAST_RCNN:
ROI_BOX_HEAD: fast_rcnn_heads.add_roi_Xconv1fc_gn_head # Note: this is a Conv GN head
ROI_XFORM_METHOD: RoIAlign
ROI_XFORM_RESOLUTION: 7
ROI_XFORM_SAMPLING_RATIO: 2
MRCNN:
ROI_MASK_HEAD: mask_rcnn_heads.mask_rcnn_fcn_head_v1up4convs_gn # Note: this is a GN mask head
RESOLUTION: 28 # (output mask resolution) default 14
ROI_XFORM_METHOD: RoIAlign
ROI_XFORM_RESOLUTION: 14 # default 7
ROI_XFORM_SAMPLING_RATIO: 2 # default 0
DILATION: 1 # default 2
CONV_INIT: MSRAFill # default GaussianFill
TRAIN:
# WEIGHTS: N/A
DATASETS: ('coco_2014_train', 'coco_2014_valminusminival')
SCALES: (800,)
MAX_SIZE: 1333
BATCH_SIZE_PER_IM: 512
RPN_PRE_NMS_TOP_N: 2000 # Per FPN level
TEST:
DATASETS: ('coco_2014_minival',)
SCALE: 800
MAX_SIZE: 1333
NMS: 0.5
RPN_PRE_NMS_TOP_N: 1000 # Per FPN level
RPN_POST_NMS_TOP_N: 1000
OUTPUT_DIR: .
52 changes: 52 additions & 0 deletions configs/04_2018_gn_baselines/scratch_e2e_mask_rcnn_R-50-FPN_3x_gn.yaml
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,52 @@
MODEL:
TYPE: generalized_rcnn
CONV_BODY: FPN.add_fpn_ResNet50_conv5_body
NUM_CLASSES: 81
FASTER_RCNN: True
MASK_ON: True
NUM_GPUS: 8
SOLVER:
WEIGHT_DECAY: 0.0001
LR_POLICY: steps_with_decay
BASE_LR: 0.02
GAMMA: 0.1
MAX_ITER: 270000
STEPS: [0, 210000, 250000]
FPN:
FPN_ON: True
MULTILEVEL_ROIS: True
MULTILEVEL_RPN: True
USE_GN: True # Note: use GN on the FPN-specific layers
RESNETS:
STRIDE_1X1: False # default True for MSRA; False for C2 or Torch models
TRANS_FUNC: bottleneck_gn_transformation # Note: this is a GN bottleneck transform
STEM_FUNC: basic_gn_stem # Note: this is a GN stem
SHORTCUT_FUNC: basic_gn_shortcut # Note: this is a GN shortcut
FAST_RCNN:
ROI_BOX_HEAD: fast_rcnn_heads.add_roi_Xconv1fc_gn_head # Note: this is a Conv GN head
ROI_XFORM_METHOD: RoIAlign
ROI_XFORM_RESOLUTION: 7
ROI_XFORM_SAMPLING_RATIO: 2
MRCNN:
ROI_MASK_HEAD: mask_rcnn_heads.mask_rcnn_fcn_head_v1up4convs_gn # Note: this is a GN mask head
RESOLUTION: 28 # (output mask resolution) default 14
ROI_XFORM_METHOD: RoIAlign
ROI_XFORM_RESOLUTION: 14 # default 7
ROI_XFORM_SAMPLING_RATIO: 2 # default 0
DILATION: 1 # default 2
CONV_INIT: MSRAFill # default GaussianFill
TRAIN:
# WEIGHTS: N/A
DATASETS: ('coco_2014_train', 'coco_2014_valminusminival')
SCALES: (800,)
MAX_SIZE: 1333
BATCH_SIZE_PER_IM: 512
RPN_PRE_NMS_TOP_N: 2000 # Per FPN level
TEST:
DATASETS: ('coco_2014_minival',)
SCALE: 800
MAX_SIZE: 1333
NMS: 0.5
RPN_PRE_NMS_TOP_N: 1000 # Per FPN level
RPN_POST_NMS_TOP_N: 1000
OUTPUT_DIR: .
Loading

0 comments on commit c7692eb

Please sign in to comment.