jbwang1997
commented
2 years ago Seem there is no mistake in your config. I consider it's because the training schedule is too short or the model doesn't converge.
Closed mZhenz closed 2 years ago
jbwang1997
commented
2 years ago Seem there is no mistake in your config. I consider it's because the training schedule is too short or the model doesn't converge.
mZhenz
commented
2 years ago Seem there is no mistake in your config. I consider it's because the training schedule is too short or the model doesn't converge.
This is the evaluation result of the first epoch, and seem it is not because training schedule is too short. Is there any suggestion to check this bug. Thanks a lot! I just modify this config file and do not modify the config file below. "../base/datasets/coco_instance.py", "../base/schedules/schedule_1x.py", "../base/default_runtime.py"
btw, could there be something problem with the setting find_unused_parameters = True?
mZhenz
commented
2 years ago Here is my training log. If anyone could give me some suggestion?
jbwang1997
commented
2 years ago find_unused_parameters = True is necessary because the final conv of moblienet isn't used in producing loss. The config of YOLOv3+ MoblieNet also adds find_unused_parameters = True.
I read the structure of moblienetv2, the strides should be [8, 16, 32] since the first layer has stride 2. You can first try to modify the strides in anchor_generator and roi_align. Then you can try to reduce the learning rate to see if the loss can decrease normally.
mZhenz
commented
2 years ago @jbwang1997 Problem solved. Thanks you so much!
Here is the config file of mask_rcnn_mobilenetv2_fpn_1x_coco.py for the convenience of others.
_base_ = [
"../_base_/datasets/coco_instance.py",
"../_base_/schedules/schedule_1x.py",
"../_base_/default_runtime.py"
]
# model settings
model = dict(
type='MaskRCNN',
backbone=dict(
type='MobileNetV2',
out_indices=(2, 4, 6),
act_cfg=dict(type='LeakyReLU', negative_slope=0.1),
init_cfg=dict(
type='Pretrained',
checkpoint='open-mmlab://mmdet/mobilenet_v2',
)),
neck=dict(
type='FPN',
in_channels=[32, 96, 320],
out_channels=256,
num_outs=3),
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=[8, 16, 32]),
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='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=[8, 16, 32]),
bbox_head=dict(
type='Shared2FCBBoxHead',
in_channels=256,
fc_out_channels=1024,
roi_feat_size=7,
num_classes=80,
bbox_coder=dict(
type='DeltaXYWHBBoxCoder',
target_means=[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)),
mask_roi_extractor=dict(
type='SingleRoIExtractor',
roi_layer=dict(type='RoIAlign', output_size=14, sampling_ratio=0),
out_channels=256,
featmap_strides=[8, 16, 32]),
mask_head=dict(
type='FCNMaskHead',
num_convs=4,
in_channels=256,
conv_out_channels=256,
num_classes=80,
loss_mask=dict(
type='CrossEntropyLoss', use_mask=True, 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=-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=True,
ignore_iof_thr=-1),
sampler=dict(
type='RandomSampler',
num=512,
pos_fraction=0.25,
neg_pos_ub=-1,
add_gt_as_proposals=True),
mask_size=28,
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,
mask_thr_binary=0.5)
# soft-nms is also supported for rcnn testing
# e.g., nms=dict(type='soft_nms', iou_threshold=0.5, min_score=0.05)
))
noreenanwar
commented
1 year ago @jbwang1997 Problem solved. Thanks you so much!
Here is the config file of
mask_rcnn_mobilenetv2_fpn_1x_coco.pyfor the convenience of others._base_ = [ "../_base_/datasets/coco_instance.py", "../_base_/schedules/schedule_1x.py", "../_base_/default_runtime.py" ] # model settings model = dict( type='MaskRCNN', backbone=dict( type='MobileNetV2', out_indices=(2, 4, 6), act_cfg=dict(type='LeakyReLU', negative_slope=0.1), init_cfg=dict( type='Pretrained', checkpoint='open-mmlab://mmdet/mobilenet_v2', )), neck=dict( type='FPN', in_channels=[32, 96, 320], out_channels=256, num_outs=3), 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=[8, 16, 32]), 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='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=[8, 16, 32]), bbox_head=dict( type='Shared2FCBBoxHead', in_channels=256, fc_out_channels=1024, roi_feat_size=7, num_classes=80, bbox_coder=dict( type='DeltaXYWHBBoxCoder', target_means=[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)), mask_roi_extractor=dict( type='SingleRoIExtractor', roi_layer=dict(type='RoIAlign', output_size=14, sampling_ratio=0), out_channels=256, featmap_strides=[8, 16, 32]), mask_head=dict( type='FCNMaskHead', num_convs=4, in_channels=256, conv_out_channels=256, num_classes=80, loss_mask=dict( type='CrossEntropyLoss', use_mask=True, 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=-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=True, ignore_iof_thr=-1), sampler=dict( type='RandomSampler', num=512, pos_fraction=0.25, neg_pos_ub=-1, add_gt_as_proposals=True), mask_size=28, 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, mask_thr_binary=0.5) # soft-nms is also supported for rcnn testing # e.g., nms=dict(type='soft_nms', iou_threshold=0.5, min_score=0.05) ))Hi, I am also facing the same issue but with HRNET backbone .Any idea how to solve this?
Hi, I try to replace the backbone of mask rcnn to mobilenetv2. And I firstly get the error like that.
Therefore, I add
find_unused_parameters = Truein my config file. Well, it can train normally. But it get all evaluation metric equals to 0 when evalution like below.And here is my config files. Is there anything wrong?