|
|
model = dict( |
|
|
type="BMN", |
|
|
projection=dict( |
|
|
type="ConvSingleProj", |
|
|
in_channels=400, |
|
|
out_channels=256, |
|
|
num_convs=2, |
|
|
conv_cfg=dict(groups=4), |
|
|
), |
|
|
rpn_head=dict( |
|
|
type="TemporalEvaluationHead", |
|
|
in_channels=256, |
|
|
num_classes=2, |
|
|
conv_cfg=dict(groups=4), |
|
|
loss=dict(pos_thresh=0.5, gt_type=["startness", "endness"]), |
|
|
), |
|
|
roi_head=dict( |
|
|
type="StandardProposalMapHead", |
|
|
proposal_generator=dict(type="DenseProposalMap", tscale=128, dscale=128), |
|
|
proposal_roi_extractor=dict( |
|
|
type="BMNExtractor", |
|
|
in_channels=256, |
|
|
roi_channels=512, |
|
|
out_channels=128, |
|
|
tscale=128, |
|
|
dscale=128, |
|
|
prop_extend_ratio=0.5, |
|
|
), |
|
|
proposal_head=dict( |
|
|
type="PEMHead", |
|
|
in_channels=128, |
|
|
feat_channels=128, |
|
|
num_convs=2, |
|
|
num_classes=2, |
|
|
loss=dict( |
|
|
cls_loss=dict(type="BalancedBCELoss", pos_thresh=0.9), |
|
|
reg_loss=dict(type="BalancedL2Loss", high_thresh=0.7, low_thresh=0.3, weight=5.0), |
|
|
), |
|
|
), |
|
|
), |
|
|
) |
|
|
|