|
|
model = dict( |
|
|
type="TadTR", |
|
|
projection=dict( |
|
|
type="ConvSingleProj", |
|
|
in_channels=2048, |
|
|
out_channels=256, |
|
|
num_convs=1, |
|
|
conv_cfg=dict(kernel_size=1, padding=0), |
|
|
norm_cfg=dict(type="GN", num_groups=32), |
|
|
act_cfg=None, |
|
|
), |
|
|
transformer=dict( |
|
|
type="TadTRTransformer", |
|
|
num_proposals=40, |
|
|
num_classes=20, |
|
|
with_act_reg=True, |
|
|
roi_size=16, |
|
|
roi_extend_ratio=0.25, |
|
|
aux_loss=True, |
|
|
position_embedding=dict( |
|
|
type="PositionEmbeddingSine", |
|
|
num_pos_feats=256, |
|
|
temperature=10000, |
|
|
offset=-0.5, |
|
|
normalize=True, |
|
|
), |
|
|
encoder=dict( |
|
|
type="DeformableDETREncoder", |
|
|
embed_dim=256, |
|
|
num_heads=8, |
|
|
num_points=4, |
|
|
attn_dropout=0.1, |
|
|
ffn_dim=1024, |
|
|
ffn_dropout=0.1, |
|
|
num_layers=4, |
|
|
num_feature_levels=1, |
|
|
post_norm=False, |
|
|
), |
|
|
decoder=dict( |
|
|
type="DeformableDETRDecoder", |
|
|
embed_dim=256, |
|
|
num_heads=8, |
|
|
num_points=4, |
|
|
attn_dropout=0.1, |
|
|
ffn_dim=1024, |
|
|
ffn_dropout=0.1, |
|
|
num_layers=4, |
|
|
num_feature_levels=1, |
|
|
return_intermediate=True, |
|
|
), |
|
|
loss=dict( |
|
|
type="TadTRSetCriterion", |
|
|
num_classes=20, |
|
|
matcher=dict( |
|
|
type="HungarianMatcher", |
|
|
cost_class=6.0, |
|
|
cost_bbox=5.0, |
|
|
cost_giou=2.0, |
|
|
cost_class_type="focal_loss_cost", |
|
|
iou_type="iou", |
|
|
use_multi_class=False, |
|
|
), |
|
|
loss_class_type="focal_loss", |
|
|
weight_dict=dict( |
|
|
loss_class=2.0, |
|
|
loss_bbox=5.0, |
|
|
loss_iou=2.0, |
|
|
loss_actionness=4.0, |
|
|
), |
|
|
use_multi_class=False, |
|
|
), |
|
|
), |
|
|
) |
|
|
|
