your code:
box_head = TransformerHead(
cfg=cfg,
trans_names=cfg.MODEL.TRANSFORMER.NAMES_1ST,
kernel_size=cfg.MODEL.TRANSFORMER.KERNEL_SIZE_1ST,
use_feature_mask=cfg.MODEL.TRANSFORMER.USE_MASK_1ST,
)
box_head_2nd = TransformerHead(
cfg=cfg,
trans_names=cfg.MODEL.TRANSFORMER.NAMES_2ND,
kernel_size=cfg.MODEL.TRANSFORMER.KERNEL_SIZE_2ND,
use_feature_mask=cfg.MODEL.TRANSFORMER.USE_MASK_2ND,
)
box_head_3rd = TransformerHead(
cfg=cfg,
trans_names=cfg.MODEL.TRANSFORMER.NAMES_3RD,
kernel_size=cfg.MODEL.TRANSFORMER.KERNEL_SIZE_3RD,
use_feature_mask=cfg.MODEL.TRANSFORMER.USE_MASK_3RD,
)
definition: box_head box_head_2nd box_head_3rd
then:
box_features = self.roi_heads.box_roi_pool(features, boxes, images.image_sizes) #MultiScaleRoIAlign
box_features_2nd = self.roi_heads.box_head_2nd(box_features)
embeddings_2nd, _ = self.roi_heads.embedding_head_2nd(box_features_2nd) #NormAwareEmbedding
box_features_3rd = self.roi_heads.box_head_3rd(box_features) #box_features
embeddings_3rd, _ = self.roi_heads.embedding_head_3rd(box_features_3rd)

I wonder why this place doesn't have the first TransformerHead?But what is drawn in Figure 2 is to go through the first TransformerHead.