why adding nms_embedding_feat and nms_attention_1 together

[happynewya]

I don't understand codes below, in
resnet_v1_101_rcnn_attention_1024_pairwise_position_multi_head_16_learn_nms.py

You added one of Relation Module inputs and Relation Module output together as the input of later layer. But you did not explain why you do this in your paper.

nms_attention_1, nms_softmax_1 = self.attention_module_nms_multi_head(
                nms_embedding_feat, nms_position_matrix,
                num_rois=first_n, index=1, group=16,
                dim=(1024, 1024, 128), fc_dim=(64, 16), feat_dim=128
)
nms_all_feat_1 = nms_embedding_feat + nms_attention_1

I think that the network figure should add the green path

btw, nms_softmax_1 is redundant, I personally recommend deleting it in your release code.