Open Zetianuser opened 1 year ago
Hi, I think u are asking about the mask supervision pipeline. As u said, these lines allow detector loss to backpropagate to mask. https://github.com/Cuogeihong/CEASC/blob/2abfd1a99f1b0fe1ed3d51588b64549e1584da50/mmdet/models/dense_heads/sparse_conv_net.py#L115-L118 Meanwhile, the amm loss can supervise mask directly. https://github.com/Cuogeihong/CEASC/blob/2abfd1a99f1b0fe1ed3d51588b64549e1584da50/mmdet/models/dense_heads/gfl_dy_head.py#L484-L485
Hi, I think u are asking about the mask supervision pipeline. As u said, these lines allow detector loss to backpropagate to mask.
Meanwhile, the amm loss can supervise mask directly. https://github.com/Cuogeihong/CEASC/blob/2abfd1a99f1b0fe1ed3d51588b64549e1584da50/mmdet/models/dense_heads/gfl_dy_head.py#L484-L485
I wanna know what is the function of cost_cur = sum(cost_act_cur)/sum(cost_tot_cur)?
just for debug, it has no contribution to backpropagation
感谢您杰出的工作。请问SparseConv中,采样mask与feature之间的梯度是怎么传递的呢?将mask与feature相乘吗?