Release of mask RCNN and multi head FisherPruning

[itsMorteza]

Hi, The loss in Groups is computed based on bbox loss; what is the best way to integrate it with mask loss? Thanks,

[liyangliu]

The loss is as usual, I think you can use the normal mask loss such as BCE.

[itsMorteza]

I used forward_dummy format from mmdetection with using another fork however, it couldn't compute fisher in backward format(tensor size problem). is there any good implementation for two stage detector or segmentation loss? Thanks,

[mls1999725]

I used forward_dummy format from mmdetection with using another fork however, it couldn't compute fisher in backward format(tensor size problem). is there any good implementation for two stage detector or segmentation loss? Thanks,

Hi, I have meet the same problem when compute fisher of faster rcnn. Have you solved this problem?

[itsMorteza]

Unfortunately, no. Even with trial and error, the pruned model couldn't be fine-tuned. the problem initiated with the multi-pass loss; then it couldn't find the whole ancestors very well to make groups.

[liyangliu]

Have you checked whether all the layers that are supposed to be in one group are gathered correctly by the algorithm?

[itsMorteza]

Inside compute fisher backward hook it couldn't update temp fisher info for mask head and linear side wherein the temp fisher info is not same as the result of computing fisher.