Closed Yuxin-CV closed 4 years ago
BTW, I wonder the batch_size, the num_GPU and the GPU type(GPU memory) you use in your experiment(37.1 mask AP).
@Yuxin-CV Total batchsize of 24 on 8 GPUs(v100, 32G)
@Yuxin-CV Total batchsize of 24 on 8 GPUs(v100, 32G)
Thanks for your reply. @Epiphqny I think I find a bug in your implementation, detailed in https://github.com/aim-uofa/AdelaiDet/issues/39
@Yuxin-CV Thanks for your attention, but nowadays I have no time to improve the code, you can report the bugs here and I'll correct and train it in the future.
How do you solve the problem #2? Since I find no definitions for self.cls_tower and self.bbox_tower. @Yuxin-CV
How do you solve problem #2? Since I find no definitions for self.cls_tower and self.bbox_tower. @Yuxin-CV
@Epiphqny Hi,I just found the same problem with what @ziwei-zh mentioned above. Could you tell us about the implementation of self.cls_tower and self.bbox_tower.
?
Thank you.
How do you solve the problem #2? Since I find no definitions for self.cls_tower and self.bbox_tower. @Yuxin-CV
The code of CondInst is released at AdelaiDet.
Hi~ @Epiphqny Thanks for your code ~ I run your code and have some questions about your implementation.
The mask tower part is not consist with the paper. https://github.com/Epiphqny/CondInst/blob/ea3f717fce73a8e4c273f1379c9d9c3550387e1b/fcos/modeling/fcos/fcos.py#L163 I think it should be 4 x conv with num_channel = 128. Your implementation is 8 x conv with num_channel = 256.
The weight of mask tower is not initiated. https://github.com/Epiphqny/CondInst/blob/ea3f717fce73a8e4c273f1379c9d9c3550387e1b/fcos/modeling/fcos/fcos.py#L217-L221
After I correct 1 & 2, I found that the result is quite poor... As I mentioned in https://github.com/aim-uofa/AdelaiDet/issues/39 I also found that the loss of cls & reg is higher than FCOS. So I am interested in the loss behavior & Box AP of your R-101-3x experiment. Can you share some helpful information with me? Because it is quite strange that the Box AP is degenerated in my experiment, it should be benefit from the multi-task training.