Rongkun Yang
Rongkun Yang
Hi @thangvubk . I have tried the second way, if I want to train the top-down refinement network based on the pretrian model, should I set the optimizer from the...
OK, thank you for your fast response. As the procedure above, I have solved this problem. Best wishes!
Hi! When I experiments on ScanNet dataset, I'm curious whether the segment information on the 3rd axis of labels can help boost the performance, how does it work? Thank you!
Hi, I meet the same problem, and the target folder does not have any output, have you resolved it. And I think this may be caused by the llama weight...
Dear authors, I'm wondering whether the pretrain of backbone plays an importance effect in the performance of instance segmentation?
This is the error message: RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [512, 1]], which is output 0 of TBackward,...
Hello, dear authors, I follow the code of this PR to process the scannet++ dataset, and I found that the color is processed all into 0, this will cause an...
I process the scannet++ dataset use the sgiFormer to process Superpoint data
I would like to consult another question, when I train the SGIFormer network in this code, the Loss will get NaN, which would damage the training process, could you help...
OK, thank you for your reply, I guess the NaN loss is caused by large batch size and distributed training, and I found the learning rate seems to be inconsistent...