Well, in your paper, there are results on different frameworks and models, including R-50-C4 and R-101-C4. But in this repo, you don't provide relevant results.
I try to implement this by writing a file named Base-EQL-RCNN-C4.yaml, however, the training process always ends with the following bugs:
FloatingPointError: Loss became infinite or NaN at iteration=2!
loss_dict = {'loss_cls': tensor(nan, device='cuda:0', grad_fn=), 'loss_box_reg': tensor(nan, device='cuda:0', grad_fn=), 'loss_mask': tensor(0.6931, device='cuda:0', grad_fn=), 'loss_rpn_cls': tensor(nan, device='cuda:0', grad_fn=), 'loss_rpn_loc': tensor(inf, device='cuda:0', grad_fn=)}
Well, in your paper, there are results on different frameworks and models, including R-50-C4 and R-101-C4. But in this repo, you don't provide relevant results. I try to implement this by writing a file named Base-EQL-RCNN-C4.yaml, however, the training process always ends with the following bugs:
FloatingPointError: Loss became infinite or NaN at iteration=2! loss_dict = {'loss_cls': tensor(nan, device='cuda:0', grad_fn=), 'loss_box_reg': tensor(nan, device='cuda:0', grad_fn=), 'loss_mask': tensor(0.6931, device='cuda:0', grad_fn=), 'loss_rpn_cls': tensor(nan, device='cuda:0', grad_fn=), 'loss_rpn_loc': tensor(inf, device='cuda:0', grad_fn=)}
Can you give me some help? Base-EQL-RCNN-C4.zip