반응형
# Problem
I'm trying to fine-tune Yolact on 2GPU machine and getting the following error:
$ python train.py --config=yolact_plus_base_config --batch_size=2 --save_interval 1000 --save_folder weights/resnet101_construct_0112_fcon3_yolactplus/
...
KeyError: 'I'
# Solution
I have two pieces of suggestions for this problem:
- The first method, you can decrease the resolution of discard_mask_area.
- The second method is giving zero loss when the maskiou_targets is None in moltibox_loss.py, say my repo for details.
I tried the second one and it has worked.
$ code yolact/layers/modules/multibox_loss.py
...
# Mask IoU Loss
if cfg.use_maskiou:
if maskiou_targets is not None:
losses['I'] = self.mask_iou_loss(net, maskiou_targets)
else:
losses['I'] = torch.tensor(0, device=loc_data.device, dtype=loc_data.dtype)
...
# Reference
[Github Issue] KeyError: 'I' #442: https://github.com/dbolya/yolact/issues/442#issuecomment-744411350
반응형