Closed yCobanoglu closed 2 years ago
.json.
files, so your gt_*.txt
couldn't be the cause. But you should ensure that your data annotation is consistent with how the image has been processed and that all bounding boxes are inside of the image (check notes for the explanation). The quickest way to check is to remove , color_type="color_ignore_orientation"
from train_pipeline
and test_pipeline
and check if the issue persists.learning rate was to high i removed color_type="color_ignore_orientation
and training again thanks.
@.***
On Wed, Oct 6, 2021, 5:40 AM Yunus Cobanoglu @.***> wrote:
Closed #500 https://github.com/open-mmlab/mmocr/issues/500.
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/open-mmlab/mmocr/issues/500#event-5418207561, or unsubscribe https://github.com/notifications/unsubscribe-auth/ATUUQHVFJS5WQMCDYZFEMA3UFQKSRANCNFSM5EKC344A .
2.
train_pipeline
andtest_pipeline
where is the train_pipeline and test_pipeline?
FCENet with dcvn2 leads to loss beeing NaN. If instead FCNet without dcnv2 is used it starts out with loss beeing a number but after a couple of epoch it also is NaN.
Reproduction
I will post the code for training (
train.py
) and the two configurations file one for fcenet_dcvn2 (fcenet_dcvn.py
) one config for fcenet_no_dcvn2 (fcenet_no_dcvn.py
). The only difference is the dcvn2 part.Can this be an issue with my data ? I created my own
coco.json
icdar2015 style. The only difference is that the gt_image.txt have bounding boxes which are just x,y,width,height. But the json files should be similar to the ones created by you from gt_image.txt which have x1,y1,x2,y2,x3,y3,x4,y4. Are the gt_.txt used in training ? I was expecting the ground truth labels coming from the json file not from the .txt files. In that case this is a problem unrelated to my data/json files.Error traceback
fcenet_dcvn2.py
Running
train.py
and changing the config (from cuneiform_sign_detection.configs.fcenet_default import fcenet_no_dcvn as model_config
) and checkpoint (checkpoint=None
) and loss is a number but after a couple of epochs its also Nan:fcenet_dcvn2.py
Train.py
fcenet_dcvn.py
fcenet_no_dcvn.py