ajjimeno
commented
3 years ago Thanks Daquan. It is definitely an interesting result! We have organised a competition in ICDAR ( https://icdar2021.org/program-2/competitions/competition-on-scientific-literature-parsing), which just finished. It will be interesting to check the performance of your implementation with the performance of participants.
On Wed, Apr 7, 2021 at 1:43 PM Daquan Lin @.***> wrote:
Hi authors,
I am very interested in your work, I can't imagine EDD will perform so well with a ResNet18 backbone and two LSTM cells. Now, I am implementing the EDD with PyTorch. It can train a table image with a maximal length of tag tokens until to 1024 without CUDA out of memory in a Tesla V100 16GB after I utilized gradient checkpointing. It can achieve a ~0.9508 TEDS score in the development set. Hope it will useful to you, thanks. Here is the link https://github.com/Line290/EDD-third-party.
My Settings where different from the paper are as follows: backbone: resnext101_32x8d input image size: 640 640 feature map size: 4040*512 clip the length of characters in a cell: 100
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/ibm-aur-nlp/EDD/issues/2, or unsubscribe https://github.com/notifications/unsubscribe-auth/AA6BZDNQ4PAVM3MLAOESCRDTHPIG3ANCNFSM42P6EKVA .
Hi authors,
I am very interested in your work, I can't imagine EDD will perform so well with a ResNet18 backbone and two LSTM cells. Now, I am implementing the EDD with PyTorch. It can train a table image with a maximal length of tag tokens until to 1024 without CUDA out of memory in a Tesla V100 16GB after I utilized gradient checkpointing. It can achieve a ~0.9508 TEDS score in the development set. Hope it will useful to you, thanks. Here is the link.
My Settings where different from the paper are as follows: backbone: resnext101_32x8d input image size: 640x640 feature map size: 40x40x512 clip the length of characters in a cell: 100