search

Academic-Hammer / SciTSR

Table structure recognition dataset of the paper: Complicated Table Structure Recognition
https://arxiv.org/pdf/1908.04729.pdf
MIT License
350 stars 57 forks source link

Some problems on reproducing #17

Closed Pyten closed 4 years ago

Pyten commented 4 years ago

Hi, thanks for your great work! During reproducing this work, I have encountered several problems. First, since there was no "rel" dir in the test data dir, so I just tried to produce it with rel_gen.py you provided, during which a lot "no match" errors were printed. Second, assertion error that "test_path/json" is not dir occurred when running the train.py. Then I just change the "with_cells" parameter to False when creating test dataset. So I am wondering whether I was doing right and were there some procedures I missed? And, what is the difference between with_cells=True and with_cells=False? Btw, does the "json" dir in your code mostly mean the "structure" dir containing json files? Please give me some hints. Thanks again!

CZWin32768 commented 4 years ago

Yes, the "json" dir in our code means the "structure" dir in the released SciTSR dataset.

I rename this folder from "json" to "structure" to make it more meaningful, but I forgot to update the related code at here.

A walkaround is to rename the "structure" folder to "json".

CZWin32768 commented 4 years ago

with_cells means to load structure labels.

The model is trained on the constructed graphs, so it is ok to set with_cells as False. During the evaluation, with_cells should be set as True to load structure labels.

Pyten commented 4 years ago

@CZWin32768 Thanks for your detailed replay. During reproducing your work, I got 98%acc after 15 epochs, is this similar to your result? And if I want to evaluate Precision and Recall during training, how should I do it? I noticed there is eval_relations API for evaluation, but can I directly use it with label and model output? Please offer some tips. Thanks again!

kbrajwani commented 3 years ago

@CZWin32768 @Pyten hey can you help me to inference image through trained model. I have trained model now i have new image so what steps i can perform to get the table structure.