lquirosd / order_relation_operator Goto Github PK

In trying to reproduce, I stumbled across several problems:

• in get_data.sh, OHG and ABP datasets have been deactivated (but they are needed for run.sh)
• in run.sh, the CLI of MLP.py has changed
• in run.sh, the file RDF.py is nowhere to be found (I guess it should be RTC.py)
• in the training data PAGE XML files, some regions/lines have no type in @custom; but text_line_features_from_xml falls back to TextRegion, which is not in the configured list of categories and therefore fails with ValueError: 'TextRegion' is not in list
• also, 2 regions/lines in the OHG dataset have paragraph as type, which likewise fails
• also, in the calls for the ABP dataset, (since there is no type in @custom) there should be a --level=region, otherwise no categories match at all
• in RTC.py, the import text_line_dataset.TextLineInMemory should have changed to text_line_dataset.PairsInMemoryDataset
• in metrics.py, implementation of Kendall's Tau must compensate for the case where there's just 1 sample (to prevent ZeroDivisionError)
• src/basic_ro.py now requires an additional argument lines|regions|lines_hier, and its 2nd argument contains the level in its name (e.g. test_line.pickle instead of test.pickle)
• in MLP.py, when region level was introduced, the number of features changed (from len(categories) + 6 to len(categories) + 7), so the shape of the augmentation mask must as well
• in run.sh, the plot commands do not work; it looks like plot_reading_order.py should be replaced with the new CLI plot_reading_order_2.py; but also, in both files, the calculation of the center coordinates from the feature vectors is wrong (the line-level calculation has been commented out in favour of a region-level calculation)
• how do I train a hierarchical model? In run.sh, there's no --hierarchical...

Spoiler: I have already worked on workarounds and fixes, so I am likely going to create a PR, but I would really appreciate some directions in advance.

@lquirosd hi,
did you upload the pdf of the paper