manucarbonell / gcn-form-understanding Goto Github PK

Hi,
Thank you for releasing your code. I am trying to run this code but there is no text_data file, can you please upload that or the embedding you created from fasttext. I am just trying to do some inferencing to understand the basics, it would be really helpful if you can release your model file as well.

Can you please provide the text_data.txt . It will help me in doing the inference.

Hi Manuel Carbonell,

when i try to run this code in dataset.py line 58 "self.embeddings = fasttext.train_unsupervised('text_data.txt', model='skipgram')" i am getting an error because of text_data.txt... i dont have that file...what is that file where can i get this file....or how can i create it on my own.

I found that you are calculating group score instead of entity link score for entity linking (only_entity_linking branch) - is that correct?

        groups_score = self.calc_score(g)
        
        entity_states = []
        entity_positions = []
        return groups_score#,entity_class,entity_positions,entity_link_score

While trying to run the model with the FUNSD dataset in the training script we are facing error in test_linking function due to mismatch of shape of link_scores and link_labels.

IndexError: The shape of the mask [360] at indexed 0 does not match the shape of the indexed tensor [170] at indexed 0

Hi, thank you for the great work! I'm trying to reproduce your results on the FUNSD dataset however something is troubling me.
When I checked how you computed entity recognition metrics I realized that on the inference mode connected components are computed based on the edge prediction model which makes sense. However, there is no guarantee that predicted components will match with the ground truth ones in terms of length and order. Could you please explain it further if I'm missing something?

def test_labeling(entity_class,entity_labels,threshold=0.5):
    labels = entity_labels[0][:,0]
    entity_class = torch.argmax(entity_class,dim=-1)
    true_positives = float((entity_class==labels).sum())
    
    total = int(labels.numel())
    acc =(true_positives/total)
    return acc,acc