Skip to content

Problem about the GCN results of Webnlg task #9

Description

@ghostyoona

I'm interest in your paper, thank you for your work.
GCN result in Table 1 of the paper shows 53.5 bleu score. But I can't get the score(highest score is 51.75) after running the project serval times(I have tried to change the batch_size and use gcn gates).
My procedure lists as follow:
(1)Using preprocess.py to preprocess by train-delex files and dev-delex files, e.g train-webnlg-all-delex-src-nodes.txt.

(2)Then I run the train.py. python3 train.py -data data/gcn_exp -save_model data/tmp_ -rnn_size 256 -word_vec_size 256 -layers 1 -epochs 30 -optim adam -learning_rate 0.001 -encoder_type gcn -gcn_num_inputs 256 -gcn_num_units 256 -gcn_in_arcs -gcn_out_arcs -gcn_num_layers 4 -gcn_num_labels 5 -gcn-residual residual -seed 42(43 and 44)

(3)Using translate.py to generate text by test-delex files and relexicalise the text.

(4)Using calculate_bleu_dev.sh to evaluation relexicalised text with only three test-notdelex references(test-all-notdelex-reference0.lex, test-all-notdelex-reference1.lex, test-all-notdelex-reference2.lex).

Is my procedure right? Could you please provide me a exact hyparameters?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions