Training a baseline pointer generator model with OpenNMT

paws · April 26, 2019, 5:23am

The tutorial given for summarization is that of bottom up summarization, which has a pg component but also a contextual encoder.
I wanted to know what part of the code should i consider looking at if i wanted to make a pointer generator model and train it on CNN/Daily mail to get a baseline ready.

guillaumekln · April 28, 2019, 8:37pm

Are you looking for this file?

github.com

OpenNMT/OpenNMT-py/blob/master/onmt/modules/copy_generator.py

import torch
import torch.nn as nn

from onmt.utils.misc import aeq
from onmt.utils.loss import LossComputeBase


def collapse_copy_scores(scores, batch, tgt_vocab, src_vocabs,
                         batch_dim=1, batch_offset=None):
    """
    Given scores from an expanded dictionary
    corresponeding to a batch, sums together copies,
    with a dictionary word when it is ambiguous.
    """
    offset = len(tgt_vocab)
    for b in range(scores.size(batch_dim)):
        blank = []
        fill = []
        batch_id = batch_offset[b] if batch_offset is not None else b
        index = batch.indices.data[batch_id]

This file has been truncated. show original