Getting alignment/attention during translation in OpenNMT-py

Question

Does anyone know how to get the alignments weights when translating in Opennmt-py? Usually the only output are the resulting sentences and I have tried to find a debugging flag or similar for the attention weights. So far, I have been unsuccessful.

Answer 1

You can get the attention matrices. Note that it is not the same as alignment which is a term from statistical (not neural) machine translation.

There is a thread on github discussing it. Here is a snippet from the discussion. When you get the translations from the mode, the attentions are in the attn field.

import onmt
import onmt.io
import onmt.translate
import onmt.ModelConstructor
from collections import namedtuple

# Load the model.
Opt = namedtuple('Opt', ['model', 'data_type', 'reuse_copy_attn', "gpu"])

opt = Opt("PATH_TO_SAVED_MODEL", "text", False, 0)
fields, model, model_opt =  onmt.ModelConstructor.load_test_model(
    opt, {"reuse_copy_attn" : False})

# Test data
data = onmt.io.build_dataset(
    fields, "text", "PATH_TO_DATA", None, use_filter_pred=False)
data_iter = onmt.io.OrderedIterator(
    dataset=data, device=0,
    batch_size=1, train=False, sort=False,
    sort_within_batch=True, shuffle=False)

# Translator
translator = onmt.translate.Translator(
    model, fields, beam_size=5, n_best=1,
    global_scorer=None, cuda=True)

builder = onmt.translate.TranslationBuilder(
   data, translator.fields, 1, False, None)

batch = next(data_iter)
batch_data = translator.translate_batch(batch, data)
translations = builder.from_batch(batch_data)
translations[0].attn # <--- here are the attentions