From 21be4a378eb03c7545f6f58f1b6daba313cc8555 Mon Sep 17 00:00:00 2001
From: Simon Will <will@cl.uni-heidelberg.de>
Date: Tue, 6 Apr 2021 13:07:29 +0200
Subject: [PATCH] Make training work by using non-determinstic random shuffler
---
joeynmt_server/joey_model.py | 5 ++++-
joeynmt_server/trainer.py | 3 +++
joeynmt_server/utils/batching.py | 13 ++++++++++++-
joeynmt_server/views/errors.py | 1 -
joeynmt_server/views/feedback.py | 4 ++--
5 files changed, 21 insertions(+), 5 deletions(-)
diff --git a/joeynmt_server/joey_model.py b/joeynmt_server/joey_model.py
index d7545f8..f944aa8 100644
--- a/joeynmt_server/joey_model.py
+++ b/joeynmt_server/joey_model.py
@@ -217,7 +217,10 @@ class JoeyModel:
for i, batch in enumerate(batches):
logging.info('Training on batch {}.'.format(i + 1))
- logging.info('IDs in batch: {}'.format(batch.id))
+ logging.info('Size of batch: {}'.format(len(batch)))
+ if hasattr(batch, 'id'):
+ logging.info('IDs in batch: {}'.format(batch.id))
+
joey_batch = Batch(batch, self.model.pad_index,
use_cuda=trainer.use_cuda)
trainer._train_step(joey_batch)
diff --git a/joeynmt_server/trainer.py b/joeynmt_server/trainer.py
index ee08747..f1df7ee 100644
--- a/joeynmt_server/trainer.py
+++ b/joeynmt_server/trainer.py
@@ -74,6 +74,9 @@ def train(config_basename, smallest_usage_count, segment_1, segment_2):
train_iterator = merge_iterators(*iterators)
def increment_train_usages(batch):
+ if not hasattr(batch, 'id'):
+ return
+
ids = {int(id_) for id_ in batch.id if id_ >= 0}
existing_usages = {
usage.feedback_id: usage
diff --git a/joeynmt_server/utils/batching.py b/joeynmt_server/utils/batching.py
index 8595f2f..dcc333f 100644
--- a/joeynmt_server/utils/batching.py
+++ b/joeynmt_server/utils/batching.py
@@ -1,6 +1,8 @@
import itertools
+import random
from torchtext.data import Batch, BucketIterator, Dataset, Example, Field
+from torchtext.data.utils import RandomShuffler
ID_FIELD = Field(sequential=False, use_vocab=False, batch_first=True)
@@ -21,12 +23,21 @@ def make_dataset(src, src_field, trg_field=None, trg=None, ids=None):
return dataset
+class MyRandomShuffler(RandomShuffler):
+
+ def __call__(self, data):
+ return random.sample(data, len(data))
+
+
class MyBucketIterator(BucketIterator):
- def __init__(self, *args, max_epochs=None, **kwargs):
+ def __init__(self, *args, max_epochs=None, deterministic=False, **kwargs):
self.max_epochs = max_epochs
super().__init__(*args, **kwargs)
+ if not deterministic:
+ self.random_shuffler = MyRandomShuffler()
+
def iter_batches_as_lists(self):
"""Mostly a copy of BucketIterator.__iter__, but yielding the
batches as lists of sentences instead of Batch objects.
diff --git a/joeynmt_server/views/errors.py b/joeynmt_server/views/errors.py
index f8a78f8..434d489 100644
--- a/joeynmt_server/views/errors.py
+++ b/joeynmt_server/views/errors.py
@@ -1,7 +1,6 @@
import traceback
from flask import current_app, jsonify, render_template, request
-from flask_login import current_user
@current_app.errorhandler(403)
diff --git a/joeynmt_server/views/feedback.py b/joeynmt_server/views/feedback.py
index 50b6d51..06e7a40 100644
--- a/joeynmt_server/views/feedback.py
+++ b/joeynmt_server/views/feedback.py
@@ -109,7 +109,7 @@ def query_feedback():
'id': piece.id, 'created': piece.created.isoformat(timespec='seconds'),
'nl': piece.nl, 'correct_lin': piece.correct_lin,
'original_model': piece.model, 'original_lin': piece.system_lin,
- 'parent_id': piece.parent_id,
+ 'parent_id': piece.parent_id, 'split': piece.split,
'model': model, 'model_lin': model_lin
}
joined_feedback.append(joined)
@@ -130,7 +130,7 @@ def edit_feedback():
elif not isinstance(data['id'], int):
return jsonify({'error': 'Feedback id is not an int'}), 400
elif not {'editor_id', 'id', 'nl', 'system_lin', 'correct_lin',
- 'model'}.issuperset(data.keys()):
+ 'model', 'split'}.issuperset(data.keys()):
return jsonify({'error': 'Illegal keys given'}), 400
id = data.pop('id')
--
GitLab