[WIP] Support source and target features

data/data_features/src-test-with-feats.txt

@@ -0,0 +1 @@
+she￨C is￨B a￨A hard-working.￨B

data/data_features/src-train-with-feats.txt

@@ -0,0 +1,3 @@
+however,￨A according￨A to￨A the￨A logs,￨B she￨A is￨A a￨A hard-working.￨C
+however,￨A according￨B to￨C the￨D logs,￨E
+she￨C is￨B a￨A hard-working.￨B

data/data_features/src-val-with-feats.txt

@@ -0,0 +1 @@
+she￨C is￨B a￨A hard-working.￨B

data/data_features/tgt-test-with-feats.txt

@@ -0,0 +1 @@
+she￨C is￨B a￨A hard-working.￨B

data/data_features/tgt-test.txt

@@ -0,0 +1 @@
+she is a hard-working.

data/data_features/tgt-train-with-feats.txt

@@ -0,0 +1,3 @@
+however,￨A according￨A to￨A the￨A logs,￨B she￨A is￨A a￨A hard-working.￨C
+however,￨A according￨B to￨C the￨D logs,￨E
+she￨C is￨B a￨A hard-working.￨B

data/data_features/tgt-val-with-feats.txt

@@ -0,0 +1 @@
+she￨C is￨B a￨A hard-working.￨B

data/features_configs/source_and_target_features.yaml

@@ -0,0 +1,20 @@
+# Corpus opts:
+data:
+    corpus_1:
+        path_src: data/data_features/src-train-with-feats.txt
+        path_tgt: data/data_features/tgt-train-with-feats.txt
+        transforms: [inferfeats]
+    corpus_2:
+        path_src: data/data_features/src-train.txt
+        path_tgt: data/data_features/tgt-train.txt
+        transforms: [inferfeats]
+    valid:
+        path_src: data/data_features/src-val-with-feats.txt
+        path_tgt: data/data_features/tgt-val-with-feats.txt
+        transforms: [inferfeats]
+
+# # Feats options
+n_src_feats: 1
+n_tgt_feats: 1
+src_feats_defaults: "0"
+tgt_feats_defaults: "1"

data/features_configs/source_features_only.yaml

@@ -0,0 +1,18 @@
+# Corpus opts:
+data:
+    corpus_1:
+        path_src: data/data_features/src-train-with-feats.txt
+        path_tgt: data/data_features/tgt-train.txt
+        transforms: [inferfeats]
+    corpus_2:
+        path_src: data/data_features/src-train.txt
+        path_tgt: data/data_features/tgt-train.txt
+        transforms: [inferfeats]
+    valid:
+        path_src: data/data_features/src-val-with-feats.txt
+        path_tgt: data/data_features/tgt-val.txt
+        transforms: [inferfeats]
+
+# # Feats options
+n_src_feats: 1
+src_feats_defaults: "0"

data/features_configs/target_features_only.yaml

@@ -0,0 +1,18 @@
+# Corpus opts:
+data:
+    corpus_1:
+        path_src: data/data_features/src-train.txt
+        path_tgt: data/data_features/tgt-train-with-feats.txt
+        transforms: [inferfeats]
+    corpus_2:
+        path_src: data/data_features/src-train.txt
+        path_tgt: data/data_features/tgt-train.txt
+        transforms: [inferfeats]
+    valid:
+        path_src: data/data_features/src-val.txt
+        path_tgt: data/data_features/tgt-val-with-feats.txt
+        transforms: [inferfeats]
+
+# # Feats options
+n_tgt_feats: 1
+tgt_feats_defaults: "0"

onmt/bin/build_vocab.py

@@ -9,7 +9,7 @@
 from onmt.inputters.text_utils import process
 from onmt.transforms import make_transforms, get_transforms_cls
 from onmt.constants import CorpusName, CorpusTask
-from collections import Counter, defaultdict
+from collections import Counter
 import multiprocessing as mp
 
 
@@ -42,19 +42,26 @@ def write_files_from_queues(sample_path, queues):
 
 # Just for debugging purposes
 # It appends features to subwords when dumping to file
-def append_features_to_example(example, features):
-    ex_toks = example.split(' ')
-    feat_toks = features.split(' ')
-    toks = [f"{subword}￨{feat}" for subword, feat in
-            zip(ex_toks, feat_toks)]
-    return " ".join(toks)
+def append_features_to_text(text, features):
+    text_tok = text.split(' ')
+    feats_tok = [x.split(' ') for x in features]
+
+    pretty_toks = []
+    for tok, *feats in zip(text_tok, *feats_tok):
+        feats = '￨'.join(feats)
+        if feats:
+            pretty_toks.append(f"{tok}￨{feats}")
+        else:
+            pretty_toks.append(tok)
+    return " ".join(pretty_toks)
 
 
 def build_sub_vocab(corpora, transforms, opts, n_sample, stride, offset):
     """Build vocab on (strided) subpart of the data."""
     sub_counter_src = Counter()
     sub_counter_tgt = Counter()
-    sub_counter_src_feats = defaultdict(Counter)
+    sub_counter_src_feats = [Counter() for _ in range(opts.n_src_feats)]
+    sub_counter_tgt_feats = [Counter() for _ in range(opts.n_tgt_feats)]
     datasets_iterables = build_corpora_iters(
         corpora, transforms, opts.data,
         skip_empty_level=opts.skip_empty_level,
@@ -68,28 +75,35 @@ def build_sub_vocab(corpora, transforms, opts, n_sample, stride, offset):
                 if opts.dump_samples:
                     build_sub_vocab.queues[c_name][offset].put("blank")
                 continue
-            src_line, tgt_line = (maybe_example['src']['src'],
-                                  maybe_example['tgt']['tgt'])
-            src_line_pretty = src_line
-            for feat_name, feat_line in maybe_example["src"].items():
-                if feat_name not in ["src", "src_original"]:
-                    sub_counter_src_feats[feat_name].update(
-                        feat_line.split(' '))
-                    if opts.dump_samples:
-                        src_line_pretty = append_features_to_example(
-                            src_line_pretty, feat_line)
+            src_line = maybe_example.src
+            tgt_line = maybe_example.tgt
+            src_feats_lines = maybe_example.src_feats
+            tgt_feats_lines = maybe_example.tgt_feats
+
             sub_counter_src.update(src_line.split(' '))
             sub_counter_tgt.update(tgt_line.split(' '))
+            for i in range(opts.n_src_feats):
+                sub_counter_src_feats[i].update(src_feats_lines[i].split(' '))
+            for i in range(opts.n_tgt_feats):
+                sub_counter_tgt_feats[i].update(tgt_feats_lines[i].split(' '))
+
             if opts.dump_samples:
+                src_pretty_line = append_features_to_text(
+                    src_line, src_feats_lines)
+                tgt_pretty_line = append_features_to_text(
+                    tgt_line, tgt_feats_lines)
                 build_sub_vocab.queues[c_name][offset].put(
-                    (i, src_line_pretty, tgt_line))
+                    (i, src_pretty_line, tgt_pretty_line))
             if n_sample > 0 and ((i+1) * stride + offset) >= n_sample:
                 if opts.dump_samples:
                     build_sub_vocab.queues[c_name][offset].put("break")
                 break
         if opts.dump_samples:
             build_sub_vocab.queues[c_name][offset].put("break")
-    return sub_counter_src, sub_counter_tgt, sub_counter_src_feats
+    return (sub_counter_src,
+            sub_counter_tgt,
+            sub_counter_src_feats,
+            sub_counter_tgt_feats)
 
 
 def init_pool(queues):
@@ -113,7 +127,8 @@ def build_vocab(opts, transforms, n_sample=3):
     corpora = get_corpora(opts, task=CorpusTask.TRAIN)
     counter_src = Counter()
     counter_tgt = Counter()
-    counter_src_feats = defaultdict(Counter)
+    counter_src_feats = [Counter() for _ in range(opts.n_src_feats)]
+    counter_tgt_feats = [Counter() for _ in range(opts.n_tgt_feats)]
     from functools import partial
     queues = {c_name: [mp.Queue(opts.vocab_sample_queue_size)
                        for i in range(opts.num_threads)]
@@ -130,14 +145,18 @@ def build_vocab(opts, transforms, n_sample=3):
         func = partial(
             build_sub_vocab, corpora, transforms,
             opts, n_sample, opts.num_threads)
-        for sub_counter_src, sub_counter_tgt, sub_counter_src_feats in p.imap(
-                func, range(0, opts.num_threads)):
+        for (sub_counter_src, sub_counter_tgt,
+             sub_counter_src_feats, sub_counter_tgt_feats) \
+                in p.imap(func, range(0, opts.num_threads)):
             counter_src.update(sub_counter_src)
             counter_tgt.update(sub_counter_tgt)
-            counter_src_feats.update(sub_counter_src_feats)
+            for i in range(opts.n_src_feats):
+                counter_src_feats[i].update(sub_counter_src_feats[i])
+            for i in range(opts.n_tgt_feats):
+                counter_tgt_feats[i].update(sub_counter_tgt_feats[i])
     if opts.dump_samples:
         write_process.join()
-    return counter_src, counter_tgt, counter_src_feats
+    return counter_src, counter_tgt, counter_src_feats, counter_tgt_feats
 
 
 def build_vocab_main(opts):
@@ -163,13 +182,16 @@ def build_vocab_main(opts):
     transforms = make_transforms(opts, transforms_cls, None)
 
     logger.info(f"Counter vocab from {opts.n_sample} samples.")
-    src_counter, tgt_counter, src_feats_counter = build_vocab(
+    (src_counter, tgt_counter,
+     src_feats_counter, tgt_feats_counter) = build_vocab(
         opts, transforms, n_sample=opts.n_sample)
 
-    logger.info(f"Counters src:{len(src_counter)}")
-    logger.info(f"Counters tgt:{len(tgt_counter)}")
-    for feat_name, feat_counter in src_feats_counter.items():
-        logger.info(f"Counters {feat_name}:{len(feat_counter)}")
+    logger.info(f"Counters src: {len(src_counter)}")
+    logger.info(f"Counters tgt: {len(tgt_counter)}")
+    for i, feat_counter in enumerate(src_feats_counter):
+        logger.info(f"Counters src feat_{i}: {len(feat_counter)}")
+    for i, feat_counter in enumerate(tgt_feats_counter):
+        logger.info(f"Counters tgt feat_{i}: {len(feat_counter)}")
 
     def save_counter(counter, save_path):
         check_path(save_path, exist_ok=opts.overwrite, log=logger.warning)
@@ -186,8 +208,10 @@ def save_counter(counter, save_path):
         save_counter(src_counter, opts.src_vocab)
         save_counter(tgt_counter, opts.tgt_vocab)
 
-    for k, v in src_feats_counter.items():
-        save_counter(v, opts.src_feats_vocab[k])
+    for i, c in enumerate(src_feats_counter):
+        save_counter(c, f"{opts.src_vocab}_feat{i}")
+    for i, c in enumerate(tgt_feats_counter):
+        save_counter(c, f"{opts.tgt_vocab}_feat{i}")
 
 
 def _get_parser():

onmt/decoders/ensemble.py

@@ -99,14 +99,31 @@ def forward(self, hidden, attn=None, src_map=None):
         by averaging distributions from models in the ensemble.
         All models in the ensemble must share a target vocabulary.
         """
-        distributions = torch.stack(
-                [mg(h) if attn is None else mg(h, attn, src_map)
-                 for h, mg in zip(hidden, self.model_generators)]
-            )
+
+        distributions, feats_distributions = [], []
+        n_feats = len(self.model_generators[0].feats_generators)
+        for h, mg in zip(hidden, self.model_generators):
+            scores, feats_scores = \
+                (mg(h) if attn is None else mg(h, attn, src_map))
+            distributions.append(scores)
+            feats_distributions.append(feats_scores)
+
+        distributions = torch.stack(distributions)
+
+        stacked_feats_distributions = []
+        for i in range(n_feats):
+            stacked_feats_distributions.append(
+                torch.stack([feat_distribution[i]
+                             for feat_distribution in feats_distributions
+                             for i in range(n_feats)]))
+
         if self._raw_probs:
-            return torch.log(torch.exp(distributions).mean(0))
+            return (torch.log(torch.exp(distributions).mean(0)),
+                    [torch.log(torch.exp(d).mean(0))
+                     for d in stacked_feats_distributions])
         else:
-            return distributions.mean(0)
+            return (distributions.mean(0),
+                    [d.mean(0) for d in stacked_feats_distributions])
 
 
 class EnsembleModel(NMTModel):

onmt/inputters/__init__.py

@@ -4,15 +4,14 @@
 e.g., from a line of text to a sequence of vectors.
 """
 from onmt.inputters.inputter import build_vocab, IterOnDevice
-from onmt.inputters.text_utils import text_sort_key, process,\
-    numericalize, tensorify
+from onmt.inputters.text_utils import text_sort_key, process, tensorify
 from onmt.inputters.text_corpus import ParallelCorpus, ParallelCorpusIterator
 from onmt.inputters.dynamic_iterator import MixingStrategy, SequentialMixer,\
     WeightedMixer, DynamicDatasetIter
 
 
 __all__ = ['IterOnDevice', 'build_vocab', 'text_sort_key',
-           'process', 'numericalize', 'tensorify',
+           'process', 'tensorify',
            'ParallelCorpus', 'ParallelCorpusIterator',
            'MixingStrategy', 'SequentialMixer', 'WeightedMixer',
            'DynamicDatasetIter']

onmt/inputters/dynamic_iterator.py

@@ -4,7 +4,7 @@
 from onmt.constants import CorpusTask
 from onmt.inputters.text_corpus import get_corpora, build_corpora_iters
 from onmt.inputters.text_utils import text_sort_key, process,\
-    numericalize, tensorify, _addcopykeys
+    tensorify
 from onmt.transforms import make_transforms
 from onmt.utils.logging import init_logger, logger
 from onmt.utils.misc import RandomShuffler
@@ -209,8 +209,9 @@ def _tuple_to_json_with_tokIDs(self, tuple_bucket):
         for example in tuple_bucket:
             if example is not None:
                 if self.copy:
-                    example = _addcopykeys(self.vocabs, example)
-                bucket.append(numericalize(self.vocabs, example))
+                    example.addcopykeys(self.vocabs)
+                example.numericalize(self.vocabs)
+                bucket.append(example)
         return bucket
 
     def _bucketing(self):
@@ -252,11 +253,8 @@ def batch_size_fn(nbsents, maxlen):
 
         minibatch, maxlen, size_so_far, seen = [], 0, 0, []
         for ex in data:
-            if (
-                   (ex['src']['src'] not in seen) or
-                   (self.task != CorpusTask.TRAIN)
-            ):
-                seen.append(ex['src']['src'])
+            if ((ex.src not in seen) or (self.task != CorpusTask.TRAIN)):
+                seen.append(ex.src)
                 minibatch.append(ex)
                 nbsents = len(minibatch)
                 maxlen = max(text_sort_key(ex), maxlen)