pickable Vocab / v3.0.2 (#2268)

* pickable Vocab / v3.0.2

CHANGELOG.md

@@ -4,7 +4,16 @@
 
 ## [Unreleased]
 
-## [3.0.0](https://github.com/OpenNMT/OpenNMT-py/tree/3.0.1) (2022-11-23)
+## [3.0.2](https://github.com/OpenNMT/OpenNMT-py/tree/3.0.2) (2022-12-07)
+* pyonmttok.Vocab is now pickable. dataloader switched to spawn. (MacOS/Windows compatible)
+* fix scoring with specific metrics (BLEU, TER)
+* fix tensorboard logging
+* fix dedup in batch iterator (only for TRAIN, was happening at inference also)
+* New: Change: tgt_prefix renamed to tgt_file_prefix
+* New: tgt_prefix / src_prefix used for "prefix" Transform (onmt/transforms/misc.py)
+* New: process transforms of buckets in batches (vs per example) / faster
+
+## [3.0.1](https://github.com/OpenNMT/OpenNMT-py/tree/3.0.1) (2022-11-23)
 
 * fix dynamic scoring
 * reinstate apex.amp level O1/O2 for benchmarking

onmt/__init__.py

@@ -15,4 +15,4 @@
 __all__ = [onmt.inputters, onmt.encoders, onmt.decoders, onmt.models,
            onmt.utils, onmt.modules, "Trainer"]
 
-__version__ = "3.0.1"
+__version__ = "3.0.2"

onmt/inputters/dynamic_iterator.py

@@ -341,7 +341,7 @@ def build_dynamic_dataset_iter(opt, transforms_cls, vocabs, copy=False,
     else:
         data_loader = DataLoader(data_iter, batch_size=None,
                                  pin_memory=True,
-                                 multiprocessing_context="fork",
+                                 multiprocessing_context="spawn",
                                  num_workers=data_iter.num_workers,
                                  worker_init_fn=data_iter._init_datasets,
                                  prefetch_factor=opt.prefetch_factor)

onmt/opts.py

@@ -648,7 +648,7 @@ def _add_train_dynamic_data(parser):
               type=int, default=0,
               help="""The bucket size is incremented with this
               amount of examples (optional)""")
-    group.add("-prefetch_factor", "--prefetch_factor", type=int, default=100,
+    group.add("-prefetch_factor", "--prefetch_factor", type=int, default=200,
               help="""number of mini-batches loaded in advance to avoid the
                    GPU waiting during the refilling of the bucket.""")
 

onmt/utils/scoring_utils.py

@@ -185,11 +185,12 @@ def translate(self, model, sources, refs, gpu_rank, step, mode):
                     file.write("PRED: {}\n\n".format(preds[i]))
         # we deactivate the decoder's cache
         # as we use teacher forcing at training time.
-        for layer in model.decoder.transformer_layers:
-            layer.self_attn.layer_cache = (False,
-                                           {'keys': torch.tensor([]),
-                                            'values': torch.tensor([])})
-            layer.context_attn.layer_cache = (False,
-                                              {'keys': torch.tensor([]),
-                                               'values': torch.tensor([])})
+        if hasattr(model.decoder, 'transformer_layers'):
+            for layer in model.decoder.transformer_layers:
+                layer.self_attn.layer_cache = (False,
+                                               {'keys': torch.tensor([]),
+                                                'values': torch.tensor([])})
+                layer.context_attn.layer_cache = (False,
+                                                  {'keys': torch.tensor([]),
+                                                   'values': torch.tensor([])})
         return preds, texts_ref

setup.py

@@ -11,7 +11,7 @@
     description='A python implementation of OpenNMT',
     long_description=long_description,
     long_description_content_type='text/markdown',
-    version='3.0.1',
+    version='3.0.2',
     packages=find_packages(),
     project_urls={
         "Documentation": "http://opennmt.net/OpenNMT-py/",
@@ -27,7 +27,7 @@
         "tensorboard>=2.3",
         "flask",
         "waitress",
-        "pyonmttok>=1.34,<2",
+        "pyonmttok>=1.35,<2",
         "pyyaml",
         "sacrebleu"
     ],