def __init__(self, config: Dict, vocab: Optional[Vocabulary] = None):
super().__init__(vocab=vocab or vocabulary.create_empty_vocabulary())
# saves the config in the pl checkpoints
self.save_hyperparameters("config")
config = PipelineConfiguration.from_dict(config)
tokenizer = config.build_tokenizer()
featurizer = config.features.compile_featurizer(tokenizer)
embedder = config.build_embedder(self.vocab)
head = config.head.compile(backbone=ModelBackbone(
self.vocab,
featurizer=featurizer,
embedder=embedder,
encoder=config.encoder,
))
self.name = config.name
self._head = None
self.set_head(head)
self.file_path: Optional[str] = None
self.optimizer: Optional[torch.optim.Optimizer] = None
# The lr_scheduler dict follows the Lightning format:
# https://pytorch-lightning.readthedocs.io/en/stable/common/optimizers.html#learning-rate-scheduling
self.lr_scheduler: Optional[Dict] = None
self.best_metrics: Optional[Dict[str, torch.Tensor]] = None
# This is set by our trainer to figure out the best_metrics
# what metric to monitor?
self.monitor: Optional[str] = None
# shall the metric increase ("max") or decrease ("min")?
self.monitor_mode: Optional[str] = None
def test_pipeline_without_word_features():
tokenizer_config = TokenizerConfiguration()
char_features = CharFeatures(
embedding_dim=2,
encoder={
"type": "gru",
"hidden_size": 2,
"num_layers": 1,
"bidirectional": True,
},
dropout=0.1,
)
features_config = FeaturesConfiguration(char=char_features)
encoder_spec = Seq2SeqEncoderConfiguration(type="gru",
hidden_size=2,
num_layers=1,
bidirectional=True)
head_spec = TaskHeadConfiguration(
type="TextClassification",
labels=["duplicate", "not_duplicate"],
pooler={"type": "boe"},
)
pipeline_config = PipelineConfiguration(
name="no_word_features",
head=head_spec,
features=features_config,
tokenizer=tokenizer_config,
encoder=encoder_spec,
)
pl = Pipeline.from_config(pipeline_config)
assert "word" not in pl.backbone.featurizer.indexer
assert "char" in pl.backbone.featurizer.indexer
def from_config(
cls,
config: Union[PipelineConfiguration, dict],
) -> "Pipeline":
"""Creates a pipeline from a `PipelineConfiguration` object or a configuration dictionary
Parameters
----------
config
A `PipelineConfiguration` object or a configuration dict
Returns
-------
pipeline
A configured pipeline
"""
if isinstance(config, PipelineConfiguration):
config = config.as_dict()
model = PipelineModel(config=config)
if not isinstance(model, PipelineModel):
raise TypeError(f"Cannot load model. Wrong format of {model}")
cls._add_transformers_vocab_if_needed(model)
return cls(model, PipelineConfiguration.from_dict(config))
def from_config(
cls,
config: Union[PipelineConfiguration, dict],
vocab_path: Optional[str] = None,
) -> "Pipeline":
"""Creates a pipeline from a `PipelineConfiguration` object or a configuration dictionary
Parameters
----------
config: `Union[PipelineConfiguration, dict]`
A `PipelineConfiguration` object or a configuration dict
vocab_path: `Optional[str]`
If provided, the pipeline vocabulary will be loaded from this path
Returns
-------
pipeline: `Pipeline`
A configured pipeline
"""
if isinstance(config, dict):
config = PipelineConfiguration.from_dict(config)
model = PipelineModel.from_params(
Params({"config": config}),
vocab=Vocabulary.from_files(vocab_path)
if vocab_path is not None else None,
)
if not isinstance(model, PipelineModel):
raise TypeError(f"Cannot load model. Wrong format of {model}")
cls._add_transformers_vocab_if_needed(model)
return cls(model, config)
def test_pipeline_config(pipeline_yaml):
tokenizer_config = TokenizerConfiguration(
text_cleaning={"rules": ["strip_spaces"]}, use_spacy_tokens=True)
word_features = WordFeatures(embedding_dim=2, lowercase_tokens=True)
char_features = CharFeatures(
embedding_dim=2,
encoder={
"type": "gru",
"hidden_size": 2,
"num_layers": 1,
"bidirectional": True,
},
dropout=0.1,
)
features_config = FeaturesConfiguration(word=word_features,
char=char_features)
encoder_spec = Seq2SeqEncoderConfiguration(type="gru",
hidden_size=2,
num_layers=1,
bidirectional=True)
head_spec = TaskHeadConfiguration(
type=TextClassification,
labels=["duplicate", "not_duplicate"],
pooler={"type": "boe"},
)
pipeline_config = PipelineConfiguration(
name="test_pipeline_config",
head=head_spec,
features=features_config,
tokenizer=tokenizer_config,
encoder=encoder_spec,
)
pl = Pipeline.from_config(pipeline_config)
pl_yaml = Pipeline.from_yaml(pipeline_yaml)
assert pl.named_trainable_parameters == pl_yaml.named_trainable_parameters
assert pl.num_trainable_parameters == pl_yaml.num_trainable_parameters
assert pl.num_parameters == pl_yaml.num_parameters
sample_text = "My simple text"
for instance in [
pl.backbone.featurizer(sample_text),
pl_yaml.backbone.featurizer(sample_text),
]:
for key, value in instance.items():
assert key == "record"
assert isinstance(value, ListField)
assert len(value) == 1
for text in value:
assert isinstance(text, TextField)
assert all(map(lambda t: isinstance(t, Token), text.tokens))
assert sample_text == " ".join([t.text for t in text.tokens])
def from_yaml(cls, path: str) -> "Pipeline":
"""Creates a pipeline from a config yaml file
Parameters
----------
path
The path to a YAML configuration file
Returns
-------
pipeline
A configured pipeline
"""
pipeline_configuration = PipelineConfiguration.from_yaml(path)
return cls.from_config(pipeline_configuration)
def from_yaml(cls,
path: str,
vocab_path: Optional[str] = None) -> "Pipeline":
"""Creates a pipeline from a config yaml file
Parameters
----------
path : `str`
The path to a YAML configuration file
vocab_path : `Optional[str]`
If provided, the pipeline vocab will be loaded from this path
Returns
-------
pipeline: `Pipeline`
A configured pipeline
"""
pipeline_configuration = PipelineConfiguration.from_yaml(path)
return cls.from_config(pipeline_configuration, vocab_path=vocab_path)
def from_config(
cls,
config: Union[PipelineConfiguration, dict],
vocab_path: Optional[str] = None,
) -> "Pipeline":
"""Creates a pipeline from a `PipelineConfiguration` object or a configuration dictionary
Parameters
----------
config: `Union[PipelineConfiguration, dict]`
A `PipelineConfiguration` object or a configuration dict
vocab_path: `Optional[str]`
If provided, the pipeline vocabulary will be loaded from this path
Returns
-------
pipeline: `Pipeline`
A configured pipeline
"""
if isinstance(config, dict):
config = PipelineConfiguration.from_dict(config)
return _BlankPipeline(config=config,
vocab=vocabulary.load_vocabulary(vocab_path))
def _config_from_archive(archive: Archive) -> PipelineConfiguration:
config = archive.config["model"]["config"]
return PipelineConfiguration.from_params(config)