def initialize(self, resource: Resources, configs: HParams):
self.resource = resource
self.word_alphabet = resource.get("word_alphabet")
self.char_alphabet = resource.get("char_alphabet")
self.ner_alphabet = resource.get("ner_alphabet")
word_embedding_table = resource.get('word_embedding_table')
self.config_model = configs.config_model
self.config_data = configs.config_data
self.normalize_func = utils.normalize_digit_word
self.device = torch.device("cuda") if torch.cuda.is_available() \
else torch.device("cpu")
utils.set_random_seed(self.config_model.random_seed)
self.model = BiRecurrentConvCRF(
word_embedding_table, self.char_alphabet.size(),
self.ner_alphabet.size(), self.config_model).to(device=self.device)
self.optim = SGD(self.model.parameters(),
lr=self.config_model.learning_rate,
momentum=self.config_model.momentum,
nesterov=True)
self.trained_epochs = 0
self.resource.update(model=self.model)
def initialize(self, resources: Resources, configs: Config):
"""
The training pipeline will run this initialization method during
the initialization phase and send resources in as parameters.
Args:
resources: The resources shared in the pipeline.
configs: configuration object for this trainer.
Returns:
"""
self.resource = resources
self.word_alphabet = resources.get("word_alphabet")
self.char_alphabet = resources.get("char_alphabet")
self.ner_alphabet = resources.get("ner_alphabet")
word_embedding_table = resources.get("word_embedding_table")
self.config_model = configs.config_model
self.config_data = configs.config_data
self.normalize_func = utils.normalize_digit_word
self.device = (
torch.device("cuda")
if torch.cuda.is_available()
else torch.device("cpu")
)
utils.set_random_seed(self.config_model.random_seed)
self.model = BiRecurrentConvCRF(
word_embedding_table,
self.char_alphabet.size(),
self.ner_alphabet.size(),
self.config_model,
).to(device=self.device)
self.optim = SGD(
self.model.parameters(),
lr=self.config_model.learning_rate,
momentum=self.config_model.momentum,
nesterov=True,
)
self.trained_epochs = 0
self.resource.update(model=self.model)
def initialize(self, resources: Resources, configs: Config):
super().initialize(resources, configs)
if resources.get("device"):
self.device = resources.get("device")
else:
self.device = torch.device('cuda') if torch.cuda.is_available() \
else torch.device('cpu')
self.resources = resources
self.ft_configs = configs
model_path = self.ft_configs.model_path
self.model_config = AutoConfig.from_pretrained(model_path)
self.tokenizer = AutoTokenizer.from_pretrained(model_path)
self.model = AutoModelForTokenClassification.from_pretrained(
model_path,
from_tf=bool(".ckpt" in model_path),
config=self.model_config)
self.model.to(self.device)
def initialize(self, resources: Resources, configs: Config):
super().initialize(resources, configs)
self.resource = resources
self.config_model = configs.config_model
self.config_data = configs.config_data
resource_path = configs.config_model.resource_dir
keys = {
"word_alphabet",
"char_alphabet",
"ner_alphabet",
"word_embedding_table",
}
missing_keys = list(keys.difference(self.resource.keys()))
self.resource.load(keys=missing_keys, path=resource_path)
self.word_alphabet = resources.get("word_alphabet")
self.char_alphabet = resources.get("char_alphabet")
self.ner_alphabet = resources.get("ner_alphabet")
word_embedding_table = resources.get("word_embedding_table")
if resources.get("device"):
self.device = resources.get("device")
else:
self.device = (torch.device("cuda") if torch.cuda.is_available()
else torch.device("cpu"))
self.normalize_func = utils.normalize_digit_word
if "model" not in self.resource.keys():
def load_model(path):
model = BiRecurrentConvCRF(
word_embedding_table,
self.char_alphabet.size(),
self.ner_alphabet.size(),
self.config_model,
)
if os.path.exists(path):
with open(path, "rb") as f:
weights = torch.load(f, map_location=self.device)
model.load_state_dict(weights)
return model
self.resource.load(keys={"model": load_model}, path=resource_path)
self.model = resources.get("model")
self.model.to(self.device)
self.model.eval()
utils.set_random_seed(self.config_model.random_seed)
def initialize(self, resource: Resources, configs: HParams):
self.define_batcher()
self.resource = resource
self.config_model = configs.config_model
self.config_data = configs.config_data
resource_path = configs.config_model.resource_dir
keys = {
"word_alphabet", "char_alphabet", "ner_alphabet",
"word_embedding_table"
}
missing_keys = list(keys.difference(self.resource.keys()))
self.resource.load(keys=missing_keys, path=resource_path)
self.word_alphabet = resource.get("word_alphabet")
self.char_alphabet = resource.get("char_alphabet")
self.ner_alphabet = resource.get("ner_alphabet")
word_embedding_table = resource.get("word_embedding_table")
if resource.get("device"):
self.device = resource.get("device")
else:
self.device = torch.device('cuda') if torch.cuda.is_available() \
else torch.device('cpu')
self.normalize_func = utils.normalize_digit_word
if "model" not in self.resource.keys():
def load_model(path):
model = BiRecurrentConvCRF(word_embedding_table,
self.char_alphabet.size(),
self.ner_alphabet.size(),
self.config_model)
if os.path.exists(path):
with open(path, "rb") as f:
weights = pickle.load(f)
model.load_state_dict(weights)
return model
self.resource.load(keys={"model": load_model})
self.model = resource.get("model")
self.model.to(self.device)
self.model.eval()
utils.set_random_seed(self.config_model.random_seed)
def main():
config = yaml.safe_load(open("config.yml", "r"))
config = HParams(config, default_hparams=None)
resource = Resources()
query_pipeline = Pipeline(resource=resource)
query_pipeline.set_reader(reader=MultiPackTerminalReader(),
config=config.reader)
query_pipeline.add_processor(processor=MicrosoftBingTranslator(),
config=config.translator)
query_pipeline.add_processor(processor=BertBasedQueryCreator(),
config=config.query_creator)
query_pipeline.add_processor(processor=SearchProcessor(),
config=config.indexer)
query_pipeline.add_processor(
processor=NLTKSentenceSegmenter(),
selector=NameMatchSelector(
select_name=config.indexer.response_pack_name[0]))
query_pipeline.add_processor(
processor=NLTKWordTokenizer(),
selector=NameMatchSelector(
select_name=config.indexer.response_pack_name[0]))
query_pipeline.add_processor(
processor=NLTKPOSTagger(),
selector=NameMatchSelector(
select_name=config.indexer.response_pack_name[0]))
query_pipeline.add_processor(
processor=SRLPredictor(),
config=config.SRL,
selector=NameMatchSelector(
select_name=config.indexer.response_pack_name[0]))
query_pipeline.add_processor(processor=MicrosoftBingTranslator(),
config=config.back_translator)
query_pipeline.initialize()
for m_pack in query_pipeline.process_dataset():
# update resource to be used in the next conversation
query_pack = m_pack.get_pack(config.translator.in_pack_name)
if resource.get("user_utterance"):
resource.get("user_utterance").append(query_pack)
else:
resource.update(user_utterance=[query_pack])
response_pack = m_pack.get_pack(config.back_translator.in_pack_name)
if resource.get("bot_utterance"):
resource.get("bot_utterance").append(response_pack)
else:
resource.update(bot_utterance=[response_pack])
english_pack = m_pack.get_pack("pack")
print(colored("English Translation of the query: ", "green"),
english_pack.text, "\n")
pack = m_pack.get_pack(config.indexer.response_pack_name[0])
print(colored("Retrieved Document", "green"), pack.text, "\n")
print(colored("German Translation", "green"),
m_pack.get_pack("response").text, "\n")
for sentence in pack.get(Sentence):
sent_text = sentence.text
print(colored("Sentence:", 'red'), sent_text, "\n")
print(colored("Semantic role labels:", 'red'))
for link in pack.get(PredicateLink, sentence):
parent = link.get_parent()
child = link.get_child()
print(f" - \"{child.text}\" is role {link.arg_type} of "
f"predicate \"{parent.text}\"")
print()
input(colored("Press ENTER to continue...\n", 'green'))
def main():
config = yaml.safe_load(open("config.yml", "r"))
config = HParams(config, default_hparams=None)
if not os.path.exists(config.indexer.model_dir):
print(f"Creating a new index...")
encoder = BERTEncoder(pretrained_model_name="bert-base-uncased")
encoder.to(device)
feature_original_types = {
"id": ["int64", "FixedLenFeature"],
"input_ids": ["int64", "FixedLenFeature",
config.indexer.max_seq_length],
"segment_ids": ["int64", "FixedLenFeature",
config.indexer.max_seq_length],
"text": ["str", "FixedLenFeature"]
}
hparam = {
"allow_smaller_final_batch": True,
"batch_size": config.indexer.batch_size,
"dataset": {
"data_name": "data",
"feature_original_types": feature_original_types,
"files": config.indexer.pickle_data_dir
},
"shuffle": False
}
print(f"Embedding the text using BERTEncoder...")
record_data = RecordData(hparams=hparam, device=device)
data_iterator = DataIterator(record_data)
index = EmbeddingBasedIndexer(hparams={
"index_type": "GpuIndexFlatIP",
"dim": 768,
"device": "gpu0"
})
for idx, batch in enumerate(data_iterator):
ids = batch["id"]
input_ids = batch["input_ids"]
segment_ids = batch["segment_ids"]
text = batch["text"]
_, pooled_output = get_embeddings(encoder, input_ids, segment_ids)
index.add(vectors=pooled_output,
meta_data={k.item(): v for k, v in zip(ids, text)})
if (idx + 1) % 50 == 0:
print(f"Completed {idx+1} batches of size "
f"{config.indexer.batch_size}")
index.save(path=config.indexer.model_dir)
resource = Resources()
query_pipeline = Pipeline(resource=resource)
query_pipeline.set_reader(MultiPackTerminalReader())
query_pipeline.add_processor(
processor=MachineTranslationProcessor(), config=config.translator)
query_pipeline.add_processor(
processor=QueryCreator(), config=config.query_creator)
query_pipeline.add_processor(
processor=SearchProcessor(), config=config.indexer)
query_pipeline.add_processor(
processor=NLTKSentenceSegmenter(),
selector=NameMatchSelector(select_name="doc_0"))
query_pipeline.add_processor(
processor=NLTKWordTokenizer(),
selector=NameMatchSelector(select_name="doc_0"))
query_pipeline.add_processor(
processor=NLTKPOSTagger(),
selector=NameMatchSelector(select_name="doc_0"))
query_pipeline.add_processor(
processor=SRLPredictor(), config=config.SRL,
selector=NameMatchSelector(select_name="doc_0"))
# query_pipeline.add_processor(
# processor=CoNLLNERPredictor(), config=config.NER,
# selector=NameMatchSelector(select_name="doc_0"))
query_pipeline.add_processor(
processor=MachineTranslationProcessor(), config=config.back_translator)
query_pipeline.initialize()
for m_pack in query_pipeline.process_dataset():
# update resource to be used in the next conversation
query_pack = m_pack.get_pack("query")
if resource.get("user_utterance"):
resource.get("user_utterance").append(query_pack)
else:
resource.update(user_utterance=[query_pack])
response_pack = m_pack.get_pack("response")
if resource.get("bot_utterance"):
resource.get("bot_utterance").append(response_pack)
else:
resource.update(bot_utterance=[response_pack])
english_pack = m_pack.get_pack("pack")
print(colored("English Translation of the query: ", "green"),
english_pack.text, "\n")
pack = m_pack.get_pack("doc_0")
print(colored("Retrieved Document", "green"), pack.text, "\n")
print(colored("German Translation", "green"),
m_pack.get_pack("response").text, "\n")
for sentence in pack.get(Sentence):
sent_text = sentence.text
print(colored("Sentence:", 'red'), sent_text, "\n")
print(colored("Semantic role labels:", 'red'))
for link in pack.get(PredicateLink, sentence):
parent = link.get_parent()
child = link.get_child()
print(f" - \"{child.text}\" is role {link.arg_type} of "
f"predicate \"{parent.text}\"")
print()
input(colored("Press ENTER to continue...\n", 'green'))
class Pipeline(Generic[PackType]):
r"""This controls the main inference flow of the system. A pipeline is
consisted of a set of Components (readers and processors). The data flows
in the pipeline as data packs, and each component will use or add
information to the data packs.
"""
def __init__(
self,
resource: Optional[Resources] = None,
ontology_file: Optional[str] = None,
enforce_consistency: bool = False,
do_init_type_check: bool = False,
):
r"""
Args:
resource: The ``Resources`` object, which is a global registry used
in the pipeline. Objects defined as ``Resources`` will be
passed on to the processors in the
pipeline for initialization.
ontology_file: The path to the input ontology specification file,
which should be a json file, and it should have all the entries
inside with no import as key.
enforce_consistency: This boolean determines whether the
pipeline will check the content expectations specified in each
pipeline component. Each component will check whether the input
pack contains the expected data
via checking the meta-data, and throws a
:class:`~forte.common.exception.ExpectedEntryNotFound` if it
fails. When this function is called with enforce is ``True``,
all the pipeline components would check if the input datapack
record matches
with the expected types and attributes if function
``expected_types_and_attributes`` is implemented
for the processor. For example, processor A requires entry type
of ``ft.onto.base_ontology.Sentence``, and processor B would
produce this type in the output datapack, so ``record`` function
of processor B writes the record of this type in the datapack
and processor A implements ``expected_types_and_attributes`` to
add this type. Then when the pipeline runs with
`enforce_consistency=True`, processor A would check if this
type exists in the record of the output of the
previous pipeline component.
do_init_type_check: Determine whether to check records types and
attributes during pipeline initialization. Default to `False`.
If this boolean is set to `True`, each component in the
pipeline will be validated by comparing its
``expected_types_and_attributes`` with the accumulated
``records`` from all the downstream components.
"""
self._reader: BaseReader
self._reader_config: Optional[Config] = None
# These variables defines the units in the pipeline, they should be
# of the same length
self._components: List[PipelineComponent] = []
self._selectors: List[Selector] = []
self._configs: List[Optional[Config]] = []
# Maintain a set of the pipeline components to fast check whether
# the component is already there.
self.__component_set: Set[PipelineComponent] = set()
# Will initialize at `initialize` because the processors length is
# unknown.
self._proc_mgr: ProcessManager = None # type: ignore
self.evaluator_indices: List[int] = []
# needed for evaluator
self._predict_to_gold: Dict[int, PackType] = {}
if resource is None:
self.resource = Resources()
else:
self.resource = resource
if ontology_file is None:
with resources.path("forte.ontology_specs",
"base_ontology.json") as data_path:
ontology_file = str(data_path)
if ontology_file is not None:
with open(ontology_file, "r") as f:
spec_dict = json.load(f)
self.resource.update(onto_specs_path=ontology_file)
self.resource.update(onto_specs_dict=spec_dict)
# The flag indicating whether this pipeline is initialized.
self._initialized: bool = False
# The flag indicating whether we want to enforce type consistency
# between the processors.
self._check_type_consistency: bool = False
# Create one copy of the dummy selector to reduce class creation.
self.__default_selector: Selector = DummySelector()
# needed for time profiling of pipeline
self._enable_profiling: bool = False
self._profiler: List[float] = []
self._check_type_consistency = enforce_consistency
# Indicate whether do type checking during pipeline initialization
self._do_init_type_check: bool = do_init_type_check
def enforce_consistency(self, enforce: bool = True):
r"""This function determines whether the pipeline will check
the content expectations specified in each pipeline component. This
function works with :meth:`~forte.pipeline.Pipeline.initialize` called
after itself. Each component will check whether the input pack contains
the expected data via checking the meta-data, and throws a
:class:`~forte.common.exception.ExpectedEntryNotFound` if the check
fails. The example of implementation is mentioned in the docstrings of
:meth:`~forte.pipeline.Pipeline.__init__`.
Args:
enforce: A boolean of whether to enable consistency checking
for the pipeline or not.
"""
self._check_type_consistency = enforce
def init_from_config_path(self, config_path):
r"""Read the configurations from the given path ``config_path``
and build the pipeline with the config.
Args:
config_path: A string of the configuration path, which is
is a YAML file that specify the structure and parameters of the
pipeline.
"""
configs = yaml.safe_load(open(config_path))
self.init_from_config(configs)
def init_from_config(self, configs: List):
r"""Initialized the pipeline (ontology and processors) from the
given configurations.
Args:
configs: The configs used to initialize the pipeline.
"""
is_first: bool = True
for component_config in configs:
component = create_class_with_kwargs(
class_name=component_config["type"],
class_args=component_config.get("kwargs", {}),
)
if is_first:
if not isinstance(component, BaseReader):
raise ProcessorConfigError(
"The first component of a pipeline must be a reader.")
self.set_reader(component, component_config.get("configs", {}))
is_first = False
else:
# Can be processor, caster, or evaluator
self.add(component, component_config.get("configs", {}))
def _dump_to_config(self):
r"""Serialize the pipeline to an IR(intermediate representation).
The returned IR can be passed to `init_from_config` to initialize
a pipeline.
Returns:
dict: A dictionary storing IR.
"""
configs: List[Dict] = []
configs.append({
"type":
".".join([self._reader.__module__,
type(self._reader).__name__]),
"configs":
self._reader_config.todict(),
})
for component, config in zip(self.components, self.component_configs):
configs.append({
"type":
".".join([component.__module__,
type(component).__name__]),
"configs":
config.todict(),
})
return configs
def save(self, path: str):
r"""Store the pipeline as an IR(intermediate representation) in yaml.
The path can then be passed to ``init_from_config_path`` to initialize
a pipeline. Note that calling ``init_from_config`` from a different
python environment may not work for some self defined component classes
because their module name is `__main__`.
Args:
path: The file path to save configurations.
"""
with open(path, "w") as f:
yaml.safe_dump(self._dump_to_config(), f)
def _remote_service_app(self,
service_name: str = "",
input_format: str = "string"):
r"""Return a FastAPI app that can be used to serve the pipeline.
Args:
service_name: Assign a name to the pipeline service for validation.
This will appear in the `service_name` field on default page
and can be queried and validated against the expected service
name set by user. Default to `''`.
input_format: Specify format of the input for validation. It can be
`"string"` or `"DataPack"`. This will appear in the
`input_format` field on default page and can be queried and
validated against the expected input format set by user.
Default to `"string"`.
Returns:
FastAPI: A FastAPI app for remote service.
"""
# TODO: Currently we only support the `process` function, but it can
# be extended by adding new interfaces that wrap up any Pipeline
# method. Refer to https://fastapi.tiangolo.com for more info.
app = FastAPI()
records: Optional[Dict[str, Set[str]]] = None
class RequestBody(BaseModel):
args: str = "[]"
kwargs: str = "{}"
# pylint: disable=unused-variable
@app.get("/")
def default_page():
return {
"status": "OK",
"service_name": service_name,
"input_format": input_format,
"pipeline": self._dump_to_config(),
}
@app.get("/records")
def get_records():
nonlocal records
if records is None:
# Collect records of each pipeline component for validation
records = {}
for component in [self._reader] + self.components:
component.record(records)
return {"status": "OK", "records": records}
@app.get("/expectation")
def get_expectation():
expectation: Dict[str, Set[str]] = {}
if len(self.components) > 0:
expectation = self.components[0].expected_types_and_attributes(
)
return {"status": "OK", "expectation": expectation}
@app.post("/process")
def run_pipeline(body: RequestBody):
args = json.loads(body.args)
kwargs = json.loads(body.kwargs)
result = self.process(*args, **kwargs)
return {"status": "OK", "result": result.serialize()}
# pylint: enable=unused-variable
return app
def serve(
self,
host: str = "localhost",
port: int = 8008,
service_name: str = "",
input_format: str = "string",
):
r"""Start a service of the current pipeline at a specified host
and port.
Args:
host: Port number of pipeline service.
port: Host name of pipeline service.
service_name: Assign a name to the pipeline service for validation.
This will appear in the `service_name` field on default page
and can be queried and validated against the expected service
name set by user. Default to `''`.
input_format: Specify format of the input for validation. It can be
`"string"` or `"DataPack"`. This will appear in the
`input_format` field on default page and can be queried and
validated against the expected input format set by user.
Default to `"string"`.
"""
self.initialize()
uvicorn.run(
self._remote_service_app(service_name=service_name,
input_format=input_format),
host=host,
port=port,
log_level="info",
)
def set_profiling(self, enable_profiling: bool = True):
r"""Set profiling option.
Args:
enable_profiling: A boolean of whether to enable profiling
for the pipeline or not (the default is True).
"""
self._enable_profiling = enable_profiling
def initialize(self) -> "Pipeline":
"""
This function should be called before the pipeline can be used to
process the actual data. This function will call the `initialize` of
all the components inside this pipeline.
Returns:
"""
# create EntryTree type object merged_entry_tree to store the parsed
# entry tree from ontology specification file passed in as part of
# resource and add the result to resource with key of merged_entry_tree.
merged_entry_tree = EntryTree()
if self.resource.get("onto_specs_path"):
OntologyCodeGenerator().parse_schema_for_no_import_onto_specs_file(
ontology_path=self.resource.get("onto_specs_path"),
ontology_dict=self.resource.get("onto_specs_dict"),
merged_entry_tree=merged_entry_tree,
)
self.resource.update(merged_entry_tree=merged_entry_tree)
# The process manager need to be assigned first.
self._proc_mgr = ProcessManager(len(self._components))
if self._initialized:
# The pipeline has already been initialized, so we are doing
# re-initialization here.
logging.info("Re-initializing the Pipeline.")
# Reset the flags of the components before initializing them.
self._reader.reset_flags()
for c in self._components:
c.reset_flags()
# Handle the reader.
if not self._reader.is_initialized:
self._reader.initialize(self.resource, self._reader_config)
else:
logging.info(
"The reader [%s] has already initialized, "
"will skip its initialization.",
self._reader.name,
)
if self._check_type_consistency:
self.reader.enforce_consistency(enforce=True)
else:
self.reader.enforce_consistency(enforce=False)
# Handle other components.
self.initialize_components()
self._initialized = True
# Create profiler
if self._enable_profiling:
self.reader.set_profiling(True)
self._profiler = [0.0] * len(self.components)
# Check record types and attributes of each pipeline component
if self._do_init_type_check:
current_records: Dict[str, Set[str]] = {}
self._reader.record(current_records)
for component in self.components:
if hasattr(component, "expected_types_and_attributes"):
record_types_and_attributes_check(
component.expected_types_and_attributes(
), # type: ignore
current_records,
)
if hasattr(component, "record"):
component.record(current_records) # type: ignore
return self
def initialize_components(self):
"""
This function will initialize all the components in this pipeline,
except the reader. The components are initialized in a FIFO manner
based on the order of insertion,
During initialization, the component will be configured based on its
corresponding configuration. However, if the component is already
initialized (for example, being initialized manually or used twice
in the same pipeline), the new configuration will be ignored.
The pipeline will check for type dependencies between the components
inside this pipeline, see
:func:`~forte.pipeline_component.PipelineComponent.enforce_consistency`
for more details.
"""
for component, config in zip(self.components, self.component_configs):
try:
if not component.is_initialized:
component.initialize(self.resource, config)
else:
logging.info(
"The component [%s] has already initialized, "
"will skip its initialization.",
component.name,
)
except ProcessorConfigError as e:
logging.error(
"Exception occur when initializing "
"processor %s",
component.name,
)
raise e
component.enforce_consistency(enforce=self._check_type_consistency)
def set_reader(
self,
reader: BaseReader,
config: Optional[Union[Config, Dict[str, Any]]] = None,
) -> "Pipeline":
"""
Set the reader of the pipeline. A reader is the entry point of
this pipeline, data flown into the reader will be converted to the
data pack format, and being passed onto the other components for
processing.
Args:
reader: The reader to be used of the pipeline
config: The custom configuration to be passed to the reader. If
the config is not provided, the default config defined by the
reader class will be used.
Returns:
The pipeline itself, which allows you to directly chain other
pipeline construction code afterwards, i.e., you can do:
.. code-block:: python
Pipeline().set_reader(your_reader()).add(your_processor())
"""
self._reader = reader
self._reader_config = reader.make_configs(config)
return self
@property
def reader(self) -> BaseReader:
return self._reader
@property
def components(self) -> List[PipelineComponent]:
"""
Return all the components in this pipeline, except the reader.
Returns: A list containing the components.
"""
return self._components
@property
def component_configs(self) -> List[Optional[Config]]:
"""
Return the configs related to the components, except the reader.
Returns: A list containing the components configs.
"""
return self._configs
def add(
self,
component: PipelineComponent,
config: Optional[Union[Config, Dict[str, Any]]] = None,
selector: Optional[Selector] = None,
) -> "Pipeline":
"""
Adds a pipeline component to the pipeline. The pipeline components
will form a chain based on the insertion order. The customized
`config` and `selector` (:class:`~forte.data.selector.Selector`)
will be associated with this particular component. If the `config`
or the `selector` is not provided, the default ones will be used.
Here, note that the same component instance can be added multiple
times to the pipeline. In such cases, the instance will only be
setup at the first insertion (i.e. its `initialize` function will
only be called once). The subsequent insertion of the same component
instance will not change the behavior nor the states of the instance.
Thus, a different `config` cannot be provided (should be `None`) when
added the second time, otherwise a `ProcessorConfigError` will be
thrown. In the case where one want to them to behave differently, a
different instance should be used.
Args:
component (PipelineComponent): The component to be inserted next
to the pipeline.
config (Union[Config, Dict[str, Any]): The custom configuration
to be used for the added component. Default None, which means
the `default_configs()` of the component will be used.
selector (Selector): The selector used to pick the corresponding
data pack to be consumed by the component. Default None, which
means the whole pack will be used.
Returns:
The pipeline itself, which enables one to chain the creation of
the pipeline, i.e., you can do:
.. code-block:: python
Pipeline().set_reader(your_reader()).add(
your_processor()).add(anther_processor())
"""
if isinstance(component, BaseReader):
raise ProcessFlowException(
"Reader need to be set via set_reader()")
if isinstance(component, Evaluator):
# This will ask the job to keep a copy of the gold standard.
self.evaluator_indices.append(len(self.components))
if component not in self.__component_set:
# The case where the component is not found.
self._components.append(component)
self.__component_set.add(component)
self.component_configs.append(component.make_configs(config))
else:
if config is None:
self._components.append(component)
# We insert a `None` value here just to make the config list
# to match the component list, but this config should not be
# used.
self.component_configs.append(None)
else:
raise ProcessorConfigError(
f"The same instance of a component named {component.name} "
f" has already been added to"
f" the pipeline, we do not accept a different configuration"
f" for it. If you would like to use a differently"
f" configured component, please create another instance."
f" If you intend to re-use the component instance, please"
f" do not provide the `config` (or provide a `None`).")
if selector is None:
self._selectors.append(self.__default_selector)
else:
self._selectors.append(selector)
return self
def add_gold_packs(self, pack):
r"""Add gold packs to a internal dictionary used for evaluation.
This dictionary is used by the evaluator while calling
`consume_next(...)`
Args:
pack (Dict): A key, value pair containing job.id -> gold_pack
mapping
"""
self._predict_to_gold.update(pack)
def process(self, *args, **kwargs) -> PackType:
r"""Alias for :meth:`process_one`.
Args:
args: The positional arguments used to get the initial data.
kwargs: The keyword arguments used to get the initial data.
"""
return self.process_one(*args, **kwargs)
def run(self, *args, **kwargs):
r"""Run the whole pipeline and ignore all returned DataPack. This is
mostly used when you need to run the pipeline and do not require the
output but rely on the side-effect. For example, if the pipeline
writes some data to disk.
Calling this function will automatically call the :meth:`initialize`
at the beginning, and call the :meth:`finish` at the end.
Args:
args: The positional arguments used to get the initial data.
kwargs: The keyword arguments used to get the initial data.
"""
self.initialize()
for _ in self.process_dataset(*args, **kwargs):
# Process the whole dataset ignoring the return values.
# This essentially expect the processors have side effects.
pass
self.finish()
def process_one(self, *args, **kwargs) -> PackType:
r"""Process one single data pack. This is done by only reading and
processing the first pack in the reader.
Args:
kwargs: the information needed to load the data. For example, if
:attr:`_reader` is :class:`StringReader`, this should contain a
single piece of text in the form of a string variable. If
:attr:`_reader` is a file reader, this can point to the file
path.
"""
if not self._initialized:
raise ProcessFlowException(
"Please call initialize before running the pipeline")
first_pack = []
for p in self._reader.iter(*args, **kwargs):
first_pack.append(p)
break
if len(first_pack) == 1:
results = list(self._process_packs(iter(first_pack)))
return results[0]
else:
raise ValueError("Input data source contains no packs.")
def process_dataset(self, *args, **kwargs) -> Iterator[PackType]:
r"""Process the documents in the data source(s) and return an
iterator or list of DataPacks. The arguments are directly passed
to the reader to take data from the source.
"""
if not self._initialized:
raise ProcessFlowException(
"Please call initialize before running the pipeline")
data_iter = self._reader.iter(*args, **kwargs)
return self._process_packs(data_iter)
def finish(self):
"""
Call the finish method of all pipeline component. This need to be called
explicitly to release all resources.
"""
# Report time profiling of readers and processors
if self._enable_profiling:
out_header: str = "Pipeline Time Profile\n"
out_reader: str = (f"- Reader: {self.reader.component_name}, " +
f"{self.reader.time_profile} s\n")
out_processor: str = "\n".join([
f"- Component [{i}]: {self.components[i].name}, {t} s"
for i, t in enumerate(self._profiler)
])
logger.info("%s%s%s", out_header, out_reader, out_processor)
self.reader.finish(self.resource)
for p in self.components:
p.finish(self.resource)
self._initialized = False
def __update_stream_job_status(self):
q_index = self._proc_mgr.current_queue_index
u_index = self._proc_mgr.unprocessed_queue_indices[q_index]
current_queue = self._proc_mgr.current_queue
for job_i in itertools.islice(current_queue, 0, u_index + 1):
if job_i.status == ProcessJobStatus.UNPROCESSED:
job_i.set_status(ProcessJobStatus.PROCESSED)
def __update_batch_job_status(self, component: BaseBatchProcessor):
# update the status of the jobs. The jobs which were removed from
# data_pack_pool will have status "PROCESSED" else they are "QUEUED"
q_index = self._proc_mgr.current_queue_index
u_index = self._proc_mgr.unprocessed_queue_indices[q_index]
current_queue = self._proc_mgr.current_queue
data_pool_length = len(component.batcher.data_pack_pool)
for i, job_i in enumerate(
itertools.islice(current_queue, 0, u_index + 1)):
if i <= u_index - data_pool_length:
job_i.set_status(ProcessJobStatus.PROCESSED)
else:
job_i.set_status(ProcessJobStatus.QUEUED)
def __flush_batch_job_status(self):
current_queue = self._proc_mgr.current_queue
for job in current_queue:
job.set_status(ProcessJobStatus.PROCESSED)
def _process_with_component(
self,
selector: Selector,
component: PipelineComponent,
raw_job: ProcessJob,
):
for pack in selector.select(raw_job.pack):
# First, perform the component action on the pack
try:
if isinstance(component, Caster):
# Replacing the job pack with the casted version.
raw_job.alter_pack(component.cast(pack))
elif isinstance(component, BaseBatchProcessor):
pack.set_control_component(component.name)
component.process(pack)
elif isinstance(component, Evaluator):
pack.set_control_component(component.name)
component.consume_next(pack,
self._predict_to_gold[raw_job.id])
elif isinstance(component, BaseProcessor):
# Should be BasePackProcessor:
# All other processor are considered to be
# streaming processor like this.
pack.set_control_component(component.name)
component.process(pack)
# After the component action, make sure the entry is
# added into the index.
pack.add_all_remaining_entries()
except ValueError as e:
raise ProcessExecutionException(
f"Exception occurred when running "
f"{component.name}") from e
def _process_packs(self,
data_iter: Iterator[PackType]) -> Iterator[PackType]:
r"""Process the packs received from the reader by the running through
the pipeline.
Args:
data_iter (iterator): Iterator yielding jobs that contain packs
Returns:
Yields packs that are processed by the pipeline.
"""
# pylint: disable=line-too-long
# Here is the logic for the execution of the pipeline.
# The basic idea is to yield a pack as soon as it gets processed by all
# the processors instead of waiting for later jobs to get processed.
# 1) A job can be in three status
# - UNPROCESSED
# - QUEUED
# - PROCESSED
#
# 2) Each processor maintains a queue to hold the jobs to be executed
# next.
#
# 3) In case of a BatchProcessor, a job enters into QUEUED status if the
# batch is not full according to the batcher of that processor.
# In that case, the pipeline requests for additional jobs from the
# reader and starts the execution loop from the beginning.
#
# 4) At any point, while moving to the next processor, the pipeline
# ensures that all jobs are either in QUEUED or PROCESSED status. If
# they are PROCESSED, they will be moved to the next queue. This design
# ensures that at any point, while processing the job at processor `i`,
# all the jobs in the previous queues are in QUEUED status. So whenever
# a new job is needed, the pipeline can directly request it from the
# reader instead of looking at previous queues for UNPROCESSED jobs.
#
# 5) When a processor receives a poison pack, it flushes all the
# remaining batches in its memory (this actually has no effect in
# PackProcessors) and moves the jobs including the poison pack to the
# next queue. If there is no next processor, the packs are yield.
#
# 6) The loop terminates when the last queue contains only a poison pack
#
# Here is the sample pipeline and its execution
#
# Assume 1 pack corresponds to a batch of size 1
#
# After 1st step (iteration), reading from the reader,
#
# batch_size = 2 batch_size = 2
# Reader -> B1 (BatchProcessor) -> P1 (PackProcessor) -> B2(BatchProcessor)
#
# |___________|
# |___________|
# |___________|
# |___________|
# |_J1:QUEUED_|
#
# B1 needs another pack to process job J1
#
# After 2nd step (iteration),
#
# batch_size = 2 batch_size = 2
# Reader -> B1 (BatchProcessor) -> P1 (PackProcessor) -> B2(BatchProcessor)
#
# |___________| |_______________|
# |___________| |_______________|
# |___________| |_______________|
# |___________| |_J2:UNPROCESSED_|
# |___________| |_J1:UNPROCESSED_|
#
# B1 processes both the packs, the jobs are moved to the next queue.
#
# After 3rd step (iteration),
#
# batch_size = 2 batch_size = 2
# Reader -> B1 (BatchProcessor) -> P1 (PackProcessor) -> B2(BatchProcessor)
#
# |___________| |_______________| |_______________|
# |___________| |_______________| |_______________|
# |___________| |_______________| |_______________|
# |___________| |_______________| |_______________|
# |___________| |_J2:UNPROCESSED_| |_J1:UNPROCESSED_|
#
# P1 processes the first job. However, there exists one UNPROCESSED job
# J2 in the queue. Pipeline first processes this job before moving to the
# next processor
#
# After 4th step (iteration),
#
# batch_size = 2 batch_size = 2
# Reader -> B1 (BatchProcessor) -> P1 (PackProcessor) -> B2(BatchProcessor)
#
# |___________| |_______________| |_______________|
# |___________| |_______________| |_______________|
# |___________| |_______________| |_______________|
# |___________| |_______________| |_J2:UNPROCESSED_|
# |___________| |_______________| |_J1:UNPROCESSED_|
#
#
# After 5th step (iteration),
#
# batch_size = 2 batch_size = 2
# Reader -> B1 (BatchProcessor) -> P1 (PackProcessor) -> B2(BatchProcessor)
#
# |___________| |_______________| |_______________|
# |___________| |_______________| |_______________|
# |___________| |_______________| |_______________| --> Yield J1.pack and J2.pack
# |___________| |_______________| |_______________|
# |___________| |_______________| |_______________|
if not self._initialized:
raise ProcessFlowException(
"Please call initialize before running the pipeline")
buffer = ProcessBuffer(self, data_iter)
if len(self.components) == 0:
yield from data_iter
# Write return here instead of using if..else to reduce indent.
return
while not self._proc_mgr.exhausted():
# Take the raw job from the buffer, the job status now should
# be UNPROCESSED.
raw_job: ProcessJob = next(buffer)
component_index = self._proc_mgr.current_processor_index
component = self.components[component_index]
selector: Selector = self._selectors[component_index]
current_queue_index = self._proc_mgr.current_queue_index
current_queue: Deque[ProcessJob] = self._proc_mgr.current_queue
pipeline_length = self._proc_mgr.pipeline_length
unprocessed_queue_indices = self._proc_mgr.unprocessed_queue_indices
processed_queue_indices = self._proc_mgr.processed_queue_indices
next_queue_index = current_queue_index + 1
should_yield = next_queue_index >= pipeline_length
if not raw_job.is_poison:
# Start timer
if self._enable_profiling:
start_time: float = time()
self._process_with_component(selector, component, raw_job)
# Stop timer and add to time profiler
if self._enable_profiling:
self._profiler[component_index] += time() - start_time
# Then, based on component type, handle the queue.
if isinstance(component, BaseBatchProcessor):
self.__update_batch_job_status(component)
index = unprocessed_queue_indices[current_queue_index]
# Check status of all the jobs up to "index".
for i, job_i in enumerate(
itertools.islice(current_queue, 0, index + 1)):
if job_i.status == ProcessJobStatus.PROCESSED:
processed_queue_indices[current_queue_index] = i
# There are UNPROCESSED jobs in the queue.
if index < len(current_queue) - 1:
unprocessed_queue_indices[current_queue_index] += 1
elif processed_queue_indices[current_queue_index] == -1:
# Fetch more data from the reader to process the
# first job.
unprocessed_queue_indices[current_queue_index] = len(
current_queue)
self._proc_mgr.current_processor_index = 0
self._proc_mgr.current_queue_index = -1
else:
processed_queue_index = processed_queue_indices[
current_queue_index]
# Move or yield the pack.
c_queue = list(current_queue)
for job_i in c_queue[:processed_queue_index + 1]:
if job_i.status == ProcessJobStatus.PROCESSED:
if should_yield:
if job_i.id in self._predict_to_gold:
self._predict_to_gold.pop(job_i.id)
# TODO: I don't know why these are
# marked as incompatible type by mypy.
# the same happens 3 times on every yield.
# It is observed that the pack returned
# from the `ProcessJob` is considered to
# be different from `PackType`.
yield job_i.pack # type: ignore
else:
self._proc_mgr.add_to_queue(
queue_index=next_queue_index,
job=job_i)
else:
raise ProcessFlowException(
f"The job status should be "
f"{ProcessJobStatus.PROCESSED} "
f"at this point.")
current_queue.popleft()
# Set the UNPROCESSED and PROCESSED indices.
unprocessed_queue_indices[current_queue_index] = len(
current_queue)
processed_queue_indices[current_queue_index] = -1
if should_yield:
self._proc_mgr.current_processor_index = 0
self._proc_mgr.current_queue_index = -1
else:
self._proc_mgr.current_processor_index = (
next_queue_index)
self._proc_mgr.current_queue_index = (
next_queue_index)
# Besides Batch Processors, the other component type only
# deal with one pack at a time, these include: PackProcessor
# Evaluator, Caster.
# - Move them to the next queue
else:
self.__update_stream_job_status()
index = unprocessed_queue_indices[current_queue_index]
# there are UNPROCESSED jobs in the queue
if index < len(current_queue) - 1:
unprocessed_queue_indices[current_queue_index] += 1
else:
# current_queue is modified in this array
for job_i in list(current_queue):
if job_i.status == ProcessJobStatus.PROCESSED:
if should_yield:
if job_i.id in self._predict_to_gold:
self._predict_to_gold.pop(job_i.id)
yield job_i.pack # type: ignore
else:
self._proc_mgr.add_to_queue(
queue_index=next_queue_index,
job=job_i)
current_queue.popleft()
else:
raise ProcessFlowException(
f"The job status should be "
f"{ProcessJobStatus.PROCESSED} "
f"at this point.")
# set the UNPROCESSED index
# we do not use "processed_queue_indices" as the
# jobs get PROCESSED whenever they are passed
# into a PackProcessor
unprocessed_queue_indices[current_queue_index] = len(
current_queue)
# update the current queue and processor only
# when all the jobs are processed in the current
# queue
if should_yield:
self._proc_mgr.current_processor_index = 0
self._proc_mgr.current_queue_index = -1
else:
self._proc_mgr.current_processor_index = (
next_queue_index)
self._proc_mgr.current_queue_index = (
next_queue_index)
else:
component.flush()
self.__flush_batch_job_status()
# current queue is modified in the loop
for job in list(current_queue):
if (job.status != ProcessJobStatus.PROCESSED
and not job.is_poison):
raise ValueError("Job is neither PROCESSED nor is "
"a poison. Something went wrong "
"during execution.")
if not job.is_poison and should_yield:
if job.id in self._predict_to_gold:
self._predict_to_gold.pop(job.id)
yield job.pack # type: ignore
elif not should_yield:
self._proc_mgr.add_to_queue(
queue_index=next_queue_index, job=job)
if not job.is_poison:
current_queue.popleft()
if not should_yield:
# set next processor and queue as current
self._proc_mgr.current_processor_index = next_queue_index
self._proc_mgr.current_queue_index = next_queue_index
self._proc_mgr.reset()
def evaluate(self) -> Iterator[Tuple[str, Any]]:
"""
Call the evaluators in the pipeline to collect their results.
Returns:
Iterator of the evaluator results. Each element is a tuple, where
the first one is the name of the evaluator, and the second one
is the output of the evaluator (see
:func:`~forte.evaluation.base.evaluator.get_result`).
"""
for i in self.evaluator_indices:
p = self.components[i]
assert isinstance(p, Evaluator)
yield p.name, p.get_result()