class Machine:
"""
Machine: Combine modules then make a NLP Machine
* Args:
config: machine_config
"""
def __init__(self, config):
self.config = config
self.registry = Registry()
def load(self):
raise NotImplementedError("")
@classmethod
def load_from_config(cls, config_path):
with open(config_path, "r", encoding="utf-8") as in_file:
machine_config = NestedNamespace()
machine_config.load_from_json(json.load(in_file))
machine_name = machine_config.name
config = getattr(machine_config, machine_name, {})
return cls(config)
def __call__(self, text):
raise NotImplementedError("")
def make_module(self, config):
"""
Make component or experiment for claf Machine's module
* Args:
- config: module's config (claf.config.namespace.NestedNamespace)
"""
module_type = config.type
if module_type == Module.COMPONENT:
name = config.name
module_config = getattr(config, name, {})
if isinstance(module_config, Namespace):
module_config = vars(module_config)
if getattr(config, "params", None):
module_config.update(config.params)
return self.registry.get(f"component:{name}")(**module_config)
elif module_type == Module.EXPERIMENT:
experiment_config = Namespace()
experiment_config.checkpoint_path = config.checkpoint_path
experiment_config.cuda_devices = getattr(config, "cuda_devices",
None)
experiment_config.interactive = False
experiment = Experiment(Mode.PREDICT, experiment_config)
experiment.set_predict_mode(preload=True)
return experiment
else:
raise ValueError(
f"module_type is available only [component|experiment]. not '{module_type}'"
)
def test_open_qa_with_bidaf_model(open_qa_config):
claf_name = open_qa_config.name
config = getattr(open_qa_config, claf_name, {})
registry = Registry()
claf_machine = registry.get(f"machine:{claf_name}")(config)
question = utils.make_random_tokens(5)
answer = claf_machine(question)
answer = json.dumps(answer, indent=4, ensure_ascii=False)
class ModelFactory(Factory):
"""
Model Factory Class
Create Concrete model according to config.model_name
Get model from model registries (eg. @register("model:{model_name}"))
* Args:
config: model config from argument (config.model)
"""
def __init__(self, config):
self.registry = Registry()
self.name = config.name
self.model_config = {}
if getattr(config, config.name, None):
self.model_config = vars(getattr(config, config.name))
self.is_independent = getattr(config, "independent", False)
@overrides
def create(self, token_makers, **params):
model = self.registry.get(f"model:{self.name}")
if issubclass(model, ModelWithTokenEmbedder):
token_embedder = self.create_token_embedder(model, token_makers)
self.model_config["token_embedder"] = token_embedder
elif issubclass(model, ModelWithoutTokenEmbedder):
self.model_config["token_makers"] = token_makers
else:
raise ValueError(
"Model must have inheritance. (ModelWithTokenEmbedder or ModelWithoutTokenEmbedder)"
)
return model(**self.model_config, **params)
def create_token_embedder(self, model, token_makers):
# 1. Specific case
# ...
# 2. Base case
if issubclass(model, ReadingComprehension):
return token_embedder.RCTokenEmbedder(token_makers)
else:
return token_embedder.BasicTokenEmbedder(token_makers)
class TokenMakersFactory(Factory):
"""
TokenMakers Factory Class
* Args:
config: token config from argument (config.token)
"""
LANGS = ["eng", "kor"]
def __init__(self, config):
self.config = config
self.registry = Registry()
@overrides
def create(self):
if getattr(self.config, "tokenizer", None):
tokenizers = make_all_tokenizers(
convert_config2dict(self.config.tokenizer))
else:
tokenizers = {}
token_names, token_types = self.config.names, self.config.types
if len(token_names) != len(token_types):
raise ValueError(
"token_names and token_types must be same length.")
token_makers = {"tokenizers": tokenizers}
for token_name, token_type in sorted(zip(token_names, token_types)):
token_config = getattr(self.config, token_name, {})
if token_config != {}:
token_config = convert_config2dict(token_config)
# Token (tokenizer, indexer, embedding, vocab)
token_config = {
"tokenizers": tokenizers,
"indexer_config": token_config.get("indexer", {}),
"embedding_config": token_config.get("embedding", {}),
"vocab_config": token_config.get("vocab", {}),
}
token_makers[token_name] = self.registry.get(
f"token:{token_type}")(**token_config)
return token_makers
class DataReaderFactory(Factory):
"""
DataReader Factory Class
Create Concrete reader according to config.dataset
Get reader from reader registries (eg. @register("reader:{reader_name}"))
* Args:
config: data_reader config from argument (config.data_reader)
"""
def __init__(self, config):
self.registry = Registry()
self.dataset_name = config.dataset
file_paths = {}
if getattr(config, "train_file_path",
None) and config.train_file_path != "":
file_paths["train"] = config.train_file_path
if getattr(config, "valid_file_path",
None) and config.valid_file_path != "":
file_paths["valid"] = config.valid_file_path
self.reader_config = {"file_paths": file_paths}
if "params" in config and type(config.params) == dict:
self.reader_config.update(config.params)
if "tokenizers" in config:
self.reader_config["tokenizers"] = config.tokenizers
dataset_config = getattr(config, config.dataset, None)
if dataset_config is not None:
dataset_config = vars(dataset_config)
self.reader_config.update(dataset_config)
@overrides
def create(self):
reader = self.registry.get(f"reader:{self.dataset_name.lower()}")
return reader(**self.reader_config)
# -*- coding: utf-8 -*-
import json
from claf.config import args
from claf.config.registry import Registry
from claf.learn.mode import Mode
from claf import utils as common_utils
if __name__ == "__main__":
registry = Registry()
machine_config = args.config(mode=Mode.MACHINE)
machine_name = machine_config.name
config = getattr(machine_config, machine_name, {})
claf_machine = registry.get(f"machine:{machine_name}")(config)
while True:
question = common_utils.get_user_input(
f"{getattr(machine_config, 'user_input', 'Question')}")
answer = claf_machine.get_answer(question)
answer = json.dumps(answer, indent=4, ensure_ascii=False)
print(
f"{getattr(machine_config, 'system_response', 'Answer')}: {answer}"
)