# yapf: disable
import asyncio
from copy import deepcopy
from data_juicer.config import get_init_configs, prepare_side_configs
from data_juicer.core.data.dj_dataset import nested_query
from data_juicer.utils.constant import JobRequiredKeys
from jsonargparse import dict_to_namespace
from loguru import logger
from data_juicer_sandbox.context_infos import ContextInfos, JobInfos
from data_juicer_sandbox.factories import (
data_analyzer_factory,
data_evaluator_factory,
data_executor_factory,
data_pool_manipulator_factory,
general_data_executor_factory,
general_probe_factory,
model_evaluator_factory,
model_infer_evaluator_factory,
model_infer_executor_factory,
model_train_executor_factory,
)
from data_juicer_sandbox.utils import add_iter_subdir_to_paths
[文档]
class BaseHook:
[文档]
def __init__(self, job_cfg, watcher, *args, **kwargs):
self.job_cfg = job_cfg
self.watcher = watcher
self.hook_type = job_cfg.get(JobRequiredKeys.hook.value)
# hook inputs mapping for config update
self.input_mapping = job_cfg.get(JobRequiredKeys.input.value, {})
# names of hook output for record the results of hook running
self.output_keys = job_cfg.get(JobRequiredKeys.output.value, None)
if self.output_keys and not isinstance(self.output_keys, list):
self.output_keys = [self.output_keys]
# hook local settings for configs
self.local_settings = job_cfg.get(JobRequiredKeys.local.value, {})
# unique meta name for this job
self.meta_name = job_cfg.get(JobRequiredKeys.meta_name.value, 'anonymous_meta_name')
# data-juicer config for some jobs based on Data-Juicer
self.dj_cfg = job_cfg.get(JobRequiredKeys.dj_configs.value, None)
self.inited_dj_cfg = None
# extra config for some other specific jobs
self.extra_cfg = job_cfg.get(JobRequiredKeys.extra_configs.value, None)
# the current iteration number
self.curr_iter = None
[文档]
def run(self, context_infos: ContextInfos):
# record the current iteration number
self.curr_iter = context_infos.iter
# update the args of this hook with input mappings and local settings
self._input_updating_hook(context_infos)
# updating export_path of dj config.
# we assume that for non-dj hooks, the output paths are specified by "export_path" field always.
if self.inited_dj_cfg and 'export_path' in self.inited_dj_cfg:
self.inited_dj_cfg['export_path'] = add_iter_subdir_to_paths(
[self.inited_dj_cfg['export_path']], self.curr_iter)[0]
if self.extra_cfg and 'export_path' in self.extra_cfg:
if isinstance(self.extra_cfg['export_path'], list):
self.extra_cfg['export_path'] = add_iter_subdir_to_paths(
self.extra_cfg['export_path'], self.curr_iter)
else:
self.extra_cfg['export_path'] = add_iter_subdir_to_paths(
[self.extra_cfg['export_path']], self.curr_iter)[0]
outputs = self.hook()
if outputs:
return self._output_recording_hook(outputs)
else:
return None
def _input_updating_hook(self, context_infos: ContextInfos):
self.specify_dj_and_extra_configs(allow_fail=True)
prev_dj_cfg = deepcopy(self.dj_cfg) if self.dj_cfg else None
prev_extra_cfg = deepcopy(self.extra_cfg) if self.extra_cfg else None
logger.debug(f"Previous DJ Cfg: {prev_dj_cfg}")
logger.debug(f"Previous Extra Cfg: {prev_extra_cfg}")
logger.debug(f"Local Settings: {self.local_settings}")
logger.debug(f"Input Mappings: {self.input_mapping}")
# update configs according to local settings
for key, value in self.local_settings.items():
key_to_updated_parts = key.split('.')
cfg_type = key_to_updated_parts[0]
# check which config group to update
if cfg_type == JobRequiredKeys.dj_configs.value:
target_config = self.dj_cfg
elif cfg_type == JobRequiredKeys.extra_configs.value:
target_config = self.extra_cfg
else:
raise ValueError(f'The key {cfg_type} to update is not supported.')
key_to_updated_parts = key_to_updated_parts[1:]
# update the target key
if len(key_to_updated_parts) > 0:
if len(key_to_updated_parts) > 1:
target_config = nested_query(target_config, '.'.join(key_to_updated_parts[:-1]))
target_config[key_to_updated_parts[-1]] = value
else:
if cfg_type == JobRequiredKeys.dj_configs.value:
self.dj_cfg = value
elif cfg_type == JobRequiredKeys.extra_configs.value:
self.extra_cfg = value
# update configs according to input mapping and context_infos
for key_to_updated, key_in_history in self.input_mapping.items():
key_to_updated_parts = key_to_updated.split('.')
key_in_history_parts = key_in_history.split('.')
# check if it's -1
history_job_infos = context_infos
if key_in_history_parts[0] == '-1':
# get the latest infos
if len(key_in_history_parts) <= 1:
raise ValueError(f'Need to specify the job result keys precisely for inputs to '
f'find the target values. Only got [{key_in_history}].')
# find the last non-empty job_infos
if len(context_infos) == 0:
raise ValueError(f'Cannot find the previous non-empty job infos for [{key_in_history}].')
history_job_infos = context_infos.get_the_last_job_infos()
key_in_history_parts = key_in_history_parts[1:]
# query target values
target_value = history_job_infos['.'.join(key_in_history_parts)]
# check which config group to update
cfg_type = key_to_updated_parts[0]
if cfg_type == JobRequiredKeys.dj_configs.value:
target_config = self.dj_cfg
elif cfg_type == JobRequiredKeys.extra_configs.value:
target_config = self.extra_cfg
else:
raise ValueError(f'The key {key_to_updated_parts[0]} to update is not supported.')
key_to_updated_parts = key_to_updated_parts[1:]
# update the target key
if len(key_to_updated_parts) > 0:
if len(key_to_updated_parts) > 1:
target_config = nested_query(target_config, '.'.join(key_to_updated_parts[:-1]))
target_config[key_to_updated_parts[-1]] = target_value
else:
if cfg_type == JobRequiredKeys.dj_configs.value:
self.dj_cfg = target_value
elif cfg_type == JobRequiredKeys.extra_configs.value:
self.extra_cfg = target_value
if self.dj_cfg != prev_dj_cfg or self.extra_cfg != prev_extra_cfg:
logger.info('Configs are updated according to input and local settings. Re-initializing them...')
self.specify_dj_and_extra_configs()
def _output_recording_hook(self, outputs):
if isinstance(outputs, tuple):
# multiple outputs
outputs = list(outputs)
else:
# single output
if isinstance(outputs, list) and len(outputs) == 1:
# if there is only 1 ele in the returned list, unpack it
outputs = outputs[0]
outputs = [outputs]
if self.output_keys is None:
self.output_keys = list(range(len(outputs)))
if len(outputs) != len(self.output_keys):
raise ValueError(
f'## HOOK [{self.hook_type}]: The number of outputs does not match the number of output keys. '
f'Expected {len(self.output_keys)} but got {len(outputs)}'
)
curr_job_infos = JobInfos(self.meta_name, self.output_keys, outputs)
return curr_job_infos
[文档]
def hook(self, **kwargs):
raise NotImplementedError
[文档]
def specify_dj_and_extra_configs(self, allow_fail=False):
# prepare data-juicer config
if self.dj_cfg:
prev_dj_cfg = self.dj_cfg
prev_inited_dj_cfg = self.inited_dj_cfg
try:
logger.info('Parsing Data-Juicer configs in the job.')
self.dj_cfg = prepare_side_configs(self.dj_cfg)
# require Data-Juicer data process in some jobs
# so we need to init the Data-Juicer data process configs
self.inited_dj_cfg = get_init_configs(self.dj_cfg)
self.dj_cfg = dict_to_namespace(self.dj_cfg)
except Exception as e:
if allow_fail:
self.dj_cfg = prev_dj_cfg
self.inited_dj_cfg = prev_inited_dj_cfg
else:
raise e
# prepare extra configs
if self.extra_cfg:
prev_extra_cfg = self.extra_cfg
try:
logger.info('Parsing extra configs in the job.')
self.extra_cfg = prepare_side_configs(self.extra_cfg)
self.extra_cfg = dict_to_namespace(self.extra_cfg)
except Exception as e:
if allow_fail:
self.extra_cfg = prev_extra_cfg
else:
raise e
[文档]
class ProbeViaAnalyzerHook(BaseHook):
"""
The hook to probe dataset via Data-Juicer Analyzer.
Input:
- A data-juicer config.
Output:
- the path to export the analyzed dataset.
"""
[文档]
def __init__(self, job_cfg, watcher, *args, **kwargs):
"""
Initialize the hook for probing the data via Analyzer
:param job_cfg: the job configs
:param watcher: for watching the result
"""
super(ProbeViaAnalyzerHook, self).__init__(job_cfg, watcher, *args,
**kwargs)
[文档]
def hook(self, **kwargs):
analyzer = data_analyzer_factory(self.inited_dj_cfg)
# probe the data via Analyzer
logger.info('Begin to analyze data')
analyzer.run()
analyzer_res = analyzer.overall_result
# drop string rows to avoid unaligned dtypes
string_rows = ['unique', 'top', 'freq']
for row_name in string_rows:
if row_name in analyzer_res.index:
analyzer_res = analyzer_res.drop(row_name)
self.watcher.watch(analyzer_res, self.meta_name)
return analyzer.exporter.export_path
[文档]
class ProbeViaModelInferHook(BaseHook):
[文档]
def __init__(self, job_cfg, watcher, *args, **kwargs):
"""
Initialize the hook for probing the data via Model Infer
:param job_cfg: the job configs
:param watcher: for watching the result
"""
super(ProbeViaModelInferHook, self).__init__(job_cfg, watcher, *args,
**kwargs)
[文档]
def hook(self, **kwargs):
data_executor = data_executor_factory(self.inited_dj_cfg)
model_infer_executor = model_infer_evaluator_factory(self.extra_cfg)
# TODO
# probe the model (calling inference sub-pipeline) based on
# original data, such that we know what is the "hard" data for
# the model and how to process the data accordingly
sampled_data = data_executor.sample_data(
sample_ratio=self.inited_dj_cfg.data_probe_ratio,
sample_algo=self.inited_dj_cfg.data_probe_algo,
)
res_type, infer_res = model_infer_executor.run(
model_infer_executor.model_config['type'], sampled_data)
self.watcher.watch(infer_res, self.meta_name)
return infer_res
[文档]
class GeneralProbeHook(BaseHook):
[文档]
def __init__(self, job_cfg, watcher, *args, **kwargs):
super(GeneralProbeHook, self).__init__(job_cfg, watcher, *args, **kwargs)
[文档]
def hook(self, **kwargs):
data_probe = general_probe_factory(self.extra_cfg)
logger.info('Begin to probe data.')
ret = data_probe.run()
return ret
[文档]
class RefineRecipeViaKSigmaHook(BaseHook):
[文档]
def __init__(self, job_cfg, watcher, *args, **kwargs):
"""
Initialize the hook for refining the recipe via K Sigma
:param job_cfg: the job configs
:param watcher: for watching the result
"""
super(RefineRecipeViaKSigmaHook,
self).__init__(job_cfg, watcher, *args, **kwargs)
[文档]
def hook(self, **kwargs):
path_k_sigma_recipe = self.extra_cfg.path_k_sigma_recipe
# use k-sigma strategy to modify the data recipe
# TODO: cannot import tools correctly if DJ is installed by pypi. Maybe we need
# other importing methods.
from data_juicer.tools.hpo.execute_hpo_3sigma import modify_recipe_k_sigma
modify_recipe_k_sigma(self.dj_cfg, self.watcher.query(self.meta_name),
path_k_sigma_recipe)
return path_k_sigma_recipe
[文档]
class RefineRecipeViaModelFeedbackHook(BaseHook):
[文档]
def __init__(self, job_cfg, watcher, *args, **kwargs):
"""
Initialize the hook for refining the recipe via Model Feedback
:param job_cfg: the job configs
:param watcher: for watching the result
"""
super(RefineRecipeViaModelFeedbackHook,
self).__init__(job_cfg, watcher, *args, **kwargs)
[文档]
def hook(self, **kwargs):
# TODO
# use model-feedback-based strategy to modify the data recipe,
# e.g., more mapper on the "hard" or "sensitive" data, those were
# ranked by user-interested measurement after model inference
if self.sandbox_cfg.path_model_feedback_recipe is not None:
# modify_recipe_model_feedback(
# self.sandbox_cfg,
# self.watcher.query("measure_on_infer_res"),
# self.sandbox_cfg.path_model_feedback_recipe)
raise NotImplementedError('Not implemented yet.')
return None
[文档]
class ProcessDataHook(BaseHook):
[文档]
def __init__(self, job_cfg, watcher, *args, **kwargs):
"""
Initialize the hook for processing the data via Data-Juicer
:param job_cfg: the job configs
:param watcher: for watching the result
"""
super(ProcessDataHook, self).__init__(job_cfg, watcher, *args,
**kwargs)
[文档]
def hook(self, **kwargs):
data_executor = data_executor_factory(self.inited_dj_cfg)
# basic routine to process data, users can customize this freely
logger.info('Begin to process the data with given dj recipe')
data_executor.run()
return self.inited_dj_cfg.export_path
[文档]
class DataPoolManipulationHook(BaseHook):
"""
Hook for data pool manipulation, including construction, combination, ranking, etc.
"""
[文档]
def __init__(self, job_cfg, watcher, *args, **kwargs):
super(DataPoolManipulationHook, self).__init__(job_cfg, watcher, *args, **kwargs)
[文档]
def hook(self, **kwargs):
data_pool_manipulator = data_pool_manipulator_factory(self.extra_cfg)
logger.info('Begin to manipulate data pools.')
ret = data_pool_manipulator.run()
return ret
[文档]
class GeneralDataExecutorHook(BaseHook):
[文档]
def __init__(self, job_cfg, watcher, *args, **kwargs):
super(GeneralDataExecutorHook, self).__init__(job_cfg, watcher, *args, **kwargs)
[文档]
def hook(self, **kwargs):
data_executor = general_data_executor_factory(self.extra_cfg)
logger.info('Begin to execute general data executor.')
ret = data_executor.run()
return ret
[文档]
class TrainModelHook(BaseHook):
[文档]
def __init__(self, job_cfg, watcher, *args, **kwargs):
"""
Initialize the hook for model training
:param job_cfg: the job configs
:param watcher: for watching the result
"""
super(TrainModelHook, self).__init__(job_cfg, watcher, *args, **kwargs)
[文档]
def hook(self, **kwargs):
# try to update train dataset
model_trainer = model_train_executor_factory(self.extra_cfg,
watcher=self.watcher)
# basic routine to train model via the processed data,
# users can customize this freely
logger.info('Begin to train the model with given model config')
# update training dataset path
ret = asyncio.run(
model_trainer.run(model_trainer.model_config['type'], **kwargs))
return ret
[文档]
class InferModelHook(BaseHook):
[文档]
def __init__(self, job_cfg, watcher, *args, **kwargs):
"""
Initialize the hook for model training
:param job_cfg: the job configs
:param watcher: for watching the result
"""
super(InferModelHook, self).__init__(job_cfg, watcher, *args, **kwargs)
[文档]
def hook(self, **kwargs):
model_infer = model_infer_executor_factory(self.extra_cfg,
watcher=self.watcher)
logger.info('Begin to infer the model with given model config')
ret = asyncio.run(model_infer.run(model_infer.model_config['type'],
**kwargs))
return ret
[文档]
class EvaluateDataHook(BaseHook):
[文档]
def __init__(self, job_cfg, watcher, *args, **kwargs):
"""
Initialize the hook for data evaluation
:param job_cfg: the job configs
:param watcher: for watching the result
"""
super(EvaluateDataHook, self).__init__(job_cfg, watcher, *args,
**kwargs)
[文档]
def hook(self, **kwargs):
data_evaluator = data_evaluator_factory(self.extra_cfg)
# basic routine to evaluate the given data,
# users can customize this freely
logger.info('Begin to evaluate the data with given evaluator config')
eval_res = data_evaluator.run(eval_type='data', **kwargs)
self.watcher.watch(eval_res, self.meta_name)
return eval_res
[文档]
class EvaluateModelHook(BaseHook):
[文档]
def __init__(self, job_cfg, watcher, *args, **kwargs):
"""
Initialize the hook for model evaluation
:param job_cfg: the job configs
:param watcher: for watching the result
"""
super(EvaluateModelHook, self).__init__(job_cfg, watcher, *args,
**kwargs)
[文档]
def hook(self, **kwargs):
model_evaluator = model_evaluator_factory(self.extra_cfg)
# basic routine to evaluate the given model,
# users can customize this freely
logger.info('Begin to evaluate the model with given evaluator config')
eval_res = model_evaluator.run(kwargs)
self.watcher.watch(eval_res, self.meta_name)
return eval_res
HOOK_MAPPING = {
# Data/Model Probe hooks
'ProbeViaAnalyzerHook': ProbeViaAnalyzerHook,
'ProbeViaModelInferHook': ProbeViaModelInferHook,
'GeneralProbeHook': GeneralProbeHook,
# Data-Recipe Refinement hooks
'RefineRecipeViaKSigmaHook': RefineRecipeViaKSigmaHook,
'RefineRecipeViaModelFeedbackHook': RefineRecipeViaModelFeedbackHook,
# Data/Model Execution hooks
'ProcessDataHook': ProcessDataHook,
'DataPoolManipulationHook': DataPoolManipulationHook,
'GeneralDataExecutorHook': GeneralDataExecutorHook,
'TrainModelHook': TrainModelHook,
'InferModelHook': InferModelHook,
# Evaluation hooks
'EvaluateDataHook': EvaluateDataHook,
'EvaluateModelHook': EvaluateModelHook,
}
[文档]
def register_hook(job_cfg, watcher):
if job_cfg.hook not in HOOK_MAPPING:
raise ValueError('Undefined hook: [{job_cfg.hook}].')
return HOOK_MAPPING[job_cfg.hook](job_cfg, watcher)
