def dump_train_result(train_id, scenario_tag, result):
model = result.pop('model', None)
preprocessor = result.pop('preprocessor', None)
model_name = 'train_result_model_{}'.format(train_id)
result_name = 'train_result_meta_{}'.format(train_id)
dirpath = pathjoin(settings.operation_results_dir, scenario_tag)
to_pickle(pathjoin(dirpath, '{}.pkl'.format(result_name)),
result) # dump result
if model is not None:
model.dump(dirpath, model_name)
if preprocessor is not None:
preprocessor.dump_with_operation_rule(dirpath, train_id)
def load(load_func_kwargs, param):
logger.debug('bigquery loader invoked.')
if param['dataset_name'] is None:
raise ValueError('dataset_name must be set for csv loader')
fname = param['dataset_name'] + '.csv'
df = pd_read_csv(pathjoin(settings.datasource_dir, fname),
**load_func_kwargs)
return df
def get_preprocessor_for_prediction(scenario_tag, train_id, train_result=None, dirpath=None):
if train_result is None:
logger.debug('train_result is None .. load from scenario_tag: {}, train_id: {}'.format(scenario_tag, train_id))
train_result = get_train_result(scenario_tag=scenario_tag, train_id=train_id)
preprocessor = get_preprocessor(train_result['preprocessor_config'])
preprocessor.set_operation_mode('predict')
if dirpath is None:
dirpath = pathjoin(settings.operation_results_dir, scenario_tag)
preprocessor.load_with_operation_rule(dirpath, train_id)
return preprocessor
def dump_predicted_result(predict_id, scenario_tag, dumper_config, df, meta):
dirpath = pathjoin(settings.operation_results_dir, scenario_tag)
fname_meta = 'predict_result_meta_{}'.format(predict_id)
fname = 'predict_result_{}'.format(predict_id)
dump_result_format = dumper_config['name']
if dump_result_format == 'csv':
pd_to_csv(df, pathjoin(dirpath, fname + '.csv'), index=False)
elif dump_result_format == 'pickle':
to_pickle(pathjoin(dirpath, fname + '.pkl'), df)
elif dump_result_format == 'bigquery':
import pandas_gbq
destination_table = dumper_config.get('destination_table')
if destination_table is None:
raise ValueError(
'destination_table must be set for bigquery dumper.')
if dumper_config.get('add_predict_id_enabled', True):
destination_table += ('_' + predict_id)
pandas_gbq.to_gbq(df, destination_table,
**dumper_config.get('kwargs', {}))
else:
raise Exception('invalid format')
to_pickle(pathjoin(dirpath, fname_meta + '.pkl'), meta)
def get_scenario_ids(scenario_tag):
dirpath = pathjoin(settings.operation_results_dir, scenario_tag)
filenames = list_directory(dirpath, mode='filename')
train_regexps = [
re.search(_train_result_meta_pattern, fn) for fn in filenames
]
train_ids = [tr.group(1) for tr in train_regexps if tr is not None]
predict_regexps = [
re.search(_predict_result_meta_pattern, fn) for fn in filenames
]
predict_ids = [tr.group(1) for tr in predict_regexps if tr is not None]
return {
'train': train_ids,
'predict': predict_ids,
}
def load_train_results(scenario_tag='default', train_ids='all'):
dirpath = pathjoin(settings.operation_results_dir, scenario_tag)
logger.debug('dirpath: {}'.format(dirpath))
file_paths = list_directory(dirpath)
result_paths = [
fp for fp in file_paths
if re.search('.+\{}train_result_meta_[0-9]+\.pkl$'.format(os.sep), fp)
is not None
]
results = [from_pickle(rp) for rp in result_paths]
if not (isinstance(train_ids, list) or train_ids == 'all'):
raise ValueError('train_ids must be list type or str "all"')
if not train_ids == 'all':
results = [r for r in results if r['id'] in train_ids]
# convert evaluate result to pandas.DataFrame
for idx, r in enumerate(results):
if 'evaluate' in r:
results[idx]['evaluate']['metrics'] = pd.DataFrame(
[{o['name']: o['value']
for o in met} for met in r['evaluate']['metrics']])
return results
def load_sklearn_model(model, dirpath, model_name):
model._value = from_pickle(pathjoin(dirpath, '{}.pkl'.format(model_name)))
if model._value is None:
raise Exception('load {} failed.'.format(model_name))
return model
def dump_sklearn_model(obj, dirpath, model_name):
return to_pickle(pathjoin(dirpath, '{}.pkl'.format(model_name)), obj.value)
def predict(predict_id,
scenario_tag,
method_type='predict',
dataset_config=None,
train_id='0',
dump_result_enabled=False,
append_evacuated_columns_enabled=False,
dumper_config={},
result_target_columns='all',
result_predict_column='predicted'):
"""
予測を実行する関数
:param predict_id: シナリオ中における予測実行の識別子
:type predict_id: str
:param scenario_tag: シナリオに付与されるタグ
:type scenario_tag: str
:param method_type: 予測のタイプ。設定可能なタイプは `predict` or `predict_proba`
:type method_type: str
:param dataset_config: Datasetの設定。:class:`akebono.dataset.get_dataset` の引数。
:type dataset_config: dict
:param train_id: 予測で使うモデルのtrain_id
:type train_id: str
:param dump_result_enabled: 予測結果の永続化を実行するかのフラグ
:type dump_result_enabled: bool
:param append_evacuated_columns_enabled: Dataset中で退避したカラムをpredictの結果に加えるかを決めるフラグ
:type append_evacuated_columns_enabled: bool
:param dumper_config: 予測結果の設定。
:type dumper_config: dict
:param result_target_columns: 予測結果に含めるべき説明変数のカラム名のリスト。全ての場合は'all'とする
:type result_target_columns: str or list(str)
:param result_predict_column: 予測結果が格納されるカラム名
:type result_predict_column: str
"""
if dataset_config is None:
raise ValueError('dataset_config must be set.')
if dump_result_enabled and 'name' not in dumper_config:
raise ValueError('`name` key must be contained in dumper_config.')
train_id = str(train_id)
tr = get_train_result(scenario_tag=scenario_tag, train_id=train_id)
model, model_config = get_trained_model(scenario_tag,
train_id,
train_result=tr)
ret = {
'type': 'predict',
'method_type': method_type,
'dataset_config': dataset_config,
'train_id': train_id,
'dump_result_enabled': dump_result_enabled,
'dumper_config': dumper_config,
'result_target_columns': result_target_columns,
'result_predict_column': result_predict_column,
'train_result': tr,
'model_config': model_config,
}
dataset_config[
'target_column'] = None # target_columnがNoneだと、predict用のDatasetが返ってくる
dataset = get_dataset(dataset_config)
dirpath = pathjoin(settings.operation_results_dir, scenario_tag)
preprocessor = get_preprocessor_for_prediction(scenario_tag,
train_id,
train_result=tr,
dirpath=dirpath)
X = dataset.get_predictor()
if X.index.size == 0:
raise EmptyDatasetError('empty record')
fX, _ = preprocessor.process(X, None)
gc.collect()
predict_func = getattr(model, method_type, None)
if predict_func is None:
raise Exception('{} is not defined.'.format(method_type))
rawresult = predict_func(fX)
gc.collect()
predict_result = fX.copy()
if not result_target_columns == 'all':
if not isinstance(result_target_columns, list):
raise TypeError('result_target_columns must be list.')
predict_result = predict_result[result_target_columns]
# len(rawresult.shape) > 1でもpredict_resultのdfに格納できるようにするためlistにしている
# そもそもpredict_resultを１つのdfにするべきなのかは考え直しても良いと思う
predict_result.loc[:, result_predict_column] = list(rawresult)
if append_evacuated_columns_enabled:
predict_result = pd.concat([dataset.get_evacuated(), predict_result],
axis=1)
if dump_result_enabled:
logger.debug('dump_predicted_result start.')
dump_predicted_result(predict_id, scenario_tag, dumper_config,
predict_result, ret)
logger.debug('dump_predicted_result done.')
ret['predict_result'] = predict_result
return ret
def load(self, dirpath, name):
self._value.load_model(pathjoin(dirpath, name + '.bin'))
return self
def dump(self, dirpath, name):
self.value.save_model(pathjoin(dirpath, name + '.bin'))
return self
def get_dataset(dataset_config):
"""
Datasetを生成するための関数
:param dataset_config: Datasetについての設定
:type dataset_config: dict
:return: :class:`Dataset` object
Usage:
>>> from akebono.dataset import get_dataset
>>> dataset_config = {
'loader_config': {
'name': '*****@*****.**',
'kwargs': {
'n_features': 1,
'noise': 30.0,
'random_state': 0,
},
},
'target_column': 'target',
'cache_enabled': False,
}
>>> ds = get_dataset(dataset_config)
>>> ds
<akebono.dataset.model.Dataset object at 0x11291acc0>
"""
dataset_name = dataset_config.get('name')
target_column = dataset_config.get('target_column', 'target')
cache_enabled = dataset_config.get('cache_enabled', False)
evacuated_columns = dataset_config.get('evacuated_columns', [])
if not isinstance(evacuated_columns, list):
raise TypeError('evacuated_columns must be list.')
loader_config = dataset_config.get('loader_config')
if not isinstance(loader_config, dict):
raise Exception(
'loader_config must be specified and this type is dict.')
load_func = loader_config.get('name')
load_func = _loader_name_alias.get(load_func, load_func) # エイリアスがあったらそれを使う
if load_func is None:
raise Exception('loader_config.name must be specified.')
load_func = load_object_by_str(load_func)
load_func_kwargs = Param(loader_config.get('kwargs', {}))
loader_param = loader_config.get('param', {})
_reserved_params = (
'dataset_name',
'target_column',
)
for rp in _reserved_params:
if rp in loader_param:
raise KeyError('{} is reserved param.'.format(rp))
loader_param['dataset_name'] = dataset_name
loader_param['target_column'] = target_column
preprocess_func_str = dataset_config.get(
'preprocess_func', '*****@*****.**')
preprocess_func_hash = get_hash(preprocess_func_str)
preprocess_func = load_object_by_str(preprocess_func_str)
preprocess_func_kwargs = Param(
dataset_config.get('preprocess_func_kwargs', {}))
def _core_func():
return preprocess_func(
load_func(copy.deepcopy(load_func_kwargs.value), loader_param),
**copy.copy(preprocess_func_kwargs.value))
fname = '{}_{}_{}_{}'.format(
dataset_name, load_func_kwargs.get_hashed_id(length=24),
preprocess_func_hash[:24],
preprocess_func_kwargs.get_hashed_id(length=24))
dataset_loading_cache_enabled = dataset_config.get(
'dataset_loading_cache_enabled', True)
if dataset_loading_cache_enabled:
ds = datasetholder.get(fname)
if ds is not None:
logger.debug(
'dataset_loading_cache enabled .. {} get done.'.format(
ds.name))
return ds
pkl_fname = fname + '.pkl'
if cache_enabled:
if dataset_name is not None:
logger.info('dataset cache enabled')
_core_func = cache_located_at(
pathjoin(settings.cache_dir, pkl_fname))(_core_func)
else:
raise Exception(
'dataset_config.cache_enabled is True, but dataset_config.name is None'
)
ds = Dataset(fname, _core_func(), target_column, evacuated_columns)
if dataset_loading_cache_enabled:
datasetholder.set(ds)
logger.debug('dataset_loading_cache enabled .. {} set done.'.format(
ds.name))
return ds
