def pd_dsa2_custom(df: pd.DataFrame, col: list = None, pars: dict = None):
"""
Example of custom Processor Combining
Usage :
,{"uri": THIS_FILEPATH + "::pd_dsa2_custom", "pars": {'coldate': 'date'}, "cols_family": "coldate", "cols_out": "coldate_features1", "type": "" },
"""
prefix = "tseries_feat" ### Used acolumn index
#### Inference time LOAD previous pars ###########################################
from prepro import prepro_load, prepro_save
prepro, pars_saved, cols_saved = prepro_load(prefix, pars)
###################################################################################
#### Do something #################################################################
from source.prepro_tseries import pd_ts_date, pd_ts_rolling
if prepro is None: ### Training time
dfy, coly = pars['dfy'], pars['coly']
coldate = pars['coldate']
#df = df.set_index(coldate)
#### time features
dfi, coli = pd_ts_date(
df,
cols=[coldate],
pars={'col_add': ['day', 'month', 'year', 'weekday']})
df_new = dfi
#### Rolling features
dfi, coli = pd_ts_rolling(df,
cols=['date', 'item', 'store', 'sales'],
pars={
'col_groupby': ['store', 'item'],
'col_stat': 'sales',
'lag_list': [7, 30]
})
df_new = pd.concat([df_new, dfi], axis=1)
else: ### predict time
"""
y is NOT provided, need to calculate y based on past values.
Auto-regressive feature engineering.
"""
pars = pars_saved ##merge
###################################################################################
###################################################################################
### Clean up the df ###############################################################
df_new.index = df.index ### Impt for JOIN
df_new.columns = [col + f"_{prefix}" for col in df_new.columns]
cols_new = list(df_new.columns)
###### Training time save all #####################################################
df_new, col_pars = prepro_save(prefix, pars, df_new, cols_new, prepro)
return df_new, col_pars
def pd_prepro_custom2(df: pd.DataFrame, cols: list = None, pars: dict = None):
""" Generic template for feature generation
'colnum' : ['sales1' 'units' ]
'pars_function_list' : [
{ 'name': 'deltapy.transform::robust_scaler', 'pars': {'drop':["Close_1"]} },
{ 'name': 'deltapy.transform::standard_scaler', 'pars': {'drop':["Close_1"]} },
]e
:param df:
:param col:
:param pars:
:return:
"""
prefix = "coltseries_custom2"
#### Inference time LOAD previous pars ###########################################
from prepro import prepro_load, prepro_save
prepro, pars_saved, cols_saved = prepro_load(prefix, pars)
#### Do something #################################################################
df = df[col]
coldate = pars['coldate']
colnum = pars['colnum']
colcat = pars['colcat']
colgroups = pars['colgroup']
colgstat = pars['colstat']
log("### Only dates")
df1 = pd_ts_date(df, coldate, pars)
coldate1 = list(df1.columns)
log("### Initial features")
df1 = df1.join(df, on=coldate, how='left')
log("### Groupby features")
df2 = pd_ts_groupby(df, col, pars)
df1 = df1.join(df2, on=coldate, how='left')
log("### Numerical features")
colnum2 = list(df2.columns) + colnum
df1 = df1.set_index(coldate1)
log("### Deltapy features")
for pars_function_dict_i in pars.get('pars_function_list', []):
dfi = pd_ts_deltapy_generic(df1,
col=colnum2,
pars=pars_function_dict_i)
df1 = df1.join(dfi, on=coldate, how='left')
df_new = df1
### Transform features ###################################
df_new.columns = [col + f"_{prefix}" for col in df_new.columns]
cols_new = list(df_new.columns)
###### Training time save all #####################################################
df_new, col_pars = prepro_save(prefix, pars, df_new, cols_new, prepro)
return df_new, col_pars
def test_prepro_v1():
df = test_get_sampledata()
time_eng = pd_ts_date(df, ['Date'], pars={})
onehot = pd_ts_onehot(df, ['Name'], {})
trendless = pd_ts_difference(df, ['Close'], {})