def prepare_data_features_kr(df_feat_kr):
'''
Prepare DafaFrame to be used with model
Reduce timeframe to usable dates without NaN values for features
'''
df_out = df_feat_kr.copy()
# correct age
df_age_pos = df_out["age_pos"].copy()
df_age_pos.fillna(method="pad", inplace=True)
df_out["age_pos"] = df_age_pos
# check by drop if nan
df_out.dropna(inplace=True,
subset=[
"date", 'nb_cases', 'T_min', 'T_max', 'H_mean',
'W_speed', 'pos', 'test', 'day_num', 'age_pos'
])
date_old = add_days(df_out.index[0].strftime("%Y-%m-%d"), -1)
dates_index = df_out.index.strftime("%Y-%m-%d")
for date_curr in dates_index:
if date_curr != add_days(date_old, 1):
print("ERROR : ", date_curr)
#break
assert date_curr == add_days(date_old, 1)
date_old = date_curr
return df_out
def extrap_missing_kr(df_feat_kr_tmp, list_missing):
'''
Extrapolate data if missing dates in data
'''
def fun_apply_extrap(date_curr):
return date_curr in list_missing
if len(list_missing) == 0:
return df_feat_kr_tmp
'''for dates_curr in list_missing:
list_dates_tmp = df_feat_kr_tmp["date"].tolist()
# search day-1
date_before = add_days(dates_curr, -1)
if date_before in list_dates_tmp:'''
str_date_min = min(df_feat_kr_tmp["date"])
str_date_min = add_days(str_date_min, -1)
list_range = generate_list_dates(str_date_min,
max(df_feat_kr_tmp["date"]))
new_index = pd.DatetimeIndex(list_range)
df_feat_kr_tmp = df_feat_kr_tmp.reindex(new_index, method='ffill')
df_feat_kr_tmp["date"] = \
pd.to_datetime(df_feat_kr_tmp.index).strftime("%Y-%m-%d").tolist()
df_feat_kr_tmp["extrap"] = False
df_feat_kr_tmp["extrap"] = df_feat_kr_tmp["date"].apply(fun_apply_extrap)
return df_feat_kr_tmp
def check_update_df_feat_kr(date_now=None, force_update=False):
if date_now is None:
date_now = datetime.datetime.now().strftime("%Y-%m-%d")
if force_update:
return True, DATE_FIRST_CASES_GOUV_KR, \
DATE_FIRST_FEAT_OK_KR, date_now
flag_update = True # update to be done ?
flag_update_age = True
if os.path.isfile(PATH_DF_FEAT_KR):
df_feat_kr = pd.read_csv(PATH_DF_FEAT_KR)
# first date to download for cases
date_first_extrap = get_first_day_extrap_kr(df_feat_kr)
if date_first_extrap is not None:
# take first date of extrap to try to update it
date_req_start = date_first_extrap
else:
# normal mode : take last date +1
date_req_start = add_days(df_feat_kr["date"].max(), 1)
# first date to download for ages
date_req_start_age = \
df_feat_kr[df_feat_kr["daily_age"].isna() & \
(df_feat_kr["date"] > DATE_FIRST_FEAT_OK_KR)]["date"].min()
if (date_req_start_age is np.nan):
date_req_start_age = date_req_start
#df_feat_kr[df_feat_kr["age_pos"].isna()]["date"].max()
else:
date_req_start = DATE_FIRST_CASES_GOUV_KR
date_req_start_age = DATE_FIRST_FEAT_OK_KR
if date_req_start > date_now:
#print(" date_req_start : ", date_req_start)
#print("update false")
flag_update = False
if date_req_start_age > date_now:
flag_update_age = False
if not flag_update:
date_req_start = None
if not flag_update_age:
date_req_start_age = None
if (flag_update | flag_update_age):
date_req_end = date_now
else:
date_req_end = None
print("Updating Data KR...")
print("update cases : ", flag_update)
print("update age : ", flag_update_age)
print("date_req_start: ", date_req_start)
print("date_req_start_age: ", date_req_start_age)
print("date_req_end: ", date_req_end)
return flag_update, flag_update_age, \
date_req_start, date_req_start_age, date_req_end
def update_pred_pos(df_feat_fr, from_disk=False):
'''
Update prediction data positive cases France
'''
# check if last prediction is after last known date
if os.path.isfile(PATH_DF_PLOT_PRED):
df_plot_pred = pd.read_csv(PATH_DF_PLOT_PRED)
df_plot_pred.index = df_plot_pred["date"]
if df_plot_pred["date"].min() <= df_feat_fr["date"].max():
from_disk = False
# if no prediction or if from disk
if (not settings.PREDICT) | from_disk:
df_plot_pred = pd.read_csv(PATH_DF_PLOT_PRED)
df_plot_pred.index = df_plot_pred["date"]
return df_plot_pred
# prepare features
dataset, data_std, data_mean = prepare_dataset(df_feat_fr)
# predict next days
if settings.MODEL_TFLITE:
json_list_list_x = prepare_to_lambda_future(dataset)
resp = requests.post(URL_PREDICT, json=json_list_list_x)
print("status code : ", resp.status_code)
if resp.status_code == 200:
y_multi_pred = retrieve_from_lambda(resp)
else:
print("AWS Lamdba future pred ERROR!")
df_plot_pred = pd.read_csv(PATH_DF_PLOT_PRED)
df_plot_pred.index = df_plot_pred["date"]
return df_plot_pred
else:
# prepare data : very last days
x_multi = np.array([dataset[-PAST_HISTORY:, :]])
# load model
multi_step_model = tf.keras.models.load_model(PATH_MDL_MULTI_STEP)
y_multi_pred = multi_step_model.predict(x_multi)
# convert in positive cases
y_pos_pred = y_multi_pred * data_std[4] + data_mean[4]
# pos pred next 3 days from last day : date, pos, total (sum)
str_date_pred_0 = df_feat_fr.date.max()
str_date_pred_1 = add_days(str_date_pred_0, FUTURE_TARGET)
list_dates_pred = generate_list_dates(str_date_pred_0, str_date_pred_1)
# figure
df_plot_pred = pd.DataFrame(index=list_dates_pred,
columns=["date"],
data=list_dates_pred)
df_plot_pred["pos"] = y_pos_pred[0].astype(int)
arr_nb_pred = df_plot_pred["pos"].cumsum().values
df_plot_pred["nb_cases"] = df_feat_fr["nb_cases"].max() + arr_nb_pred
# save for future pred
df_plot_pred.to_csv(PATH_DF_PLOT_PRED, index=False)
return df_plot_pred
def update_pos(df_feat_fr):
'''
Update plot data positive cases France
'''
# pos last NB_DAY_PLOT days : date, pos, total (sum)
str_date_0 = add_days(df_feat_fr.date.max(), -NB_DAY_PLOT)
df_plot = df_feat_fr[df_feat_fr["date"] >= str_date_0].copy()
return df_plot
def update_pred_pos_all_kr(df_feat_kr, from_disk=False):
'''
Update prediction data positive cases France for all days
'''
if os.path.isfile(PATH_DF_PLOT_PRED_ALL_KR):
df_plot_pred = pd.read_csv(PATH_DF_PLOT_PRED_ALL_KR)
df_plot_pred.index = df_plot_pred["date"]
if df_plot_pred["date"].max() < df_feat_kr["date"].max():
from_disk = False
else:
from_disk = False
if (not settings.PREDICT) | from_disk:
df_plot_pred_all = pd.read_csv(PATH_DF_PLOT_PRED_ALL_KR)
df_plot_pred_all.index = df_plot_pred_all["date"]
return df_plot_pred_all
# prepare features
dataset, data_std, data_mean = prepare_dataset_kr(df_feat_kr)
# predict
if settings.MODEL_TFLITE:
json_list_list_x = prepare_to_lambda(dataset)
resp = requests.post(URL_PREDICT_KR, json=json_list_list_x)
print("status code : ", resp.status_code)
if resp.status_code == 200:
y_multi_pred = retrieve_from_lambda(resp)
else:
print("AWS Lamdba future pred ERROR!")
df_plot_pred_all = pd.read_csv(PATH_DF_PLOT_PRED_ALL_KR)
df_plot_pred_all.index = df_plot_pred_all["date"]
return df_plot_pred_all
else:
# load model
multi_step_model = tf.keras.models.load_model(PATH_MDL_MULTI_STEP_KR)
list_x = []
# prepare data : very last days
nb_max = NB_PERIOD_PLOT
for I in range(nb_max, 0, -1):
I_start = I * FUTURE_TARGET - PAST_HISTORY
if I_start < 0:
break
I_end = I * FUTURE_TARGET
list_x.append(np.array([dataset[I_start:I_end, :]]))
# model prediction
for I, x_multi in enumerate(list_x):
if I:
y_multi_pred = np.concatenate(
[y_multi_pred,
multi_step_model.predict(x_multi)], axis=1)
else:
y_multi_pred = multi_step_model.predict(x_multi)
# convert in positive cases
y_pos_pred = y_multi_pred * data_std[4] + data_mean[4]
# list of dates
K_days = y_pos_pred.shape[1]
print("K_days = ", K_days)
print("y_pos_pred.shape = ", y_pos_pred.shape)
str_date_pred_1 = df_feat_kr.date.max()
str_date_pred_0 = add_days(str_date_pred_1, -1 * K_days)
list_dates_pred = generate_list_dates(str_date_pred_0, str_date_pred_1)
# create df output
df_plot_pred_all = pd.DataFrame(index=list_dates_pred,
columns=["date"],
data=list_dates_pred)
# daily
df_plot_pred_all["pos"] = y_pos_pred[0].astype(int)
# Total : cumulate sum
list_nb_cases = []
str_date_nb_0 = str_date_pred_0
for I in range(0, df_plot_pred_all["pos"].shape[0], FUTURE_TARGET):
str_date_nb_0 = add_days(str_date_pred_0, I)
nb_0 = df_feat_kr[df_feat_kr["date"] == str_date_nb_0]["nb_cases"][0]
arr_nb = nb_0 + \
df_plot_pred_all.iloc[I:I+FUTURE_TARGET]["pos"].cumsum().values
list_nb_cases = list_nb_cases + arr_nb.tolist()
df_plot_pred_all["nb_cases"] = list_nb_cases
# save for future pred
df_plot_pred_all.to_csv(PATH_DF_PLOT_PRED_ALL_KR, index=False)
return df_plot_pred_all
