def trainClf(clf, param_grid_rf):
data_dir = os.path.join(os.path.join('.', 'data'), 'cs-train')
work_dir = os.path.join(os.path.join('.', 'data'), 'work-data')
aDf = ingestTrainData(data_dir)
data = getAllTS(aDf, work_dir)
X, y, dates = engineer_features(data['united_kingdom'])
# Perform a train-test split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.25,
shuffle=True, random_state=42)
# train a random forest model
pipe_rf = Pipeline(steps=[('scaler', StandardScaler()),
('clf', clf)])
grid = GridSearchCV(pipe_rf, param_grid=param_grid_rf,
cv=5, iid=False, n_jobs=-1)
grid.fit(X_train, y_train)
y_train_pred = grid.predict(X_train)
y_pred = grid.predict(X_test)
eval_train_rmse = round(np.sqrt(mean_squared_error(y_train, y_train_pred)))
eval_rmse = round(np.sqrt(mean_squared_error(y_test, y_pred)))
print('eval_rmse_test', eval_rmse)
print('eval_rmse_train', eval_train_rmse)
print(grid.best_estimator_)
return grid
def load_data(inp_dir=None, work_dir=None, training=None):
# load data
idf = ingestTrainData(inp_dir)
ts_data = getAllTS(idf, work_dir)
all_data = {}
for country, df in ts_data.items():
X, y, dates = engineer_features(df, training=training)
dates = np.array([str(d) for d in dates])
all_data[country] = {"X": X, "y": y, "dates": dates}
return all_data
def getFeatures():
# load time series
work_dir = join(data_dir, 'work-data')
ts_file_path = join(work_dir, 'ts-data-all.csv')
df = pd.DataFrame()
if not exists(ts_file_path):
# create time series
idf = ingestTrainData(join('.', 'data'))
df = getTimeSeries(idf)
else:
df = pd.read_csv(ts_file_path)
X, y, dates = engineer_features(df)
return X, y, dates
def model_train(data_dir, test=False, model_dir=None, force_data_load=True):
"""
funtion to train model given a df
'mode' - can be used to subset data essentially simulating a train
"""
work_dir = os.path.join(data_dir, TSDIR)
inp_dir = os.path.join(data_dir, 'cs-train')
if not model_dir:
model_dir = MODEL_DIR
if not os.path.isdir(MODEL_DIR):
os.mkdir(MODEL_DIR)
if test:
print("running training in test mode only uk will be trained")
idf = pd.DataFrame()
if (force_data_load):
# print('loading data from ', inp_dir)
idf = ingestTrainData(inp_dir)
else:
train_path = os.path.join(work_dir, 'train-data-cleaned.csv')
idf = pd.read_csv(train_path)
# fetch time-series formatted data
ts_data = getAllTS(idf, work_dir)
# train a different model for each data sets
for country, df in ts_data.items():
if test and country not in ['all', 'united_kingdom']:
continue
_model_train(df, country, model_dir=model_dir, test=test)
def monitoring():
# load time series
data_dir = join('.', 'data')
model_dir = join('.', 'models')
monitor_dir = join('.', 'monitor')
if not exists(monitor_dir):
os.mkdir(monitor_dir)
work_dir = join(data_dir, 'work-data')
ts_file_path = join(work_dir, 'ts-data-all.csv')
df = pd.DataFrame()
if not exists(ts_file_path):
# create time series
idf = ingestTrainData(join('.', 'data'))
df = getTimeSeries(idf)
else:
df = pd.read_csv(ts_file_path)
X, y, dates = engineer_features(df)
all_data, all_models = model_load(training=False,
data_dir=data_dir,
model_dir=model_dir,
test=False)
results = pd.DataFrame(columns=['date', 'y_pred', 'y', 'diff'])
for idx, d in enumerate(dates):
date = pd.to_datetime(d)
error = False
answ = None
try:
answ = model_predict('all',
str(date.year),
str(date.month),
str(date.day),
test=False,
all_data=all_data,
all_models=all_models)
except:
print('system error:' + str(sys.exc_info()[1]))
error = True
y_pred = None
diff = None
yt = y[idx]
if not error:
y_pred = answ['y_pred'][0]
diff = abs(y_pred - yt)
results = results.append(
{
'date': date,
'y_pred': y_pred,
'y': yt,
'diff': diff
},
ignore_index=True)
# take only the last dates
today = dt.datetime.today()
monname = "model-monitoring-{}-{}-{}".format(today.year, today.month,
today.day)
results.to_csv(join(monitor_dir, monname + ".csv"))
fig, ax = plt.subplots(1, 1)
fig.set_size_inches(15, 8)
ax.set_title('prediction error distribution')
sns.distplot(results['diff'], bins=50, color='#008899', ax=ax)
fig.savefig(join(monitor_dir, monname) + '.png', dpi=200)
statistics_path = join(monitor_dir, 'monitor_statistics.csv')
statDF = pd.DataFrame()
today_iso = today.strftime('%y-%m-%d')
mse = mean_squared_error(results['y'].values, results['y_pred'].values)
if exists(statistics_path):
statDF = pd.read_csv(statistics_path)
found = statDF[statDF['date'] == today_iso]
if (found.shape[0] > 0):
statDF.loc[statDF['date'] == today_iso, ['mse']] = mse
else:
statDF = statDF.append({
'date': today_iso,
'mse': mse
},
ignore_index=True)
statDF.to_csv(statistics_path)