def get_predictions(output):
# Load campaign data
data = pd.read_csv('campaigns.csv')
# Trim data to only include desired output metric
data = tra.trim(data, output)
# Preprocess data without train/test or y/X splitting
data, _ = pre.data_pipeline(data, output)
# Load primary models
direct_model = joblib.load('./models/' + output + '_model.pkl')
direct_columns = joblib.load('./models/' + output + '_columns.pkl')
cpx_model = \
joblib.load('./models/' + 'cost_per_' + output[0:-1] + '_model.pkl')
cpx_columns = \
joblib.load('./models/' + 'cost_per_' + output[0:-1] + '_columns.pkl')
# Calculate and save primary predictions
predictions = pd.DataFrame(columns=[output, 'direct', 'cpx'])
for index, row in data.iterrows():
direct_row = pd.DataFrame([dict(row)]) \
.reindex(columns=direct_columns, fill_value=0)
cpx_row = pd.DataFrame([dict(row)]) \
.reindex(columns=cpx_columns, fill_value=0)
direct_prediction = int(direct_model.predict(direct_row)[0])
cpx_prediction = int(row['cost'] / cpx_model.predict(cpx_row)[0])
predictions.loc[index] = [
row[output], direct_prediction, cpx_prediction
]
return predictions
def train(output,
update=False,
upload=False,
models=['linear', 'forest', 'svr']):
"""Complete training pipeline"""
if update:
postgres_to_csv()
print('Data updated.')
data = pd.read_csv('campaigns.csv')
print('Data loaded.')
data = trim(data, output)
print('Data trimmed.')
data, data_cat = pre.data_pipeline(data, output)
[X, y, X_train, y_train, X_test, y_test, X_scaled, y_scaled,
X_train_scaled, y_train_scaled, X_test_scaled, y_scaler] \
= pre.split_pipeline(data, output, encoded=True)
[X_cat, y_cat, X_train_cat, y_train_cat, X_test_cat, y_test_cat] = \
pre.split_pipeline(data_cat, output, encoded=False)
print('Data preprocessed.')
regressors = build(X_train, y_train, X_train_scaled, y_train_scaled,
X_train_cat, y_train_cat, models)
best_regressor = evaluate(regressors, X_train, y_train, X_train_scaled,
y_train_scaled, X_test, y_test, X_test_scaled,
y_scaler, X_train_cat, y_train_cat, X_test_cat,
y_test_cat)
print('Regressors evaluated. Best regressor is:\n' + str(best_regressor))
if 'SVR' in str(best_regressor):
best_regressor.fit(X_scaled, y_scaled)
elif 'Cat' in str(best_regressor):
best_regressor.fit(X_cat, y_cat)
else:
best_regressor.fit(X, y)
print('Regressor fit.')
print_results(best_regressor, X, X_scaled, y, y_scaler, X_cat)
save(best_regressor, X, output)
print('Regressor saved.')
if upload:
upload_to_s3(output)
print('Regressor uploaded.')
def train(output, source='pg', models=['linear', 'tree', 'forest', 'svr']):
"""Complete training pipeline"""
if source == 'pg':
data = load_data_from_postgres(output)
print('Data loaded from Postgres.')
elif source == 'csv':
data = load_data_from_csv(output)
print('Data loaded from CSV.')
else:
print('Source not available.')
return
data = pre.data_pipeline(data, output)
[X, y, X_train, y_train, X_test, y_test, X_scaled, y_scaled,
X_train_scaled, y_train_scaled, X_test_scaled, y_scaler] \
= pre.split_pipeline(data, output)
print('Data preprocessed.')
regressors = build(X_train, y_train, X_train_scaled, y_train_scaled,
models)
best_regressor = evaluate(regressors, X_train, y_train, X_train_scaled,
y_train_scaled, X_test, y_test, X_test_scaled,
y_scaler)
print('Regressors evaluated. Best regressor is:\n' + str(best_regressor))
if 'SVR' in str(best_regressor):
best_regressor.fit(X_scaled, y_scaled)
else:
best_regressor.fit(X, y)
print('Regressor fit.')
print_results(best_regressor, X, X_scaled, y, y_scaler)
save(best_regressor, X, output)
print('Regressor saved.')
upload(output)
print('Regressor uploaded.')
def calculate(output, model):
"""Determine feature importance for specified model"""
data = pd.read_csv('campaigns.csv')
print('Data loaded.')
data = tra.trim(data, output)
print('Data trimmed.')
# Add random column to data
np.random.seed(seed=0)
data['random'] = np.random.random(size=len(data))
data, data_cat = pre.data_pipeline(data, output)
[_, _, X_train, y_train, _, _, _, _,
X_train_scaled, y_train_scaled, _, y_scaler] \
= pre.split_pipeline(data, output, encoded=True)
[_, _, X_train_cat, y_train_cat, _, _] = \
pre.split_pipeline(data_cat, output, encoded=False)
print('Data preprocessed.')
regressor = tra.build(X_train, y_train, X_train_scaled, y_train_scaled,
X_train_cat, y_train_cat, [model])[0]
model_clone = clone(regressor)
# Set random_state for comparability
model_clone.random_state = 0
# Train and score the benchmark model
if 'SVR' in str(regressor):
model_clone.fit(X_train_scaled, y_train_scaled)
benchmark_score = hel.mean_relative_accuracy(
y_scaler.inverse_transform(model_clone.predict(X_train_scaled)),
y_train)
else:
model_clone.fit(X_train, y_train)
benchmark_score = \
hel.mean_relative_accuracy(model_clone.predict(X_train), y_train)
# Calculate and store feature importance benchmark deviation
importances = []
columns = X_train.columns
i = 1
for column in columns:
model_clone = clone(regressor)
model_clone.random_state = 0
if 'SVR' in str(regressor):
model_clone.fit(X_train_scaled.drop(column, axis=1),
y_train_scaled)
drop_col_score = hel.mean_relative_accuracy(
model_clone.predict(X_train_scaled.drop(column, axis=1)),
y_train_scaled)
else:
model_clone.fit(X_train.drop(column, axis=1), y_train)
drop_col_score = hel.mean_relative_accuracy(
model_clone.predict(X_train.drop(column, axis=1)), y_train)
importances.append(benchmark_score - drop_col_score)
i += 1
importances_df = \
pd.DataFrame({'column': X_train.columns, 'value': importances}) \
.sort_values('value', ascending=False).reset_index(drop=True)
print('Importances:')
for i in range(0, len(importances_df)):
print(
str(importances_df.iloc[i].column) + ': ' +
str(importances_df.iloc[i].value))