def main():
"""
Get data from db and save it as csv
"""
# Helpers
bq = bqhandler.BQHandler()
io = IO(gs_bucket=options.gs_bucket)
viz = Viz(io)
predictor = Predictor(io, ModelLoader(io), options)
### OPTIONS ################################################################
# Configuration
starttime, endtime = io.get_dates(options)
logging.info('Using dataset {} and time range {} - {}'.format(options.feature_dataset,
starttime.strftime('%Y-%m-%d'),
endtime.strftime('%Y-%m-%d')))
all_param_names = options.label_params + options.feature_params + options.meta_params
aggs = io.get_aggs_from_param_names(options.feature_params)
### MODELS #################################################################
# Initialise classifier
if hasattr(options, 'classifier_file'):
classifier = io.load_scikit_model(options.classifier_file)
else:
if options.classifier == 'svc':
params = {'kernel': options.kernel, 'gamma': options.gamma, 'C': options.penalty, 'probability': options.probability}
#classifier = SVC(**params)
classifier = SVCClassifier(params, limit=options.class_limit)
elif options.classifier == 'graphsvc':
classifier = GraphSVCClassifier()
graph_data = pd.read_csv(options.graph_data, names=['date', 'start_hour', 'src', 'dst', 'type', 'sum_delay','sum_ahead','add_delay','add_ahead','train_count'])
classifier.fetch_connections(graph_data)
elif options.classifier == 'gaussiannb':
classifier = GaussianNBClassifier()
elif options.classifier == 'lstm':
num_of_features = len(options.feature_params)
if options.month: num_of_features += 1
class_weight=None
if hasattr(options, 'class_weight'):
class_weight=eval(options.class_weight)
params = {'length': options.time_steps, 'batch_size': options.batch_size, 'epochs': options.epochs, 'num_of_features': num_of_features, 'log_dir': options.log_dir, 'class_weight':class_weight}
classifier = LSTMClassifier(**params)
else:
raise('Model not specificied or wrong. Add "classifier: bgm" to config file.')
# Initialise regression model
if options.regression == 'rfr':
regressor = RandomForestRegressor(n_estimators=options.n_estimators,
n_jobs=-1,
min_samples_leaf=options.min_samples_leaf,
min_samples_split=options.min_samples_split,
max_features=options.max_features,
max_depth=options.max_depth,
bootstrap=options.bootstrap
)
#regressor = _trans.Regressor(model=model)
else:
raise('Model not specificied or wrong. Add "classifier: bgm" to config file.')
# Initialise transformer
#transformer = _trans.Selector(classifier=classifier)
# Initialise pipeline
#pipe = Pipeline(
# [('selector', transformer),
# ('regression', regressor)]
#)
### DATA ###################################################################
sum_columns = ['delay']
if 'train_count' in options.meta_params:
sum_columns.append('train_count')
# Pick only selected month
where = {}
if options.pick_month is not None:
where = {'EXTRACT(MONTH from time)': options.pick_month}
logging.info('Reading data...')
bq.set_params(loc_col='trainstation',
project=options.project,
dataset=options.feature_dataset,
table=options.feature_table,
parameters=all_param_names,
locations=options.locations,
only_winters=options.only_winters,
reason_code_table=options.reason_code_table,
where=where)
data = bq.get_rows(starttime,
endtime)
data = io.filter_train_type(labels_df=data,
train_types=options.train_types,
sum_types=True,
train_type_column='train_type',
location_column='trainstation',
time_column='time',
sum_columns=sum_columns,
aggs=aggs)
data['delay'] = data.loc[:, 'delay'].replace(-99, np.nan)
data.sort_values(by=['trainstation', 'time'], inplace=True)
logging.info('Processing {} rows...'.format(len(data)))
# Filter only timesteps with large distribution in the whole network
if options.filter_delay_limit is not None:
data = io.filter_delay_with_limit(data, options.filter_delay_limit)
# Binary class
logging.info('Adding binary class to the dataset with limit {}...'.format(options.delay_limit))
def set_class(x):
if x > options.delay_limit:
return binary_labels[1]
elif x < options.delay_limit:
return binary_labels[0]
return np.nan
data['class'] = data['delay'].map(lambda x: set_class(x))
# Separate train and validation sets
data_train, data_test = train_test_split(data, test_size=0.3, shuffle=False)
# Balance
if options.balance:
logging.info('Balancing training data...')
count = data_train.groupby('class').size().min()
# SVC can't handle more than 50 000 samples
if options.classifier == 'svc': count = min(count, 50000)
data_train = pd.concat([data_train.loc[data_train['class'] == 0].sample(n=count),
data_train.loc[data_train['class'] == 1].sample(n=count)])
logging.info('Train data:')
io.log_class_dist(data_train.loc[:, 'class'].values, labels=binary_labels)
logging.info('Test data:')
io.log_class_dist(data_test.loc[:, 'class'].values, labels=binary_labels)
# Adding month
if options.month:
logging.info('Adding month to the datasets...')
data_train['month'] = data_train.loc[:,'time'].map(lambda x: x.month)
data_test['month'] = data_test.loc[:,'time'].map(lambda x: x.month)
options.feature_params.append('month')
#data_train.set_index('time', inplace=True)
#y_train_class = data_train.loc[:,['class']].astype(np.int32).values.ravel()
#y_train_delay = data_train.loc[:,['delay']].astype(np.int32).values.ravel()
y_train_class = data_train.loc[:,['class']].values.ravel()
y_train_delay = data_train.loc[:,['delay']].values.ravel()
#y_test_class = data_test.loc[:,['class']].astype(np.int32).values.ravel()
#y_test_delay = data_test.loc[:,['delay']].astype(np.int32).values.ravel()
y_test_class = data_test.loc[:,['class']].values.ravel()
y_test_delay = data_test.loc[:,['delay']].values.ravel()
X_train = data_train.loc[:,options.feature_params].astype(np.float32).values
X_test = data_test.loc[:,options.feature_params].astype(np.float32).values
if options.smote:
logging.info('Smoting...')
sm = SMOTE()
X_train_class, y_class = sm.fit_resample(X_train, y_train_class)
io.log_class_dist(y_class, labels=binary_labels)
# X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3)
# io.log_class_dist(y_train[:,1], [-1,1])
# If asked, save used train and test splits into big query
if options.save_data:
tname = options.model+'_'+options.feature_dataset+'_'+options.config_name+'_train'
columns = [options.feature_params, ['delay'], ['class']]
bq.nparray_to_table([X_train, y_train_class, y_train_delay],
columns,
options.project,
options.feature_dataset,
tname
)
tname = options.model+'_'+options.feature_dataset+'_'+options.config_name+'_test'
bq.nparray_to_table([X_test, y_test_class, y_test_delay],
columns,
options.project,
options.feature_dataset,
tname
)
if options.normalize:
logging.info('Normalizing data...')
#scale=(0,1)
if hasattr(options, 'xscaler_file'):
xscaler = io.load_scikit_model(options.xscaler_file)
X_train = xscaler.transform(X_train)
X_test = xscaler.transform(X_test)
else:
xscaler = MinMaxScaler(feature_range=(-1,1))
#xscaler = StandardScaler()
X_train = xscaler.fit_transform(X_train)
X_test = xscaler.transform(X_test)
fname = options.save_path+'/xscaler.pkl'
io.save_scikit_model(xscaler, fname, fname)
if hasattr(options, 'yscaler_file'):
yscaler = io.load_scikit_model(options.yscaler_file)
y_train_delay = yscaler.transform(y_train_delay)
y_test_delay = yscaler.transform(y_test_delay)
else:
#yscaler = MinMaxScaler(feature_range=(0,1))
yscaler=StandardScaler()
y_train_delay = yscaler.fit_transform(y_train_delay.reshape(-1,1)).ravel()
y_test_delay = yscaler.transform(y_test_delay.reshape(-1,1)).ravel()
fname = options.save_path+'/yscaler.pkl'
io.save_scikit_model(yscaler, fname, fname)
data_train.loc[:,options.feature_params].to_csv('data/x_train.csv', index=False)
data_test.loc[:,options.feature_params].to_csv('data/x_test.csv', index=False)
data_train.loc[:,['class']].fillna(-99).astype(np.int).to_csv('data/y_train.csv', index=False)
data_test.loc[:,['class']].fillna(-99).astype(np.int).to_csv('data/y_test.csv', index=False)
sys.exit()
### TRAIN ##################################################################
if options.cv:
logging.info('Doing random search for hyper parameters...')
raise("No param_grid set for given model ({})".format(options.regression))
random_search = RandomizedSearchCV(model,
param_distributions=param_grid,
n_iter=int(options.n_iter_search),
n_jobs=-1)
random_search.fit(X_train, y_train)
logging.info("RandomizedSearchCV done.")
fname = options.output_path+'/random_search_cv_results.txt'
report_cv_results(random_search.cv_results_, fname)
io._upload_to_bucket(filename=fname, ext_filename=fname)
sys.exit()
else:
logging.info('Training classifier...')
if options.classifier == 'graphsvc':
classifier.fit(X_train, y_train_class, stations=data_train.loc[:, 'trainstation'].values)
else:
history = classifier.fit(X_train, y_train_class, X_test, y_test_class)
# Save classifier
if options.classifier == 'lstm':
history_fname = options.save_path+'/history.pkl'
fname = options.save_path+'/classifier.h5'
io.save_keras_model(fname, history_fname, classifier, history.history)
else:
fname = options.save_path+'/classifier.pkl'
io.save_scikit_model(classifier, filename=fname, ext_filename=fname)
# Drop data with no delay information
X_train = X_train[~np.isnan(y_train_delay)]
y_train_delay = y_train_delay[~np.isnan(y_train_delay)]
y_train_class = y_train_class[~np.isnan(y_train_class)]
y_pred_train_bin = classifier.predict(X_train, type='bool')
# debug
#y_pred_train_bin
#indices = np.random.choice(np.arange(y_pred_train_bin.size),
# replace=False,
# size=int(y_pred_train_bin.size * 0.2))
#y_pred_train_bin[indices] = True
#print('y_pred_train_bin: {}'.format(y_pred_train_bin.shape))
#print('y_train_delay: {}'.format(y_train_delay.shape))
#print('y_train_class: {}'.format(y_train_class.shape))
# Pick only severe values
#y_train_delay_ = y_train_delay[(len(y_train_class)-len(y_pred_train_bin)):]
#X_train_ = X_train[(len(y_train_class)-len(y_pred_train_bin)):]
y_train_delay_ = y_train_delay[(len(y_train_delay)-len(y_pred_train_bin)):]
X_train_ = X_train[(len(y_train_delay)-len(y_pred_train_bin)):]
#print('y_train_delay_: {}'.format(y_train_delay_.shape))
y_train_severe = y_train_delay_[y_pred_train_bin]
X_train_severe = X_train_[y_pred_train_bin]
logging.info('Training regressor...')
regressor.fit(X_train_severe, y_train_severe)
# Save regressor
io.save_scikit_model(regressor, filename=options.save_file, ext_filename=options.save_file)
# Learning history
# fname = options.output_path+'/learning_over_time.png'
# viz.plot_nn_perf(history.history, metrics={'Error': {'mean_squared_error': 'MSE',
# 'mean_absolute_error': 'MAE'}},
# filename=fname)
### RESULTS FOR VALIDATION SET #############################################
# Drop data with missing delay
X_test = X_test[~np.isnan(y_test_class)]
y_test_class = y_test_class[~np.isnan(y_test_class)]
data_test = data_test[~np.isnan(data_test.delay)]
# Metrics
#y_pred_proba = classifier.predict_proba(X_test)
y_pred = classifier.predict(X_test)
y_pred_proba = classifier.y_pred_proba
#y_test_delay = y_test_delay[~np.isnan(y_test_delay)]
# Classification performance
# LSTM don't have first timesteps
y_test_class = y_test_class[(len(X_test)-len(y_pred)):]
acc = accuracy_score(y_test_class, y_pred)
precision = precision_score(y_test_class, y_pred, average='micro')
recall = recall_score(y_test_class, y_pred, average='micro')
f1 = f1_score(y_test_class, y_pred, average='micro')
logging.info('Classification accuracy: {}'.format(acc))
logging.info('Classification precision: {}'.format(precision))
logging.info('Classification recall: {}'.format(recall))
logging.info('Classification F1 score: {}'.format(f1))
io.log_class_dist(y_pred, binary_labels)
# Confusion matrices
fname = '{}/confusion_matrix_validation.png'.format(options.output_path)
viz.plot_confusion_matrix(y_test_class, y_pred, np.arange(2), filename=fname)
fname = '{}/confusion_matrix_validation_normalised.png'.format(options.output_path)
viz.plot_confusion_matrix(y_test_class, y_pred, np.arange(2), True, filename=fname)
# Precision-recall curve
fname = '{}/precision-recall-curve_validation.png'.format(options.output_path)
viz.prec_rec_curve(y_test_class, y_pred_proba, filename=fname)
# ROC
fname = '{}/roc_validation.png'.format(options.output_path)
viz.plot_binary_roc(y_test_class, y_pred_proba, filename=fname)
if options.regression == 'rfr':
fname = options.output_path+'/rfc_feature_importance.png'
viz.rfc_feature_importance(regressor.feature_importances_, fname, feature_names=options.feature_params)
# Regression performance
y_pred_reg, y_test_reg = predictor.pred(data=data_test)
#y_test_reg = y_test[(len(y_test)-len(y_pred)):,0]
rmse = np.sqrt(mean_squared_error(y_test_reg, y_pred))
mae = mean_absolute_error(y_test_reg, y_pred)
r2 = r2_score(y_test_reg, y_pred)
logging.info('Regression RMSE: {}'.format(rmse))
logging.info('Regression MAE: {}'.format(mae))
logging.info('Regression R2 score: {}'.format(r2))
error_data = {'acc': [acc],
'precision': [precision],
'recall': [recall],
'f1': [f1],
'rmse': [rmse],
'mae': [mae],
'r2': [r2]}
fname = '{}/training_time_classification_validation_errors.csv'.format(options.output_path)
io.write_csv(error_data, filename=fname, ext_filename=fname)
############################################################################
# EVALUATE
############################################################################
if options.evaluate:
logging.info('Loading test data...')
test_data = bq.get_rows(dt.datetime.strptime('2010-01-01', "%Y-%m-%d"),
dt.datetime.strptime('2019-01-01', "%Y-%m-%d"),
loc_col='trainstation',
project=options.project,
dataset=options.feature_dataset,
table=options.test_table,
reason_code_table=options.reason_code_table,
locations=options.locations,
parameters=all_param_names)
test_data = io.filter_train_type(labels_df=test_data,
train_types=['K','L'],
sum_types=True,
train_type_column='train_type',
location_column='trainstation',
time_column='time',
sum_columns=['delay'],
aggs=aggs)
# Sorting is actually not necessary. It's been useful for debugging.
#test_data.sort_values(by=['time', 'trainstation'], inplace=True)
# Filter only timesteps with large distribution in the whole network
if options.filter_delay_limit is not None:
test_data = io.filter_delay_with_limit(test_data, options.filter_delay_limit)
test_data.set_index('time', inplace=True)
logging.info('Test data contain {} rows...'.format(len(test_data)))
logging.info('Adding binary class to the test dataset with limit {}...'.format(options.delay_limit))
test_data['class'] = test_data['delay'].map(lambda x: binary_labels[1] if x > options.delay_limit else binary_labels[0])
io.log_class_dist(test_data.loc[:, 'class'].values, labels=binary_labels)
if options.month:
logging.info('Adding month to the test dataset...')
test_data['month'] = test_data.index.map(lambda x: x.month)
times = [('2014-01-01', '2014-02-01'), ('2016-06-01', '2016-07-01'), ('2017-02-01', '2017-03-01'), ('2011-02-01', '2011-03-01')]
for start, end in times:
try:
y_pred_proba, y_pred, y = predict_timerange(test_data, options.feature_params, classifier, xscaler, start, end)
perf_metrics(y_pred_proba, y_pred, y, start, end, viz, io)
except EmptyDataError:
logging.info('No data for {} - {}'.format(start, end))
def main():
"""
Main program
"""
# Print GPU availability
local_device_protos = device_lib.list_local_devices()
logging.info(
[x.name for x in local_device_protos if x.device_type == 'GPU'])
bq = BQHandler()
io = IO(gs_bucket=options.gs_bucket)
viz = Viz(io)
starttime, endtime = io.get_dates(options)
logging.info('Using dataset {}.{} and time range {} - {}'.format(
options.feature_dataset, options.feature_table,
starttime.strftime('%Y-%m-%d'), endtime.strftime('%Y-%m-%d')))
all_param_names = list(
set(options.label_params + options.feature_params +
options.meta_params))
aggs = io.get_aggs_from_param_names(options.feature_params)
logging.info('Building model...')
dim = len(options.feature_params)
if options.month: dim += 1
model = convlstm.Regression(options, dim).get_model()
logging.info('Reading data...')
bq.set_params(batch_size=2500000,
loc_col='trainstation',
project=options.project,
dataset=options.feature_dataset,
table=options.feature_table,
parameters=all_param_names,
locations=options.train_stations,
only_winters=options.only_winters,
reason_code_table=options.reason_code_table)
data = bq.get_rows(starttime, endtime)
data = io.filter_train_type(labels_df=data,
train_types=options.train_types,
sum_types=True,
train_type_column='train_type',
location_column='trainstation',
time_column='time',
sum_columns=['train_count', 'delay'],
aggs=aggs)
if options.y_avg_hours is not None:
data = io.calc_running_delay_avg(data, options.y_avg_hours)
if options.y_avg:
data = io.calc_delay_avg(data)
data.sort_values(by=['time', 'trainstation'], inplace=True)
if options.month:
logging.info('Adding month to the dataset...')
data['month'] = data['time'].map(lambda x: x.month)
options.feature_params.append('month')
if options.normalize:
logging.info('Normalizing data...')
xscaler = StandardScaler()
yscaler = StandardScaler()
labels = data.loc[:, options.label_params].astype(
np.float32).values.reshape((-1, 1))
yscaler.fit(labels)
scaled_labels = pd.DataFrame(yscaler.transform(labels),
columns=['delay'])
non_scaled_data = data.loc[:, options.meta_params + ['class']]
scaled_features = pd.DataFrame(xscaler.fit_transform(
data.loc[:, options.feature_params]),
columns=options.feature_params)
data = pd.concat([non_scaled_data, scaled_features, scaled_labels],
axis=1)
fname = options.save_path + '/xscaler.pkl'
io.save_scikit_model(xscaler, fname, fname)
fname = options.save_path + '/yscaler.pkl'
io.save_scikit_model(yscaler, fname, fname)
if options.pca:
logging.info('Doing PCA analyzis for the data...')
ipca = IncrementalPCA(n_components=options.pca_components,
whiten=options.whiten,
copy=False)
non_processed_data = data.loc[:, options.meta_params +
options.label_params]
processed_data = data.loc[:, options.feature_params]
ipca.fit(processed_data)
processed_features = pd.DataFrame(ipca.transform(processed_data))
data = pd.concat([non_processed_data, processed_data], axis=1)
fname = options.output_path + '/ipca_explained_variance.png'
viz.explained_variance(ipca, fname)
io._upload_to_bucket(filename=fname, ext_filename=fname)
data_train, data_test = train_test_split(data, test_size=0.33)
# Define model
batch_size = 512
logging.info('Batch size: {}'.format(batch_size))
# Initialization
losses, val_losses, accs, val_accs, steps = [], [], [], [], []
boardcb = TensorBoard(log_dir=options.log_dir + '/lstm',
histogram_freq=0,
write_graph=True,
write_images=True)
logging.info('Data shape: {}'.format(
data_train.loc[:, options.feature_params].values.shape))
data_gen = TimeseriesGenerator(
data_train.loc[:, options.feature_params].values,
data_train.loc[:, options.label_params].values,
length=24,
sampling_rate=1,
batch_size=batch_size)
data_test_gen = TimeseriesGenerator(
data_test.loc[:, options.feature_params].values,
data_test.loc[:, options.label_params].values,
length=24,
sampling_rate=1,
batch_size=batch_size)
logging.info('X batch size: {}'.format(data_gen[0][0].shape))
logging.info('Y batch size: {}'.format(data_gen[1][0].shape))
history = model.fit_generator(data_gen,
validation_data=data_test_gen,
epochs=options.epochs,
callbacks=[boardcb]) #, batch_size=64)
history_fname = options.save_path + '/history.pkl'
io.save_keras_model(options.save_file, history_fname, model,
history.history)
scores = model.evaluate_generator(data_test_gen)
i = 0
error_data = {}
for name in model.metrics_names:
logging.info('{}: {:.4f}'.format(name, scores[i]))
error_data[name] = [scores[i]]
i += 1
fname = '{}/training_time_validation_errors.csv'.format(
options.output_path)
io.write_csv(error_data, filename=fname, ext_filename=fname)
pred = model.predict_generator(data_test_gen)
#io.log_class_dist(pred, 4)
#print(history.history)
fname = options.output_path + '/learning_over_time.png'
viz.plot_nn_perf(history.history,
metrics={
'Error': {
'mean_squared_error': 'MSE',
'mean_absolute_error': 'MAE'
}
},
filename=fname)
}
logging.basicConfig(format=(
"[%(levelname)s] %(asctime)s %(filename)s:%(funcName)s:%(lineno)s %(message)s"
),
level=logging_level[options.logging_level])
logging.info('Using configuration: {} | {}'.format(options.config_filename,
options.config_name))
# Helpers
bq = bqhandler.BQHandler()
io = IO(gs_bucket=options.gs_bucket)
viz = Viz(io)
state = State()
starttime, endtime = io.get_dates(options)
logging.info('Using dataset {} and time range {} - {}'.format(
options.feature_dataset, starttime.strftime('%Y-%m-%d'),
endtime.strftime('%Y-%m-%d')))
if options.save_data:
tname = options.model + '_' + options.feature_dataset + '_' + options.config_name + '_train'
tname = tname.replace('-', '_')
bq.delete_table(options.project, options.feature_dataset, tname)
tname = options.model + '_' + options.feature_dataset + '_' + options.config_name + '_test'
tname = tname.replace('-', '_')
bq.delete_table(options.project, options.feature_dataset, tname)
if (options.station_specific_classifier
or options.station_specific_regressor
) and options.locations is not None:
def main():
"""
Get data from db and save it as csv
"""
bq = BQHandler()
io = IO(gs_bucket=options.gs_bucket)
viz = Viz(io=io)
starttime, endtime = io.get_dates(options)
logging.info('Using dataset {} and time range {} - {}'.format(
options.feature_dataset, starttime.strftime('%Y-%m-%d'),
endtime.strftime('%Y-%m-%d')))
all_param_names = options.label_params + options.feature_params + options.meta_params
aggs = io.get_aggs_from_param_names(options.feature_params)
if options.model == 'rf':
model = RandomForestRegressor(
n_estimators=options.n_estimators,
n_jobs=-1,
min_samples_leaf=options.min_samples_leaf,
min_samples_split=options.min_samples_split,
max_features=options.max_features,
max_depth=options.max_depth,
bootstrap=options.bootstrap)
elif options.model == 'lr':
model = SGDRegressor(warm_start=True,
max_iter=options.n_loops,
shuffle=options.shuffle,
power_t=options.power_t,
penalty=options.regularizer,
learning_rate=options.learning_rate,
eta0=options.eta0,
alpha=options.alpha,
tol=0.0001)
elif options.model == 'svr':
model = SVR()
elif options.model == 'ard':
model = ARDRegression(n_iter=options.n_loops,
alpha_1=options.alpha_1,
alpha_2=options.alpha_2,
lambda_1=options.lambda_1,
lambda_2=options.lambda_2,
threshold_lambda=options.threshold_lambda,
fit_intercept=options.fit_intercept,
copy_X=options.copy_X)
elif options.model == 'gp':
k_long_term = 66.0**2 * RBF(length_scale=67.0)
k_seasonal = 2.4**2 * RBF(length_scale=90.0) * ExpSineSquared(
length_scale=150, periodicity=1.0, periodicity_bounds=(0, 10000))
k_medium_term = 0.66**2 * RationalQuadratic(length_scale=1.2,
alpha=0.78)
k_noise = 0.18**2 * RBF(length_scale=0.134) + WhiteKernel(
noise_level=0.19**2)
#kernel_gpml = k_long_term + k_seasonal + k_medium_term + k_noise
kernel_gpml = k_long_term + k_seasonal + k_medium_term + k_noise
model = GaussianProcessRegressor(
kernel=kernel_gpml, #alpha=0,
optimizer=None,
normalize_y=True)
elif options.model == 'llasso':
model = LocalizedLasso(num_iter=options.n_loops,
batch_size=options.batch_size)
elif options.model == 'nlasso':
model = NetworkLasso(num_iter=options.n_loops,
batch_size=options.batch_size)
graph_data = pd.read_csv(options.graph_data,
names=[
'date', 'start_hour', 'src', 'dst',
'type', 'sum_delay', 'sum_ahead',
'add_delay', 'add_ahead', 'train_count'
])
#stations_to_pick = options.stations_to_pick.split(',')
#graph = model.fetch_connections(graph_data, stations_to_pick)
model.fetch_connections(graph_data)
if options.pca:
ipca = IncrementalPCA(n_components=options.pca_components,
whiten=options.whiten,
copy=False)
rmses, maes, r2s, skills, start_times, end_times, end_times_obj = [], [], [], [], [], [], []
X_complete = [] # Used for feature selection
start = starttime
end = start + timedelta(days=int(options.day_step),
hours=int(options.hour_step))
if end > endtime: end = endtime
while end <= endtime and start < end:
logging.info('Processing time range {} - {}'.format(
start.strftime('%Y-%m-%d %H:%M'), end.strftime('%Y-%m-%d %H:%M')))
# Load data ############################################################
try:
logging.info('Reading data...')
data = bq.get_rows(start,
end,
loc_col='trainstation',
project=options.project,
dataset=options.feature_dataset,
table=options.feature_table,
parameters=all_param_names,
only_winters=options.only_winters)
data = io.filter_train_type(labels_df=data,
train_types=options.train_types,
sum_types=True,
train_type_column='train_type',
location_column='trainstation',
time_column='time',
sum_columns=['train_count', 'delay'],
aggs=aggs)
# Filter only timesteps with large distribution in the whole network
if options.filter_delay_limit is not None:
data = io.filter_delay_with_limit(data,
options.filter_delay_limit)
if options.y_avg_hours is not None:
data = io.calc_running_delay_avg(data, options.y_avg_hours)
if options.y_avg:
data = io.calc_delay_avg(data)
data.sort_values(by=['time', 'trainstation'], inplace=True)
if options.impute:
logging.info('Imputing missing values...')
data.drop(columns=['train_type'], inplace=True)
data = imputer.fit_transform(data)
data.loc[:, 'train_type'] = None
if options.month:
logging.info('Adding month to the dataset...')
data['month'] = data['time'].map(lambda x: x.month)
if 'month' not in options.feature_params:
options.feature_params.append('month')
if options.model == 'ard' and len(data) > options.n_samples:
logging.info('Sampling {} values from data...'.format(
options.n_samples))
data = data.sample(options.n_samples)
l_data = data.loc[:, options.label_params]
f_data = data.loc[:, options.feature_params]
except ValueError as e:
f_data, l_data = [], []
if len(f_data) < 2 or len(l_data) < 2:
start = end
end = start + timedelta(days=int(options.day_step),
hours=int(options.hour_step))
continue
logging.info('Processing {} rows...'.format(len(f_data)))
train, test = train_test_split(data, test_size=0.1)
X_train = train.loc[:,
options.feature_params].astype(np.float32).values
y_train = train.loc[:, options.label_params].astype(
np.float32).values.ravel()
X_test = test.loc[:, options.feature_params].astype(np.float32).values
y_test = test.loc[:, options.label_params].astype(
np.float32).values.ravel()
logging.debug('Features shape: {}'.format(X_train.shape))
if options.normalize:
logging.info('Normalizing data...')
xscaler, yscaler = StandardScaler(), StandardScaler()
X_train = xscaler.fit_transform(X_train)
X_test = xscaler.transform(X_test)
if len(options.label_params) == 1:
y_train = yscaler.fit_transform(y_train.reshape(-1, 1)).ravel()
#y_test = yscaler.transform(y_test.reshape(-1, 1)).ravel()
else:
y_train = yscaler.fit_transform(y_train)
#y_test = yscaler.transform(y_test)
if options.pca:
logging.info('Doing PCA analyzis for the data...')
X_train = ipca.fit_transform(X_train)
fname = options.output_path + '/ipca_explained_variance.png'
viz.explained_variance(ipca, fname)
#io._upload_to_bucket(filename=fname, ext_filename=fname)
X_test = ipca.fit_transform(X_test)
if options.model == 'llasso':
graph_data = pd.read_csv(options.graph_data,
names=[
'date', 'start_hour', 'src', 'dst',
'type', 'sum_delay', 'sum_ahead',
'add_delay', 'add_ahead',
'train_count'
])
graph = model.fetch_connections(graph_data)
logging.debug('Features shape after pre-processing: {}'.format(
X_train.shape))
# FIT ##################################################################
if options.cv:
logging.info('Doing random search for hyper parameters...')
if options.model == 'rf':
param_grid = {
"n_estimators": [10, 100, 200, 800],
"max_depth": [3, 20, None],
"max_features": ["auto", "sqrt", "log2", None],
"min_samples_split": [2, 5, 10],
"min_samples_leaf": [1, 2, 4, 10],
"bootstrap": [True, False]
}
elif options.model == 'lr':
param_grid = {
"penalty": [None, 'l2', 'l1'],
"alpha": [0.00001, 0.0001, 0.001, 0.01, 0.1],
"l1_ratio": [0.1, 0.15, 0.2, 0.5],
"shuffle": [True, False],
"learning_rate": ['constant', 'optimal', 'invscaling'],
"eta0": [0.001, 0.01, 0.1],
"power_t": [0.1, 0.25, 0.5]
}
elif options.model == 'svr':
param_grid = {
"C": [0.001, 0.01, 0.1, 1, 10],
"epsilon": [0.01, 0.1, 0.5],
"kernel":
['rbf', 'linear', 'poly', 'sigmoid', 'precomputed'],
"degree": [2, 3, 4],
"shrinking": [True, False],
"gamma": [0.001, 0.01, 0.1],
"coef0": [0, 0.1, 1]
}
else:
raise ("No param_grid set for given model ({})".format(
options.model))
random_search = RandomizedSearchCV(model,
param_distributions=param_grid,
n_iter=int(
options.n_iter_search),
n_jobs=-1)
random_search.fit(X_train, y_train)
logging.info("RandomizedSearchCV done.")
fname = options.output_path + '/random_search_cv_results.txt'
io.report_cv_results(random_search.cv_results_, fname)
#io._upload_to_bucket(filename=fname, ext_filename=fname)
sys.exit()
else:
logging.info('Training...')
if options.model in ['rf', 'svr', 'ard', 'gp']:
model.fit(X_train, y_train)
if options.feature_selection:
X_complete = X_train
y_complete = y_train
meta_complete = data.loc[:, options.meta_params]
elif options.model in ['llasso']:
model.fit(X_train,
y_train,
stations=train.loc[:, 'trainstation'].values)
elif options.model in ['nlasso']:
model.partial_fit(X_train,
y_train,
stations=train.loc[:, 'trainstation'].values)
else:
model.partial_fit(X_train, y_train)
if options.feature_selection:
try:
X_complete = np.append(X_complete, X_train)
y_complete = np.append(Y_complete, y_train)
meta_complete = meta_complete.append(
data.loc[:, options.meta_params])
except (ValueError, NameError):
X_complete = X_train
y_complete = y_train
meta_complete = data.loc[:, options.meta_params]
# EVALUATE #############################################################
# Check training score to estimate amount of overfitting
# Here we assume that we have a datetime index (from time columns)
y_pred_train = model.predict(X_train)
rmse_train = np.sqrt(mean_squared_error(y_train, y_pred_train))
mae_train = np.sqrt(mean_squared_error(y_train, y_pred_train))
logging.info('Training data RMSE: {} and MAE: {}'.format(
rmse_train, mae_train))
#try:
if True:
print(train)
#range = ('2013-02-01','2013-02-28')
range = ('2010-01-01', '2010-01-02')
X_train_sample = train.loc[range[0]:range[1],
options.feature_params].astype(
np.float32).values
target = train.loc[range[0]:range[1], options.label_params].astype(
np.float32).values.ravel()
y_pred_sample = model.predict(X_train_sample)
times = train.loc[range[0]:range[1], 'time'].values
df = pd.DataFrame(times + y_pred_sample)
print(df)
sys.exit()
# Draw visualisation
fname = '{}/timeseries_training_data.png'.format(
options.output_path)
viz.plot_delay(times, target, y_pred,
'Delay for station {}'.format(stationName), fname)
fname = '{}/scatter_all_stations.png'.format(options.vis_path)
viz.scatter_predictions(times,
target,
y_pred,
savepath=options.vis_path,
filename='scatter_{}'.format(station))
#except KeyError:
# pass
# Mean delay over the whole dataset (both train and validation),
# used to calculate Brier Skill
if options.y_avg:
mean_delay = 3.375953418071136
else:
mean_delay = 6.011229358531166
if options.model == 'llasso':
print('X_test shape: {}'.format(X_test.shape))
y_pred, weights = model.predict(X_test,
test.loc[:, 'trainstation'].values)
else:
y_pred = model.predict(X_test)
if options.normalize:
y_pred = yscaler.inverse_transform(y_pred)
rmse = np.sqrt(mean_squared_error(y_test, y_pred))
mae = mean_absolute_error(y_test, y_pred)
r2 = r2_score(y_test, y_pred)
rmse_stat = math.sqrt(
mean_squared_error(y_test, np.full_like(y_test, mean_delay)))
skill = 1 - rmse / rmse_stat
rmses.append(rmse)
maes.append(mae)
r2s.append(r2)
skills.append(skill)
start_times.append(start.strftime('%Y-%m-%dT%H:%M:%S'))
end_times.append(end.strftime('%Y-%m-%dT%H:%M:%S'))
end_times_obj.append(end)
if options.model in ['rf', 'lr', 'ard', 'gp']:
logging.info('R2 score for training: {}'.format(
model.score(X_train, y_train)))
logging.info('RMSE: {}'.format(rmse))
logging.info('MAE: {}'.format(mae))
logging.info('R2 score: {}'.format(r2))
logging.info('Brier Skill Score score: {}'.format(skill))
start = end
end = start + timedelta(days=int(options.day_step),
hours=int(options.hour_step))
if end > endtime: end = endtime
# SAVE #####################################################################
io.save_scikit_model(model,
filename=options.save_file,
ext_filename=options.save_file)
if options.normalize:
fname = options.save_path + '/xscaler.pkl'
io.save_scikit_model(xscaler, filename=fname, ext_filename=fname)
fname = options.save_path + '/yscaler.pkl'
io.save_scikit_model(yscaler, filename=fname, ext_filename=fname)
if options.model == 'rf':
fname = options.output_path + '/rfc_feature_importance.png'
viz.rfc_feature_importance(model.feature_importances_,
fname,
feature_names=options.feature_params)
#io._upload_to_bucket(filename=fname, ext_filename=fname)
try:
fname = options.output_path + '/learning_over_time.png'
viz.plot_learning_over_time(end_times_obj,
rmses,
maes,
r2s,
filename=fname)
#io._upload_to_bucket(filename=fname, ext_filename=fname)
except Exception as e:
logging.error(e)
error_data = {
'start_times': start_times,
'end_times': end_times,
'rmse': rmses,
'mae': maes,
'r2': r2s,
'skill': skills
}
fname = '{}/training_time_validation_errors.csv'.format(
options.output_path)
io.write_csv(error_data, filename=fname, ext_filename=fname)
# FEATURE SELECTION ########################################################
if options.feature_selection:
logging.info('Doing feature selection...')
selector = SelectFromModel(model, prefit=True)
print(pd.DataFrame(data=X_complete))
X_selected = selector.transform(X_complete)
selected_columns = f_data.columns.values[selector.get_support()]
logging.info(
'Selected following parameters: {}'.format(selected_columns))
data_sel = meta_complete.join(
pd.DataFrame(data=y_complete, columns=options.label_params)).join(
pd.DataFrame(data=X_selected, columns=selected_columns))
print(pd.DataFrame(data=X_selected, columns=selected_columns))
print(data_sel)
def main():
"""
Get data from db and save it as csv
"""
bq = BQHandler()
io = IO(gs_bucket=options.gs_bucket)
viz = Viz(io=io)
predictor = Predictor(io, ModelLoader(io), options,
options.station_specific_classifier,
options.station_specific_regressor)
predictor.regressor_save_file = options.save_path + '/classifier.pkl'
predictor.classifier_save_file = options.save_path + '/regressor.pkl'
# Mean delay over the whole dataset (both train and validation),
# used to calculate Brier Skill
mean_delay = options.mean_delay
starttime, endtime = io.get_dates(options)
logging.info('Using dataset {} and time range {} - {}'.format(
options.feature_dataset, starttime.strftime('%Y-%m-%d'),
endtime.strftime('%Y-%m-%d')))
# Get params
all_param_names = list(
set(options.label_params + options.feature_params +
options.meta_params + options.classifier_feature_params +
options.regressor_feature_params))
# Param list is modified after retrieving data
classifier_feature_params = copy.deepcopy(
options.classifier_feature_params)
regressor_feature_params = copy.deepcopy(options.regressor_feature_params)
all_feature_params = list(
set(options.feature_params + options.meta_params +
options.classifier_feature_params +
options.regressor_feature_params))
aggs = io.get_aggs_from_param_names(all_feature_params)
# Init error dicts
avg_delay = {}
avg_pred_delay = {}
avg_proba = {}
station_count = 0
all_times = set()
station_rmse = {}
station_mae = {}
station_r2 = {}
station_skill = {}
# For aggregated binary classification metrics
time_list, target_list, y_pred_bin_list, y_pred_bin_proba_list = [], [], [], []
# If stations are given as argument use them, else use all stations
stationList = io.get_train_stations(options.stations_file)
all_data = None
if options.locations is not None:
stations = options.locations
else:
stations = stationList.keys()
# Go through stations
for station in stations:
stationName = '{} ({})'.format(stationList[station]['name'], station)
logging.info('Processing station {}'.format(stationName))
if hasattr(options, 'classifier_model_file'):
predictor.classifier_save_file = options.classifier_model_file.replace(
'{location}', station)
elif options.station_specific_classifier:
predictor.classifier_save_file = options.save_path + '/{}'.format(
station) + '/classifier.pkl'
if hasattr(options, 'regressor_model_file'):
predictor.regressor_save_file = options.regressor_model_file.replace(
'{location}', station)
elif options.station_specific_regressor:
predictor.regressor_save_file = options.save_path + '/{}'.format(
station) + '/regressor.pkl'
station_rmse[station] = {}
station_mae[station] = {}
station_r2[station] = {}
station_skill[station] = {}
# Read data and filter desired train types (ic and commuter)
table = 'features_testset'
if hasattr(options, 'test_table'):
table = options.test_table
data = bq.get_rows(starttime,
endtime,
loc_col='trainstation',
project=options.project,
dataset='trains_data',
table=table,
parameters=all_param_names,
only_winters=options.only_winters,
reason_code_table=options.reason_code_table,
reason_codes_exclude=options.reason_codes_exclude,
reason_codes_include=options.reason_codes_include,
locations=[station])
data = io.filter_train_type(labels_df=data,
train_types=['K', 'L'],
sum_types=True,
train_type_column='train_type',
location_column='trainstation',
time_column='time',
sum_columns=['train_count', 'delay'],
aggs=aggs)
if len(data) == 0:
continue
if options.y_avg_hours is not None:
data = io.calc_running_delay_avg(data, options.y_avg_hours)
if options.y_avg:
data = io.calc_delay_avg(data)
if options.month:
logging.info('Adding month to the dataset...')
data = data.assign(
month=lambda df: df.loc[:, 'time'].map(lambda x: x.month))
if 'month' not in options.feature_params:
options.feature_params.append('month')
if 'month' not in options.regressor_feature_params:
options.regressor_feature_params.append('month')
if 'month' not in options.classifier_feature_params:
options.classifier_feature_params.append('month')
data.sort_values(by=['time'], inplace=True)
logging.info('Processing {} rows...'.format(len(data)))
if all_data is None:
all_data = data
else:
all_data.append(data, ignore_index=True)
# Pick times for creating error time series
times = data.loc[:, 'time']
station_count += 1
# Run prediction
try:
#target, y_pred = predictor.pred(times, data)
y_pred, y_pred_bin, y_pred_bin_proba = predictor.pred(times, data)
# Drop first times which LSTM are not able to predict
#times = times[(len(data)-len(y_pred)):]
except (PredictionError, ModelError) as e:
logging.error(e)
continue
target = data.loc[:, options.label_params].reset_index(
drop=True).values.ravel()
if len(y_pred) < 1 or len(target) < 1:
continue
# Create timeseries of predicted and happended delay
i = 0
for t in times:
try:
if t not in avg_delay.keys():
avg_delay[t] = [target[i]]
avg_pred_delay[t] = [y_pred[i]]
if predictor.y_pred_bin_proba is not None:
avg_proba[t] = [predictor.y_pred_bin_proba[i, 1]]
else:
avg_delay[t].append(target[i])
avg_pred_delay[t].append(y_pred[i])
if predictor.y_pred_bin_proba is not None:
avg_proba[t].append(predictor.y_pred_bin_proba[i, 1])
except IndexError as e:
# LSTM don't have first time steps because it don't
# have necessary history
pass
i += 1
# For creating visualisation
all_times = all_times.union(set(times))
# If only average plots are asked, continue to next station
if options.only_avg == 1:
continue
# Calculate errors for given station, first for all periods and then for whole time range
if predictor.y_pred_bin is not None:
time_list += list(times)
#feature_list += list()
target_list += list(target)
y_pred_bin_list += list(predictor.y_pred_bin)
y_pred_bin_proba_list += list(predictor.y_pred_bin_proba)
splits = viz._split_to_parts(list(times), [
target, y_pred, predictor.y_pred_bin,
predictor.y_pred_bin_proba
], 2592000)
else:
splits = viz._split_to_parts(list(times), [target, y_pred],
2592000)
for i in range(0, len(splits)):
logging.info('Month {}:'.format(i + 1))
if predictor.y_pred_bin is not None:
times_, target_, y_pred_, y_pred_bin_, y_pred_bin_proba_ = splits[
i]
viz.classification_perf_metrics(y_pred_bin_proba_, y_pred_bin_,
target_, options, times_,
station)
else:
times_, target_, y_pred_ = splits[i]
rmse = math.sqrt(metrics.mean_squared_error(target_, y_pred_))
mae = metrics.mean_absolute_error(target_, y_pred_)
r2 = metrics.r2_score(target_, y_pred_)
rmse_stat = math.sqrt(
metrics.mean_squared_error(target_,
np.full_like(target_, mean_delay)))
skill = 1 - rmse / rmse_stat
# Put errors to timeseries
station_rmse[station][i] = rmse
station_mae[station][i] = mae
station_r2[station][i] = r2
station_skill[station][i] = skill
logging.info('RMSE of station {} month {}: {:.4f}'.format(
stationName, i + 1, rmse))
logging.info('MAE of station {} month {}: {:.4f}'.format(
stationName, i + 1, mae))
logging.info('R2 score of station {} month {}: {:.4f}'.format(
stationName, i + 1, r2))
logging.info('Skill (RMSE) of station {} month {}: {:.4f}'.format(
stationName, i + 1, skill))
mse = math.sqrt(metrics.mean_squared_error(target, y_pred))
mae = metrics.mean_absolute_error(target, y_pred)
r2 = metrics.r2_score(target, y_pred)
rmse_stat = math.sqrt(
metrics.mean_squared_error(target,
np.full_like(target, mean_delay)))
skill = 1 - rmse / rmse_stat
station_rmse[station]['all'] = rmse
station_mae[station]['all'] = mae
station_r2[station]['all'] = r2
station_skill[station]['all'] = skill
logging.info('All periods:')
logging.info('RMSE of station {} month {}: {:.4f}'.format(
stationName, i + 1, rmse))
logging.info('MAE of station {} month {}: {:.4f}'.format(
stationName, i + 1, mae))
logging.info('R2 score of station {} month {}: {:.4f}'.format(
stationName, i + 1, r2))
logging.info('Skill (RMSE) of station {} month {}: {:.4f}'.format(
stationName, i + 1, skill))
# Create csv and upload it to pucket
times_formatted = [t.strftime('%Y-%m-%dT%H:%M:%S') for t in times]
delay_data = {
'times': times_formatted,
'delay': target,
'predicted delay': y_pred
}
fname = '{}/delays_{}.csv'.format(options.vis_path, station)
io.write_csv(delay_data, fname, fname)
# Draw visualisation
if predictor.y_pred_bin_proba is not None:
fname = '{}/timeseries_proba_{}'.format(options.vis_path, station)
proba = predictor.y_pred_bin_proba[:, 1]
viz.plot_delay(times,
target,
None,
'Delay for station {}'.format(stationName),
fname,
all_proba=proba,
proba_mode='same',
color_threshold=options.class_limit)
#else:
fname = '{}/timeseries_regression_{}'.format(options.vis_path, station)
viz.plot_delay(times,
target,
y_pred,
'Delay for station {}'.format(stationName),
fname,
all_proba=None)
fname = '{}/scatter_all_stations.png'.format(options.vis_path)
viz.scatter_predictions(times,
target,
y_pred,
savepath=options.vis_path,
filename='scatter_{}'.format(station))
# Save all station related results to csv and upload them to bucket
fname = '{}/station_rmse.csv'.format(options.vis_path)
io.dict_to_csv(station_rmse, fname, fname)
fname = '{}/station_mae.csv'.format(options.vis_path)
io.dict_to_csv(station_mae, fname, fname)
fname = '{}/station_r2.csv'.format(options.vis_path)
io.dict_to_csv(station_r2, fname, fname)
fname = '{}/station_skill_rmse.csv'.format(options.vis_path)
io.dict_to_csv(station_skill, fname, fname)
# Create timeseries of avg actual delay and predicted delay
all_times = sorted(list(all_times))
for t, l in avg_delay.items():
avg_delay[t] = sum(l) / len(l)
for t, l in avg_pred_delay.items():
avg_pred_delay[t] = sum(l) / len(l)
for t, l in avg_proba.items():
avg_proba[t] = sum(l) / len(l)
avg_delay = list(
OrderedDict(sorted(avg_delay.items(), key=lambda t: t[0])).values())
avg_pred_delay = list(
OrderedDict(sorted(avg_pred_delay.items(),
key=lambda t: t[0])).values())
avg_proba = list(
OrderedDict(sorted(avg_proba.items(), key=lambda t: t[0])).values())
# Calculate average over all times and stations, first for all months separately, then for whole time range
splits = viz._split_to_parts(list(times), [avg_delay, avg_pred_delay],
2592000)
for i in range(0, len(splits)):
times_, avg_delay_, avg_pred_delay_ = splits[i]
try:
rmse = math.sqrt(
metrics.mean_squared_error(avg_delay_, avg_pred_delay_))
mae = metrics.mean_absolute_error(avg_delay_, avg_pred_delay_)
r2 = metrics.r2_score(avg_delay_, avg_pred_delay_)
rmse_stat = math.sqrt(
metrics.mean_squared_error(
avg_delay_, np.full_like(avg_delay_, mean_delay)))
skill = 1 - rmse / rmse_stat
except ValueError:
logging.warning('Zero samples in some class')
continue
logging.info('Month: {}'.format(i + 1))
logging.info(
'RMSE of average delay over all stations: {:.4f}'.format(rmse))
logging.info(
'MAE of average delay over all stations: {:.4f}'.format(mae))
logging.info(
'R2 score of average delay over all stations: {:.4f}'.format(r2))
logging.info(
'Skill score (RMSE) of average delay over all stations: {:.4f}'.
format(skill))
# Write average data into file
avg_errors = {
'rmse': rmse,
'mae': mae,
'r2': r2,
'skill': skill,
'nro_of_samples': len(avg_delay)
}
fname = '{}/avg_erros_{}.csv'.format(options.vis_path, i)
io.dict_to_csv(avg_errors, fname, fname)
rmse = math.sqrt(metrics.mean_squared_error(avg_delay, avg_pred_delay))
#rmse_mean = np.mean(list(station_rmse.values()))
mae = metrics.mean_absolute_error(avg_delay, avg_pred_delay)
#mae_mean = np.mean(list(station_mae.values()))
r2 = metrics.r2_score(avg_delay, avg_pred_delay)
rmse_stat = math.sqrt(
metrics.mean_squared_error(avg_delay,
np.full_like(avg_delay, mean_delay)))
skill = 1 - rmse / rmse_stat
#skill_mean = 1 - rmse_mean/rmse_stat
logging.info('All periods:')
logging.info(
'RMSE of average delay over all stations: {:.4f}'.format(rmse))
#logging.info('Average RMSE of all station RMSEs: {:.4f}'.format(rmse_mean))
logging.info('MAE of average delay over all stations: {:.4f}'.format(mae))
#logging.info('Average MAE of all station MAEs: {:.4f}'.format(mae_mean))
logging.info(
'R2 score of average delay over all stations: {:.4f}'.format(r2))
logging.info(
'Skill score (RMSE) of average delay over all stations: {:.4f}'.format(
skill))
#logging.info('Skill score (avg RMSE) of all stations: {:.4f}'.format(skill_mean))
# Write average data into file
avg_errors = {
'rmse': rmse,
'mae': mae,
'r2': r2,
#'rmse_mean': rmse_mean,
#'mae_mean': mae_mean,
'skill': skill,
#'skill_mean': skill_mean,
'nro_of_samples': len(avg_delay)
}
fname = '{}/avg_erros.csv'.format(options.vis_path)
io.dict_to_csv(avg_errors, fname, fname)
# Create timeseries of average delay and predicted delays over all stations
all_times_formatted = [t.strftime('%Y-%m-%dT%H:%M:%S') for t in all_times]
delay_data = {
'times': all_times_formatted,
'delay': avg_delay,
'predicted delay': avg_pred_delay
}
# write csv
fname = '{}/avg_delays_all_stations.csv'.format(options.vis_path)
io.write_csv(delay_data, fname, fname)
# visualise
if not avg_proba:
proba = None
else:
proba = avg_proba
fname = '{}/timeseries_avg_all_stations.png'.format(options.vis_path)
if predictor.y_pred_bin is not None:
viz.plot_delay(all_times,
avg_delay,
None,
'Average delay for all station',
fname,
all_proba=proba,
proba_mode='same',
color_threshold=options.class_limit)
else:
viz.plot_delay(all_times, avg_delay, avg_pred_delay,
'Average delay for all station', fname)
fname = '{}/scatter_all_stations.png'.format(options.vis_path)
viz.scatter_predictions(all_times,
avg_delay,
avg_pred_delay,
savepath=options.vis_path,
filename='scatter_all_stations')
# Binary classification metrics
if predictor.y_pred_bin is not None:
all_data.sort_values(by=['time'], inplace=True)
times = all_data.loc[:, 'time'].values
try:
target = all_data.loc[:, options.label_params].reset_index(
drop=True).values.ravel()
y_pred, y_pred_bin, y_pred_bin_proba = predictor.pred(
times, all_data)
# Drop first times which LSTM are not able to predict
times = times[(len(all_data) - len(y_pred)):]
splits = viz._split_to_parts(list(times), [
target, y_pred, predictor.y_pred_bin,
predictor.y_pred_bin_proba
], 2592000)
for i in range(0, len(splits)):
#times_, target_, y_pred_bin_, y_pred_bin_proba_ = splits[i]
times_, target_, y_pred_, y_pred_bin_, y_pred_bin_proba_ = splits[
i]
viz.classification_perf_metrics(y_pred_bin_proba_, y_pred_bin_,
target_, options, times_,
'all')
except (PredictionError, ModelError) as e:
logging.error(e)
pass
def main():
"""
Get data from db and save it as csv
"""
# Helpers
bq = bqhandler.BQHandler()
io = IO(gs_bucket=options.gs_bucket)
viz = Viz(io)
# Configuration
starttime, endtime = io.get_dates(options)
logging.info('Using dataset {} and time range {} - {}'.format(
options.feature_dataset, starttime.strftime('%Y-%m-%d'),
endtime.strftime('%Y-%m-%d')))
all_param_names = options.label_params + options.feature_params + options.meta_params
aggs = io.get_aggs_from_param_names(options.feature_params)
# Initialise classifier
if hasattr(options, 'classifier_file'):
classifier = io.load_scikit_model(options.classifier_file)
else:
if options.classifier == 'svc':
params = {
'kernel': options.kernel,
'gamma': options.gamma,
'C': options.penalty,
'probability': options.probability
}
classifier = SVC(**params)
elif options.classifier == 'rfc':
classifier = RandomForestClassifier(n_jobs=-1)
else:
raise (
'Model not specificied or wrong. Add "classifier: bgm" to config file.'
)
# Initialise regression model
if options.regression == 'rfr':
model = RandomForestRegressor(
n_estimators=options.n_estimators,
n_jobs=-1,
min_samples_leaf=options.min_samples_leaf,
min_samples_split=options.min_samples_split,
max_features=options.max_features,
max_depth=options.max_depth,
bootstrap=options.bootstrap)
regressor = _trans.Regressor(model=model)
else:
raise (
'Model not specificied or wrong. Add "classifier: bgm" to config file.'
)
# Initialise transformer
transformer = _trans.Selector(classifier=classifier)
# Initialise pipeline
pipe = Pipeline([('selector', transformer), ('regression', regressor)])
sum_columns = ['delay']
if options.reason_code_table is not None:
sum_columns = ['count']
# Pick only selected month
where = {}
if options.pick_month is not None:
where = {'EXTRACT(MONTH from time)': options.pick_month}
logging.info('Reading data...')
bq.set_params(loc_col='trainstation',
project=options.project,
dataset=options.feature_dataset,
table=options.feature_table,
parameters=all_param_names,
reason_code_table=options.reason_code_table,
where=where)
data = bq.get_rows(starttime, endtime)
data = io.filter_train_type(labels_df=data,
train_types=options.train_types,
sum_types=True,
train_type_column='train_type',
location_column='trainstation',
time_column='time',
sum_columns=sum_columns,
aggs=aggs)
data.sort_values(by=['time', 'trainstation'], inplace=True)
logging.info('Processing {} rows...'.format(len(data)))
# Filter only timesteps with large distribution in the whole network
if options.filter_delay_limit is not None:
data = io.filter_delay_with_limit(data, options.filter_delay_limit)
# Binary class
logging.info('Adding binary class to the dataset with limit {}...'.format(
options.delay_limit))
data['class'] = data['delay'].map(lambda x: 1
if x > options.delay_limit else -1)
# Separate train and validation sets
data_train, data_test = train_test_split(data, test_size=0.3)
# Balance
if options.balance:
logging.info('Balancing training data...')
count = data_train.groupby('class').size().min()
# SVC can't handle more than 50 000 samples
if options.classifier == 'svc': count = min(count, 50000)
data_train = pd.concat([
data_train.loc[data_train['class'] == -1].sample(n=count),
data_train.loc[data_train['class'] == 1].sample(n=count)
])
logging.info('Train data:')
io.log_class_dist(data_train.loc[:, 'class'].values, labels=[-1, 1])
logging.info('Test data:')
io.log_class_dist(data_test.loc[:, 'class'].values, labels=[-1, 1])
# Adding month
if options.month:
logging.info('Adding month to the datasets...')
data_train['month'] = data_train.loc[:, 'time'].map(lambda x: x.month)
data_test['month'] = data_test.loc[:, 'time'].map(lambda x: x.month)
options.feature_params.append('month')
data_train.set_index('time', inplace=True)
y_train = data_train.loc[:, ['delay', 'class']].astype(np.int32).values
y_test = data_test.loc[:, ['delay', 'class']].astype(np.int32).values
X_train = data_train.loc[:,
options.feature_params].astype(np.float32).values
X_test = data_test.loc[:, options.feature_params].astype(np.float32).values
# X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3)
# io.log_class_dist(y_train[:,1], [-1,1])
# If asked, save used train and test splits into big query
if options.save_data:
tname = options.model + '_' + options.feature_dataset + '_' + options.config_name + '_train'
bq.nparray_to_table([X_train, y_train],
[options.feature_params, ['delay', 'class']],
options.project, options.feature_dataset, tname)
tname = options.model + '_' + options.feature_dataset + '_' + options.config_name + '_test'
bq.nparray_to_table([X_test, y_test],
[options.feature_params, ['delay', 'class']],
options.project, options.feature_dataset, tname)
if options.normalize:
logging.info('Normalizing data...')
if hasattr(options, 'xscaler_file'):
scaler = io.load_scikit_model(options.xscaler_file)
X_train = scaler.transform(X_train)
X_test = scaler.transform(X_test)
else:
scaler = StandardScaler()
X_train = scaler.fit_transform(X_train)
X_test = scaler.transform(X_test)
if options.cv:
logging.info('Doing random search for hyper parameters...')
raise ("No param_grid set for given model ({})".format(
options.regression))
random_search = RandomizedSearchCV(model,
param_distributions=param_grid,
n_iter=int(options.n_iter_search),
n_jobs=-1)
random_search.fit(X_train, y_train)
logging.info("RandomizedSearchCV done.")
fname = options.output_path + '/random_search_cv_results.txt'
report_cv_results(random_search.cv_results_, fname)
io._upload_to_bucket(filename=fname, ext_filename=fname)
sys.exit()
else:
logging.info('Training...')
transformer.set_y(y_train[:, 0])
regressor.set_classifier(transformer)
pipe.fit(X_train, y_train[:, 1])
# Metrics
y_pred_proba = pipe.steps[0][1].predict_proba(X_test)
y_pred = pipe.steps[0][1].predict(X_test, type='int')
io.save_scikit_model(pipe,
filename=options.save_file,
ext_filename=options.save_file)
# Classification performance
y_test_class = y_test[:, 1]
acc = accuracy_score(y_test_class, y_pred)
precision = precision_score(y_test_class, y_pred, average='micro')
recall = recall_score(y_test_class, y_pred, average='micro')
f1 = f1_score(y_test_class, y_pred, average='micro')
logging.info('Classification accuracy: {}'.format(acc))
logging.info('Classification precision: {}'.format(precision))
logging.info('Classification recall: {}'.format(recall))
logging.info('Classification F1 score: {}'.format(f1))
io.log_class_dist(y_pred, [-1, 1])
# Confusion matrices
fname = '{}/confusion_matrix_validation.png'.format(options.output_path)
viz.plot_confusion_matrix(y_test_class,
y_pred,
np.arange(2),
filename=fname)
fname = '{}/confusion_matrix_validation_normalised.png'.format(
options.output_path)
viz.plot_confusion_matrix(y_test_class,
y_pred,
np.arange(2),
True,
filename=fname)
# Precision-recall curve
fname = '{}/precision-recall-curve.png'.format(options.output_path)
viz.prec_rec_curve(y_test_class, y_pred_proba, filename=fname)
# ROC
fname = '{}/roc.png'.format(options.output_path)
viz.plot_binary_roc(y_test_class, y_pred_proba, filename=fname)
if options.normalize:
fname = options.save_path + '/xscaler.pkl'
io.save_scikit_model(scaler, filename=fname, ext_filename=fname)
if options.regression == 'rfr':
fname = options.output_path + '/rfc_feature_importance.png'
viz.rfc_feature_importance(pipe.steps[1][1].get_feature_importances(),
fname,
feature_names=options.feature_params)
# Regression performance
#y_pred = pipe.steps[1][1].predict(X_test)
y_test_reg = y_test[:, 0]
pipe.set_params(selector__full=True)
y_pred = pipe.predict(X_test, full=True)
rmse = np.sqrt(mean_squared_error(y_test_reg, y_pred))
mae = mean_absolute_error(y_test_reg, y_pred)
r2 = r2_score(y_test_reg, y_pred)
logging.info('Regression RMSE: {}'.format(rmse))
logging.info('Regression MAE: {}'.format(mae))
logging.info('Regression R2 score: {}'.format(r2))
error_data = {
'acc': [acc],
'precision': [precision],
'recall': [recall],
'f1': [f1],
'rmse': [rmse],
'mae': [mae],
'r2': [r2]
}
fname = '{}/training_time_classification_validation_errors.csv'.format(
options.output_path)
io.write_csv(error_data, filename=fname, ext_filename=fname)
############################################################################
# EVALUATE
############################################################################
if options.evaluate:
logging.info('Loading test data...')
test_data = bq.get_rows(dt.datetime.strptime('2010-01-01', "%Y-%m-%d"),
dt.datetime.strptime('2019-01-01', "%Y-%m-%d"),
loc_col='trainstation',
project=options.project,
dataset=options.feature_dataset,
table=options.test_table,
parameters=all_param_names)
test_data = io.filter_train_type(labels_df=test_data,
train_types=['K', 'L'],
sum_types=True,
train_type_column='train_type',
location_column='trainstation',
time_column='time',
sum_columns=['delay'],
aggs=aggs)
# Sorting is actually not necessary. It's been useful for debugging.
test_data.sort_values(by=['time', 'trainstation'], inplace=True)
test_data.set_index('time', inplace=True)
logging.info('Test data contain {} rows...'.format(len(test_data)))
logging.info(
'Adding binary class to the test dataset with limit {}...'.format(
options.delay_limit))
#logging.info('Adding binary class to the dataset with limit {}...'.format(limit))
#data['class'] = data['count'].map(lambda x: 1 if x > options.delay_count_limit else -1)
test_data['class'] = test_data['delay'].map(
lambda x: 1 if x > options.delay_limit else -1)
io.log_class_dist(test_data.loc[:, 'class'].values, labels=[-1, 1])
if options.month:
logging.info('Adding month to the test dataset...')
test_data['month'] = test_data.index.map(lambda x: x.month)
times = [('2014-01-01', '2014-02-01'), ('2016-06-01', '2016-07-01'),
('2017-02-01', '2017-03-01'), ('2011-02-01', '2017-03-01')]
for start, end in times:
try:
y_pred_proba, y_pred, y = predict_timerange(
test_data, options.feature_params, pipe.steps[0][1],
scaler, start, end)
perf_metrics(y_pred_proba, y_pred, y, start, end, viz, io)
except EmptyDataError:
logging.info('No data for {} - {}'.format(start, end))
def main():
"""
Get data from db and save it as csv
"""
bq = BQHandler()
io = IO(gs_bucket=options.gs_bucket)
viz = Viz(io)
starttime, endtime = io.get_dates(options)
logging.info('Using dataset {} and time range {} - {}'.format(
options.feature_dataset, starttime.strftime('%Y-%m-%d'),
endtime.strftime('%Y-%m-%d')))
all_param_names = options.label_params + options.feature_params + options.meta_params
aggs = io.get_aggs_from_param_names(options.feature_params)
if options.model == 'bgm':
model = BayesianGaussianMixture(
weight_concentration_prior_type="dirichlet_process",
n_components=options.n_components)
elif options.model == 'gaussiannb':
model = GaussianNB()
elif options.model == 'rfc':
model = RandomForestClassifier(n_jobs=-1)
elif options.model == 'svc':
params = {'kernel': 'rbf', 'gamma': 0.5, 'C': 1, 'probability': True}
model = SVC(**params)
else:
raise (
'Model not specificied or wrong. Add for example "model: bgm" to config file.'
)
if options.pca:
ipca = IncrementalPCA(n_components=options.pca_components,
whiten=options.whiten,
copy=False)
sum_columns = ['delay']
if options.reason_code_table is not None:
sum_columns = ['count']
logging.info('Reading data...')
data = bq.get_rows(starttime,
endtime,
loc_col='trainstation',
project=options.project,
dataset=options.feature_dataset,
table=options.feature_table,
parameters=all_param_names,
reason_code_table=options.reason_code_table,
only_winters=options.only_winters)
data = io.filter_train_type(labels_df=data,
train_types=options.train_types,
sum_types=True,
train_type_column='train_type',
location_column='trainstation',
time_column='time',
sum_columns=sum_columns,
aggs=aggs)
# Sorting is actually not necessary. It's been useful for debugging.
data.sort_values(by=['time', 'trainstation'], inplace=True)
data.set_index('time', inplace=True)
logging.info('Data contain {} rows...'.format(len(data)))
logging.info('Adding binary class to the dataset with limit {}...'.format(
options.delay_limit))
#logging.info('Adding binary class to the dataset with limit {}...'.format(limit))
#data['class'] = data['count'].map(lambda x: 1 if x > options.delay_count_limit else -1)
data['class'] = data['delay'].map(lambda x: 1
if x > options.delay_limit else -1)
io.log_class_dist(data.loc[:, 'class'].values, labels=[-1, 1])
if options.balance:
logging.info('Balancing dataset...')
count = data.groupby('class').size().min()
data = pd.concat([
data.loc[data['class'] == -1].sample(n=count),
data.loc[data['class'] == 1].sample(n=count)
])
io.log_class_dist(data.loc[:, 'class'].values, labels=[-1, 1])
if options.month:
logging.info('Adding month to the dataset...')
data['month'] = data.index.map(lambda x: x.month)
options.feature_params.append('month')
target = data.loc[:, 'class'].astype(np.int32).values.ravel()
features = data.loc[:, options.feature_params].astype(np.float32).values
X_train, X_test, y_train, y_test = train_test_split(features,
target,
test_size=0.3)
if options.normalize:
logging.info('Normalizing data...')
scaler = StandardScaler()
X_train = scaler.fit_transform(X_train)
X_test = scaler.transform(X_test)
logging.debug('Features shape after pre-processing: {}'.format(
X_train.shape))
if options.cv:
logging.info('Doing random search for hyper parameters...')
if options.model == 'bgm':
param_grid = {
"n_components": [1, 2, 4, 8, 16],
"covariance_type": ['full', 'tied', 'diag', 'spherical'],
"init_params": ['kmeans', 'random']
}
elif options.model == 'rfc':
raise ("Not implemented. Get back to work!")
elif options.model == 'svc':
features_compinations = [
[
'lat', 'lon', 'pressure', 'max_temperature',
'min_temperature', 'mean_temperature', 'mean_dewpoint',
'mean_humidity', 'mean_winddirection', 'mean_windspeedms',
'max_windgust', 'max_precipitation1h', 'max_snowdepth',
'max_n', 'min_vis', 'min_clhb', 'max_precipitation3h'
],
[
'pressure', 'max_temperature', 'min_temperature',
'mean_temperature', 'mean_dewpoint', 'mean_humidity',
'mean_winddirection', 'mean_windspeedms', 'max_windgust',
'max_precipitation1h', 'max_snowdepth', 'max_n', 'min_vis',
'min_clhb', 'max_precipitation3h'
],
[
'pressure', 'min_temperature', 'mean_dewpoint',
'mean_winddirection', 'mean_windspeedms', 'max_windgust',
'max_precipitation1h', 'max_snowdepth', 'max_n', 'min_vis',
'min_clhb', 'max_precipitation3h'
],
[
'pressure', 'min_temperature', 'mean_dewpoint',
'mean_winddirection', 'mean_windspeedms', 'max_snowdepth',
'max_n', 'min_vis', 'min_clhb', 'max_precipitation3h'
],
[
'pressure', 'min_temperature', 'mean_dewpoint',
'mean_winddirection', 'mean_windspeedms', 'max_snowdepth',
'max_n', 'min_vis', 'min_clhb', 'max_precipitation1h'
],
[
'pressure', 'min_temperature', 'mean_dewpoint',
'mean_winddirection', 'mean_windspeedms', 'max_snowdepth',
'min_vis', 'max_precipitation1h'
],
[
'pressure', 'min_temperature', 'mean_winddirection',
'mean_windspeedms', 'max_snowdepth', 'max_precipitation1h'
]
]
param_grid = {
"C": [0.0001, 0.001, 0.01, 0.1, 1],
"kernel": ['rbf', 'poly'],
"degree": [2, 3],
"gamma": [0.5],
"coef0": [0.1],
"probability": [True],
"features": features_compinations
}
from lib.svc import SVCF
model = SVCF(all_features=options.feature_params)
else:
raise ("No param_grid set for given model ({})".format(
options.model))
print(model.get_params().keys())
ftwo_scorer = make_scorer(fbeta_score, beta=2)
scoring = {
'accuracy': 'accuracy',
'precision': 'precision',
'recall': 'recall',
'f1': 'f1',
'f2': ftwo_scorer
}
random_search = RandomizedSearchCV(model,
param_distributions=param_grid,
n_iter=int(options.n_iter_search),
verbose=1,
scoring=scoring,
refit='recall',
n_jobs=-1)
random_search.fit(X_train, y_train)
logging.info("RandomizedSearchCV done.")
scores = ['accuracy', 'precision', 'recall', 'f1', 'f2']
fname = options.output_path + '/random_search_cv_results.txt'
io.report_cv_results(random_search.cv_results_,
scores=scores,
filename=fname,
ext_filename=fname)
model = random_search.best_estimator_
io.save_scikit_model(model,
filename=options.save_file,
ext_filename=options.save_file)
if options.normalize:
fname = options.save_path + '/xscaler.pkl'
io.save_scikit_model(scaler, filename=fname, ext_filename=fname)
else:
logging.info('Training...')
model.fit(X_train, y_train)
# Save model and xscaler (no reason to save xscaler before the model has fitted as well)
io.save_scikit_model(model,
filename=options.save_file,
ext_filename=options.save_file)
if options.normalize:
fname = options.save_path + '/xscaler.pkl'
io.save_scikit_model(scaler, filename=fname, ext_filename=fname)
# Metrics
y_pred_proba = model.predict_proba(X_test)
y_pred = np.argmax(y_pred_proba, axis=1)
# We want [-1,1] classes as y values are
y_pred[y_pred == 0] = -1
acc = accuracy_score(y_test, y_pred)
precision = precision_score(y_test, y_pred, average='binary')
recall = recall_score(y_test, y_pred, average='binary')
f1 = f1_score(y_test, y_pred, average='binary')
logging.info('Accuracy: {}'.format(acc))
logging.info('Precision: {}'.format(precision))
logging.info('Recall: {}'.format(recall))
logging.info('F1 score: {}'.format(f1))
io.log_class_dist(y_pred, labels=[-1, 1])
error_data = {
'acc': [acc],
'precision': [precision],
'recall': [recall],
'f1': [f1]
}
fname = '{}/training_time_validation_errors.csv'.format(
options.output_path)
io.write_csv(error_data, filename=fname, ext_filename=fname)
# Confusion matrices
fname = '{}/confusion_matrix_validation.png'.format(options.output_path)
viz.plot_confusion_matrix(y_test, y_pred, np.arange(2), filename=fname)
fname = '{}/confusion_matrix_validation_normalised.png'.format(
options.output_path)
viz.plot_confusion_matrix(y_test,
y_pred,
np.arange(2),
True,
filename=fname)
# Precision-recall curve
fname = '{}/precision-recall-curve.png'.format(options.output_path)
viz.prec_rec_curve(y_test, y_pred_proba, filename=fname)
# ROC
fname = '{}/roc.png'.format(options.output_path)
viz.plot_binary_roc(y_test, y_pred_proba, filename=fname)
############################################################################
# EVALUATE
############################################################################
if options.evaluate:
logging.info('Loading test data...')
test_data = bq.get_rows(dt.datetime.strptime('2010-01-01', "%Y-%m-%d"),
dt.datetime.strptime('2019-01-01', "%Y-%m-%d"),
loc_col='trainstation',
project=options.project,
dataset=options.feature_dataset,
table=options.test_table,
parameters=all_param_names)
test_data = io.filter_train_type(labels_df=test_data,
train_types=['K', 'L'],
sum_types=True,
train_type_column='train_type',
location_column='trainstation',
time_column='time',
sum_columns=['delay'],
aggs=aggs)
# Sorting is actually not necessary. It's been useful for debugging.
test_data.sort_values(by=['time', 'trainstation'], inplace=True)
test_data.set_index('time', inplace=True)
logging.info('Test data contain {} rows...'.format(len(test_data)))
logging.info(
'Adding binary class to the test dataset with limit {}...'.format(
options.delay_limit))
#logging.info('Adding binary class to the dataset with limit {}...'.format(limit))
#data['class'] = data['count'].map(lambda x: 1 if x > options.delay_count_limit else -1)
test_data['class'] = test_data['delay'].map(
lambda x: 1 if x > options.delay_limit else -1)
io.log_class_dist(test_data.loc[:, 'class'].values, labels=[-1, 1])
if options.month:
logging.info('Adding month to the test dataset...')
test_data['month'] = test_data.index.map(lambda x: x.month)
times = [('2011-02-01', '2011-03-01'), ('2016-06-01', '2016-07-01'),
('2017-02-01', '2017-03-01'), ('2011-02-01', '2017-03-01')]
for start, end in times:
try:
y_pred_proba, y_pred, y = predict_timerange(
test_data, options.feature_params, model, scaler, start,
end)
perf_metrics(y_pred_proba, y_pred, y, start, end, viz, io)
except EmptyDataError:
logging.info('No data for {} - {}'.format(start, end))
