def save_predictions(model, tag, partition):
r"""Save the predictions to disk.
Parameters
----------
model : alphapy.Model
The model object to save.
tag : str
A unique identifier for the output files, e.g., a date stamp.
partition : alphapy.Partition
Reference to the dataset.
Returns
-------
preds : numpy array
The prediction vector.
probas : numpy array
The probability vector.
"""
# Extract model parameters.
directory = model.specs['directory']
extension = model.specs['extension']
model_type = model.specs['model_type']
separator = model.specs['separator']
# Get date stamp to record file creation
timestamp = get_datestamp()
# Specify input and output directories
input_dir = SSEP.join([directory, 'input'])
output_dir = SSEP.join([directory, 'output'])
# Read the prediction frame
file_spec = ''.join([datasets[partition], '*'])
file_name = most_recent_file(input_dir, file_spec)
file_name = file_name.split(SSEP)[-1].split(PSEP)[0]
pf = read_frame(input_dir, file_name, extension, separator)
# Cull records before the prediction date
try:
predict_date = model.specs['predict_date']
found_pdate = True
except:
found_pdate = False
if found_pdate:
pd_indices = pf[pf.date >= predict_date].index.tolist()
pf = pf.iloc[pd_indices]
else:
pd_indices = pf.index.tolist()
# Save predictions for all projects
logger.info("Saving Predictions")
output_file = USEP.join(['predictions', timestamp])
preds = model.preds[(tag, partition)].squeeze()
if found_pdate:
preds = np.take(preds, pd_indices)
pred_series = pd.Series(preds, index=pd_indices)
df_pred = pd.DataFrame(pred_series, columns=['prediction'])
write_frame(df_pred, output_dir, output_file, extension, separator)
# Save probabilities for classification projects
probas = None
if model_type == ModelType.classification:
logger.info("Saving Probabilities")
output_file = USEP.join(['probabilities', timestamp])
probas = model.probas[(tag, partition)].squeeze()
if found_pdate:
probas = np.take(probas, pd_indices)
prob_series = pd.Series(probas, index=pd_indices)
df_prob = pd.DataFrame(prob_series, columns=['probability'])
write_frame(df_prob, output_dir, output_file, extension, separator)
# Save ranked predictions
logger.info("Saving Ranked Predictions")
pf['prediction'] = pred_series
if model_type == ModelType.classification:
pf['probability'] = prob_series
pf.sort_values('probability', ascending=False, inplace=True)
else:
pf.sort_values('prediction', ascending=False, inplace=True)
output_file = USEP.join(['rankings', timestamp])
write_frame(pf, output_dir, output_file, extension, separator)
# Return predictions and any probabilities
return preds, probas
def training_pipeline(model):
r"""AlphaPy Training Pipeline
Parameters
----------
model : alphapy.Model
The model object for controlling the pipeline.
Returns
-------
model : alphapy.Model
The final results are stored in the model object.
Raises
------
KeyError
If the number of columns of the train and test data do not match,
then this exception is raised.
"""
logger.info("Training Pipeline")
# Unpack the model specifications
calibration = model.specs['calibration']
directory = model.specs['directory']
drop = model.specs['drop']
extension = model.specs['extension']
feature_selection = model.specs['feature_selection']
grid_search = model.specs['grid_search']
model_type = model.specs['model_type']
predict_mode = model.specs['predict_mode']
rfe = model.specs['rfe']
sampling = model.specs['sampling']
scorer = model.specs['scorer']
separator = model.specs['separator']
target = model.specs['target']
# Get train and test data
X_train, y_train = get_data(model, Partition.train)
X_test, y_test = get_data(model, Partition.test)
# Determine if there are any test labels
if y_test.any():
logger.info("Test Labels Found")
model.test_labels = True
model = save_features(model, X_train, X_test, y_train, y_test)
# Log feature statistics
logger.info("Original Feature Statistics")
logger.info("Number of Training Rows : %d", X_train.shape[0])
logger.info("Number of Training Columns : %d", X_train.shape[1])
if model_type == ModelType.classification:
uv, uc = np.unique(y_train, return_counts=True)
logger.info("Unique Training Values for %s : %s", target, uv)
logger.info("Unique Training Counts for %s : %s", target, uc)
logger.info("Number of Testing Rows : %d", X_test.shape[0])
logger.info("Number of Testing Columns : %d", X_test.shape[1])
if model_type == ModelType.classification and model.test_labels:
uv, uc = np.unique(y_test, return_counts=True)
logger.info("Unique Testing Values for %s : %s", target, uv)
logger.info("Unique Testing Counts for %s : %s", target, uc)
# Merge training and test data
if X_train.shape[1] == X_test.shape[1]:
split_point = X_train.shape[0]
X = pd.concat([X_train, X_test])
else:
raise IndexError(
"The number of training and test columns [%d, %d] must match." %
(X_train.shape[1], X_test.shape[1]))
# Apply treatments to the feature matrix
all_features = apply_treatments(model, X)
# Drop features
all_features = drop_features(all_features, drop)
# Save the train and test files with extracted and dropped features
datestamp = get_datestamp()
data_dir = SSEP.join([directory, 'input'])
df_train = all_features.iloc[:split_point, :]
df_train = pd.concat(
[df_train, pd.DataFrame(y_train, columns=[target])], axis=1)
output_file = USEP.join([model.train_file, datestamp])
write_frame(df_train, data_dir, output_file, extension, separator)
df_test = all_features.iloc[split_point:, :]
if y_test.any():
df_test = pd.concat(
[df_test, pd.DataFrame(y_test, columns=[target])], axis=1)
output_file = USEP.join([model.test_file, datestamp])
write_frame(df_test, data_dir, output_file, extension, separator)
# Create crosstabs for any categorical features
if model_type == ModelType.classification:
create_crosstabs(model)
# Create initial features
all_features = create_features(model, all_features)
X_train, X_test = np.array_split(all_features, [split_point])
model = save_features(model, X_train, X_test)
# Generate interactions
all_features = create_interactions(model, all_features)
X_train, X_test = np.array_split(all_features, [split_point])
model = save_features(model, X_train, X_test)
# Remove low-variance features
all_features = remove_lv_features(model, all_features)
X_train, X_test = np.array_split(all_features, [split_point])
model = save_features(model, X_train, X_test)
# Shuffle the data [if specified]
model = shuffle_data(model)
# Oversampling or Undersampling [if specified]
if model_type == ModelType.classification:
if sampling:
model = sample_data(model)
else:
logger.info("Skipping Sampling")
# Get sample weights (classification only)
model = get_class_weights(model)
# Perform feature selection, independent of algorithm
if feature_selection:
model = select_features(model)
# Get the available classifiers and regressors
logger.info("Getting All Estimators")
estimators = get_estimators(model)
# Get the available scorers
if scorer not in scorers:
raise KeyError("Scorer function %s not found" % scorer)
# Model Selection
logger.info("Selecting Models")
for algo in model.algolist:
logger.info("Algorithm: %s", algo)
# select estimator
try:
estimator = estimators[algo]
scoring = estimator.scoring
est = estimator.estimator
except KeyError:
logger.info("Algorithm %s not found", algo)
# initial fit
model = first_fit(model, algo, est)
# recursive feature elimination
if rfe:
if scoring:
model = rfecv_search(model, algo)
elif hasattr(est, "coef_"):
model = rfe_search(model, algo)
else:
logger.info("No RFE Available for %s", algo)
# grid search
if grid_search:
model = hyper_grid_search(model, estimator)
# predictions
model = make_predictions(model, algo, calibration)
# Create a blended estimator
if len(model.algolist) > 1:
model = predict_blend(model)
# Generate metrics
model = generate_metrics(model, Partition.train)
model = generate_metrics(model, Partition.test)
# Store the best estimator
model = predict_best(model)
# Generate plots
generate_plots(model, Partition.train)
if model.test_labels:
generate_plots(model, Partition.test)
# Save best features and predictions
save_model(model, 'BEST', Partition.test)
# Return the model
return model