def test_training_with_warm_start(self):
"""
Training with a user provided model for warm start.
"""
# Get trainer object, but only train 1 L-BFGS step.
binary_lr_trainer = BinaryLogisticRegressionTrainer(lambda_l2=0.0,
max_iter=1)
coefficients_warm_start = binary_lr_trainer.fit(
X=self.x_train,
y=self.y_train,
weights=None,
offsets=None,
theta_initial=self.custom_weights)[0]
# Warm start.
# The trained model should be close to initial value
# since the solution should have already converged.
self.assertAllClose(coefficients_warm_start,
self.custom_weights,
rtol=_TOLERANCE,
atol=_TOLERANCE,
msg='models mismatch')
coefficients_code_start = binary_lr_trainer.fit(X=self.x_train,
y=self.y_train,
weights=None,
offsets=None,
theta_initial=None)[0]
# Code start
# The trained model should be far from initial value since we only train 1 step,
# while the initial model was trained for 100 steps.
self.assertNotAllClose(coefficients_code_start,
self.custom_weights,
msg='models are too close')
def test_scoring_should_fail_if_not_trained(self):
"""
Inference should fail on untrained model
"""
# Reset trainer object
self.binary_lr_trainer = BinaryLogisticRegressionTrainer()
with self.assertRaises(Exception):
self.binary_lr_trainer.predict_proba(X=self.x_test, offsets=None)
def __init__(self, consumer_id, regularize_bias=False, lambda_l2=1.0, tolerance=1e-8, num_of_curvature_pairs=10,
num_iterations=100):
self.consumer_id = consumer_id
self.lr_trainer = BinaryLogisticRegressionTrainer(regularize_bias=regularize_bias, lambda_l2=lambda_l2,
precision=tolerance/np.finfo(float).eps,
num_lbfgs_corrections=num_of_curvature_pairs,
max_iter=num_iterations)
self.processed_counter = 0
def test_scoring_should_succeed_if_custom_weights_provided(self):
"""
Inference should succeed on untrained model if custom weights provided
"""
# Reset trainer object
self.binary_lr_trainer = BinaryLogisticRegressionTrainer()
validation_pred = self.binary_lr_trainer.predict_proba(
X=self.x_test, offsets=None, custom_theta=self.custom_weights)
assert (validation_pred.shape[0] == self.x_test.shape[0])
def setUp(self):
# Since grid machines may or may not have access to internet,
# using a pickled instance of popular open-source breast cancer dataset for testing
sample_dataset = pickle.load(
open(sample_dataset_path + "/sklearn_data.p", "rb"))
self.x_train, self.x_test, self.y_train, self.y_test = train_test_split(
sample_dataset.data,
sample_dataset.target,
test_size=0.25,
random_state=0)
self.binary_lr_trainer = BinaryLogisticRegressionTrainer(max_iter=1000)
self.binary_lr_trainer_without_bias = BinaryLogisticRegressionTrainer(
max_iter=1000, has_intercept=False)
self.custom_weights = self.binary_lr_trainer.fit(X=self.x_train,
y=self.y_train,
weights=None,
offsets=None)[0][0]
def test_metrics_computation_should_fail_if_model_not_trained(self):
"""
Metrics computation should fail on untrained model
"""
# Reset trainer object
self.binary_lr_trainer = BinaryLogisticRegressionTrainer()
with self.assertRaises(Exception):
self.binary_lr_trainer.compute_metrics(X=self.x_test,
y=self.y_test,
offsets=None)
def test_metrics_computation_should_succeed_if_custom_weights_provided(self):
"""
Metrics computation should succeed on untrained model if custom weights provided
"""
# Reset trainer object
self.binary_lr_trainer = BinaryLogisticRegressionTrainer()
validation_metrics = self.binary_lr_trainer.compute_metrics(X=self.x_test,
y=self.y_test,
offsets=None,
custom_theta=self.custom_weights)
assert (0.0 <= validation_metrics['auc'] <= 1.0)
def test_scoring_should_fail_if_custom_weights_not_of_known_type(self):
"""
Inference should fail if custom weights are neither Numpy ndarray or Scipy sparse amtrix
"""
# Reset trainer object
self.binary_lr_trainer = BinaryLogisticRegressionTrainer()
# Run inference using a Python list, which is neither a numpy ndarray nor a scipy matrix
with self.assertRaises(Exception):
self.binary_lr_trainer.predict_proba(X=self.x_test,
offsets=None,
custom_theta=self.custom_weights.tolist())
def _predict(self, pool, input_path, metadata, tensor_metadata, output_file, schema_params, num_features, metadata_file, model_weights, use_local_index):
logger.info(f"Start inference for {input_path}.")
# Create LR model object for inference
lr_model = BinaryLogisticRegressionTrainer(regularize_bias=True, lambda_l2=self.model_params.l2_reg_weight)
consumer = InferenceJobConsumer(lr_model, num_features, schema_params, use_local_index, name=input_path)
results = self._pooled_action(pool, consumer, input_path, schema_params, model_weights, num_features, metadata_file, gen_index_map=False)
# Set up output schema
output_schema = fastavro.parse_schema(get_inference_output_avro_schema(
metadata,
has_logits_per_coordinate=True, # Always true for custom scipy-based LR
schema_params=schema_params,
has_weight=any(schema_params.sample_weight == feature.name for feature in tensor_metadata.get_features())))
batched_write_avro(itertools.chain.from_iterable(results), output_file, output_schema)
logger.info(f"Inference complete: {input_path}.")
def _train(self, pool, input_path, metadata_file, model_weights: dict, num_features, schema_params, output_model_file):
logger.info(f"Start training with {f'loaded {len(model_weights)} previous models' if model_weights else 'zeros'} as the model initial value.")
lr_model = BinaryLogisticRegressionTrainer(regularize_bias=self.model_params.regularize_bias, lambda_l2=self.model_params.l2_reg_weight,
precision=self.model_params.lbfgs_tolerance / np.finfo(float).eps,
num_lbfgs_corrections=self.model_params.num_of_lbfgs_curvature_pairs,
max_iter=self.model_params.num_of_lbfgs_iterations)
consumer = TrainingJobConsumer(lr_model, name=input_path)
results = self._pooled_action(pool, consumer, input_path, schema_params, model_weights, num_features, metadata_file,
self.model_params.enable_local_indexing)
model_weights.update(results)
logger.info(f"{len(model_weights)} models in total after training/refreshing.")
# Dump results to model output directory.
if self.model_params.feature_file:
self._save_model(output_model_file, model_coefficients=model_weights, num_features=num_features, feature_file=self.model_params.feature_file)
else:
logger.info("Both feature file and avro model output directory required to export model. Skipping export")
return model_weights
def _train(self, pool, input_path, metadata_file, model_weights: dict,
num_features, schema_params, output_model_file):
logger.info(
f"Start training with {f'loaded {len(model_weights)} previous models' if model_weights else 'zeros'} as the model initial value."
)
lr_model = BinaryLogisticRegressionTrainer(
regularize_bias=self.model_params.regularize_bias,
lambda_l2=self.model_params.l2_reg_weight,
precision=self.model_params.lbfgs_tolerance / np.finfo(float).eps,
num_lbfgs_corrections=self.model_params.
num_of_lbfgs_curvature_pairs,
max_iter=self.model_params.num_of_lbfgs_iterations,
has_intercept=self.has_intercept)
consumer = TrainingJobConsumer(
lr_model,
name=input_path,
job_queue=self.job_queue,
enable_local_indexing=self.model_params.enable_local_indexing,
sparsity_threshold=self.model_params.sparsity_threshold,
variance_mode=self.model_params.random_effect_variance_mode)
# Make sure the queue is empty
assert (self.job_queue.empty())
results = self._pooled_action(pool, consumer, input_path,
schema_params, model_weights,
num_features, metadata_file,
self.model_params.enable_local_indexing)
# The trained model should be updated by the prior model to cover the following two cases:
# (1) the prior model has extra features that are not present in the current datasets.
# (2) the prior model has extra model_ids that are not present in the current datasets.
# In both cases, the extra features/model_id needs to be copied to the current models.
# This is needed especially incremental learning is implemented.
# It is not needed when the prior model and current model share the same features / model_ids.
# Revisit this when we start working on more advanced warm start.
model_weights.update(results)
logger.info(
f"{len(model_weights)} models in total after training/refreshing.")
# Dump results to model output directory.
self._save_model(output_model_file,
model_coefficients=model_weights,
num_features=num_features,
feature_file=self.feature_file)
return model_weights
def _predict(self, inference_dataset, model_coefficients, metadata,
tensor_metadata, output_file, prediction_params):
# Create LR trainer object for inference
lr_trainer = BinaryLogisticRegressionTrainer(
regularize_bias=True,
lambda_l2=self.model_params[constants.L2_REG_WEIGHT])
# Create PhotonMLWriter object
prediction_params.update(self.model_params)
inference_runner = PhotonMLWriter(schema_params=prediction_params)
# Delegate inference to PhotonMLWriter object
inference_runner.run_custom_scipy_re_inference(
inference_dataset=inference_dataset,
model_coefficients=model_coefficients,
lr_model=lr_trainer,
metadata=metadata,
tensor_metadata=tensor_metadata,
output_file=output_file)
def _predict(self, pool, input_path, metadata, tensor_metadata,
output_file, schema_params, num_features, metadata_file,
model_weights):
logger.info(f"Start inference for {input_path}.")
# Create LR model object for inference
lr_model = BinaryLogisticRegressionTrainer(
regularize_bias=True,
lambda_l2=self.model_params.l2_reg_weight,
has_intercept=self.has_intercept)
consumer = InferenceJobConsumer(lr_model,
num_features,
schema_params,
name=input_path,
job_queue=self.job_queue)
# Make sure the queue is empty
assert (self.job_queue.empty())
# Prediction does not use local indexing since it can work on sparse coefficients directly.
results = self._pooled_action(pool,
consumer,
input_path,
schema_params,
model_weights,
num_features,
metadata_file,
enable_local_indexing=False)
# Set up output schema
output_schema = fastavro.parse_schema(
get_inference_output_avro_schema(
metadata,
has_logits_per_coordinate=
True, # Always true for custom scipy-based LR
schema_params=schema_params,
has_weight=any(schema_params.weight_column_name == feature.name
for feature in tensor_metadata.get_features())))
batched_write_avro(itertools.chain.from_iterable(results), output_file,
output_schema)
logger.info(f"Inference complete: {input_path}.")
def test_fit_with_variance_computation_without_intercept(self):
"""
Test fit when the variance computation is required but no intercept is used
"""
# Generate the dataset
num_features = 10
num_samples = 100
X = np.random.randn(num_samples, num_features)
y = np.random.randint(2, size=num_samples)
weights = np.random.rand(num_samples)
offsets = np.random.randn(num_samples)
lambda_l2 = 0.0
binary_lr_trainer = BinaryLogisticRegressionTrainer(
lambda_l2=lambda_l2,
max_iter=1000,
regularize_bias=True,
has_intercept=False)
expected_simple = compute_coefficients_and_variance(
X=X,
y=y,
weights=weights,
offsets=offsets,
variance_mode=constants.SIMPLE,
lambda_l2=lambda_l2,
has_intercept=False)
expected_full = compute_coefficients_and_variance(
X=X,
y=y,
weights=weights,
offsets=offsets,
variance_mode=constants.FULL,
lambda_l2=lambda_l2,
has_intercept=False)
actual_simple = binary_lr_trainer.fit(X=sparse.csr_matrix(X),
y=y,
weights=weights,
offsets=offsets,
variance_mode=constants.SIMPLE)
actual_full = binary_lr_trainer.fit(X=sparse.csr_matrix(X),
y=y,
weights=weights,
offsets=offsets,
variance_mode=constants.FULL)
self.assertAllClose(expected_simple[0],
actual_simple[0][0],
rtol=1e-02,
atol=1e-02,
msg='simple mean mismatch')
self.assertAllClose(expected_simple[1],
actual_simple[1],
rtol=1e-02,
atol=1e-02,
msg='simple variance mismatch')
self.assertAllClose(expected_full[0],
actual_full[0][0],
rtol=1e-02,
atol=1e-02,
msg='full mean mismatch')
self.assertAllClose(expected_full[1],
actual_full[1],
rtol=1e-02,
atol=1e-02,
msg='full variance mismatch')