def _test_optimizer(optimizer, target=0.75):
x_train, y_train = get_test_data()
model = get_model(x_train.shape[1], 10, y_train.shape[1])
model.compile(loss='categorical_crossentropy',
optimizer=optimizer,
metrics=['accuracy'])
history = model.fit(x_train, y_train, epochs=2, batch_size=16, verbose=0)
assert history.history['acc'][-1] >= target
config = optimizers.serialize(optimizer)
custom_objects = {optimizer.__class__.__name__: optimizer.__class__}
optim = optimizers.deserialize(config, custom_objects)
new_config = optimizers.serialize(optim)
assert config == new_config
def test_functional_model(spark_context, classification_model_functional,
mnist_data):
batch_size = 64
epochs = 1
x_train, y_train, x_test, y_test = mnist_data
x_train = x_train[:1000]
y_train = y_train[:1000]
df = to_data_frame(spark_context, x_train, y_train, categorical=True)
test_df = to_data_frame(spark_context, x_test, y_test, categorical=True)
sgd = optimizers.SGD()
sgd_conf = optimizers.serialize(sgd)
estimator = ElephasEstimator()
estimator.set_keras_model_config(classification_model_functional.to_yaml())
estimator.set_optimizer_config(sgd_conf)
estimator.set_mode("synchronous")
estimator.set_loss("categorical_crossentropy")
estimator.set_metrics(['acc'])
estimator.set_epochs(epochs)
estimator.set_batch_size(batch_size)
estimator.set_validation_split(0.1)
estimator.set_categorical_labels(True)
estimator.set_nb_classes(10)
pipeline = Pipeline(stages=[estimator])
fitted_pipeline = pipeline.fit(df)
prediction = fitted_pipeline.transform(test_df)
pnl = prediction.select("label", "prediction")
pnl.show(100)
prediction_and_label = pnl.rdd.map(lambda row: (row.label, row.prediction))
metrics = MulticlassMetrics(prediction_and_label)
print(metrics.accuracy)
def test_custom_objects(spark_context, boston_housing_dataset):
def custom_activation(x):
return 2 * relu(x)
model = Sequential()
model.add(Dense(64, input_shape=(13, )))
model.add(Dense(64, activation=custom_activation))
model.add(Dense(1, activation='linear'))
x_train, y_train, x_test, y_test = boston_housing_dataset
df = to_data_frame(spark_context, x_train, y_train)
test_df = to_data_frame(spark_context, x_test, y_test)
sgd = optimizers.SGD(lr=0.00001)
sgd_conf = optimizers.serialize(sgd)
estimator = ElephasEstimator()
estimator.set_keras_model_config(model.to_yaml())
estimator.set_optimizer_config(sgd_conf)
estimator.set_mode("synchronous")
estimator.set_loss("mae")
estimator.set_metrics(['mae'])
estimator.set_epochs(10)
estimator.set_batch_size(32)
estimator.set_validation_split(0.01)
estimator.set_categorical_labels(False)
estimator.set_custom_objects({'custom_activation': custom_activation})
pipeline = Pipeline(stages=[estimator])
fitted_pipeline = pipeline.fit(df)
prediction = fitted_pipeline.transform(test_df)
def test_regression_model(spark_context, regression_model,
boston_housing_dataset):
batch_size = 64
epochs = 10
x_train, y_train, x_test, y_test = boston_housing_dataset
df = to_data_frame(spark_context, x_train, y_train)
test_df = to_data_frame(spark_context, x_test, y_test)
sgd = optimizers.SGD(lr=0.00001)
sgd_conf = optimizers.serialize(sgd)
estimator = ElephasEstimator()
estimator.set_keras_model_config(regression_model.to_yaml())
estimator.set_optimizer_config(sgd_conf)
estimator.set_mode("synchronous")
estimator.set_loss("mae")
estimator.set_metrics(['mae'])
estimator.set_epochs(epochs)
estimator.set_batch_size(batch_size)
estimator.set_validation_split(0.01)
estimator.set_categorical_labels(False)
pipeline = Pipeline(stages=[estimator])
fitted_pipeline = pipeline.fit(df)
prediction = fitted_pipeline.transform(test_df)
pnl = prediction.select("label", "prediction")
pnl.show(100)
prediction_and_observations = pnl.rdd.map(lambda row:
(row.label, row.prediction))
metrics = RegressionMetrics(prediction_and_observations)
print(metrics.r2)
def get_config(self):
config = {
'optimizer': optimizers.serialize(self.optimizer),
'multipliers': self.multipliers
}
base_config = super(LRMultiplier, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def test_spark_ml_model_classification(spark_context, classification_model,
mnist_data):
batch_size = 64
nb_classes = 10
epochs = 1
x_train, y_train, x_test, y_test = mnist_data
x_train = x_train[:1000]
y_train = y_train[:1000]
df = to_data_frame(spark_context, x_train, y_train, categorical=True)
test_df = to_data_frame(spark_context, x_test, y_test, categorical=True)
sgd = optimizers.SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
sgd_conf = optimizers.serialize(sgd)
# Initialize Spark ML Estimator
estimator = ElephasEstimator()
estimator.set_keras_model_config(classification_model.to_yaml())
estimator.set_optimizer_config(sgd_conf)
estimator.set_mode("synchronous")
estimator.set_loss("categorical_crossentropy")
estimator.set_metrics(['acc'])
estimator.set_epochs(epochs)
estimator.set_batch_size(batch_size)
estimator.set_validation_split(0.1)
estimator.set_categorical_labels(True)
estimator.set_nb_classes(nb_classes)
# Fitting a model returns a Transformer
pipeline = Pipeline(stages=[estimator])
fitted_pipeline = pipeline.fit(df)
# Evaluate Spark model by evaluating the underlying model
prediction = fitted_pipeline.transform(test_df)
pnl = prediction.select("label", "prediction")
pnl.show(100)
# since prediction in a multiclass classification problem is a vector, we need to compute argmax
# the casting to a double is just necessary for using MulticlassMetrics
pnl = pnl.select(
'label',
argmax('prediction').astype(DoubleType()).alias('prediction'))
prediction_and_label = pnl.rdd.map(lambda row: (row.label, row.prediction))
metrics = MulticlassMetrics(prediction_and_label)
print(metrics.accuracy)
def test_batch_predict_classes_probability(spark_context, classification_model,
mnist_data):
batch_size = 64
nb_classes = 10
epochs = 1
x_train, y_train, x_test, y_test = mnist_data
x_train = x_train[:1000]
y_train = y_train[:1000]
df = to_data_frame(spark_context, x_train, y_train, categorical=True)
test_df = to_data_frame(spark_context, x_test, y_test, categorical=True)
sgd = optimizers.SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
sgd_conf = optimizers.serialize(sgd)
# Initialize Spark ML Estimator
estimator = ElephasEstimator()
estimator.set_keras_model_config(classification_model.to_yaml())
estimator.set_optimizer_config(sgd_conf)
estimator.set_mode("synchronous")
estimator.set_loss("categorical_crossentropy")
estimator.set_metrics(['acc'])
estimator.set_epochs(epochs)
estimator.set_batch_size(batch_size)
estimator.set_validation_split(0.1)
estimator.set_categorical_labels(True)
estimator.set_nb_classes(nb_classes)
# Fitting a model returns a Transformer
fitted_pipeline = estimator.fit(df)
results = fitted_pipeline.transform(test_df)
# Set inference batch size and do transform again on the same test_df
inference_batch_size = int(len(y_test) / 10)
fitted_pipeline.set_params(inference_batch_size=inference_batch_size)
fitted_pipeline.set_params(outputCol="prediction_via_batch_inference")
results_with_batch_prediction = fitted_pipeline.transform(results)
# we should have an array of 10 elements in the prediction column, since we have 10 classes
# and therefore 10 probabilities
results_np = results_with_batch_prediction.take(1)[0]
assert len(results_np.prediction) == 10
assert len(results_np.prediction_via_batch_inference) == 10
assert np.array_equal(results_np.prediction,
results_np.prediction_via_batch_inference)
def test_save_pipeline(spark_context, classification_model):
sgd = optimizers.SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
sgd_conf = optimizers.serialize(sgd)
# Initialize Spark ML Estimator
estimator = ElephasEstimator()
estimator.set_keras_model_config(classification_model.to_yaml())
estimator.set_optimizer_config(sgd_conf)
estimator.set_mode("synchronous")
estimator.set_loss("categorical_crossentropy")
estimator.set_metrics(['acc'])
estimator.set_epochs(10)
estimator.set_batch_size(10)
estimator.set_validation_split(0.1)
estimator.set_categorical_labels(True)
estimator.set_nb_classes(10)
# Fitting a model returns a Transformer
pipeline = Pipeline(stages=[estimator])
pipeline.save('tmp')
def test_set_cols_deprecated(spark_context, regression_model,
boston_housing_dataset):
with pytest.deprecated_call():
batch_size = 64
epochs = 10
x_train, y_train, x_test, y_test = boston_housing_dataset
df = to_data_frame(spark_context, x_train, y_train)
df = df.withColumnRenamed('features', 'scaled_features')
df = df.withColumnRenamed('label', 'ground_truth')
test_df = to_data_frame(spark_context, x_test, y_test)
test_df = test_df.withColumnRenamed('features', 'scaled_features')
test_df = test_df.withColumnRenamed('label', 'ground_truth')
sgd = optimizers.SGD(lr=0.00001)
sgd_conf = optimizers.serialize(sgd)
estimator = ElephasEstimator()
estimator.set_keras_model_config(regression_model.to_yaml())
estimator.set_optimizer_config(sgd_conf)
estimator.setFeaturesCol('scaled_features')
estimator.setOutputCol('output')
estimator.setLabelCol('ground_truth')
estimator.set_mode("synchronous")
estimator.set_loss("mae")
estimator.set_metrics(['mae'])
estimator.set_epochs(epochs)
estimator.set_batch_size(batch_size)
estimator.set_validation_split(0.01)
estimator.set_categorical_labels(False)
pipeline = Pipeline(stages=[estimator])
fitted_pipeline = pipeline.fit(df)
prediction = fitted_pipeline.transform(test_df)
pnl = prediction.select("ground_truth", "output")
pnl.show(100)
prediction_and_observations = pnl.rdd.map(
lambda row: (row['ground_truth'], row['output']))
metrics = RegressionMetrics(prediction_and_observations)
print(metrics.r2)
def get_config(self):
config = {
'units': self.units,
'learning_rate': self.learning_rate,
'online': self.online,
'n_passes': self.n_passes,
'return_hidden': self.return_hidden,
'visible_activation': activations.serialize(self.visible_activation),
'hidden_activation': activations.serialize(self.hidden_activation),
'use_bias': self.use_bias,
'kernel_initializer': initializers.serialize(self.kernel_initializer),
'bias_initializer': initializers.serialize(self.bias_initializer),
'kernel_regularizer': regularizers.serialize(self.kernel_regularizer),
'bias_regularizer': regularizers.serialize(self.bias_regularizer),
'activity_regularizer': regularizers.serialize(self.activity_regularizer),
'kernel_constraint': constraints.serialize(self.kernel_constraint),
'bias_constraint': constraints.serialize(self.bias_constraint),
'optimizer': optimizers.serialize(self.optimizer)
}
base_config = super(OnlineBolzmannCell, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def test_predict_classes_probability(spark_context, classification_model,
mnist_data):
batch_size = 64
nb_classes = 10
epochs = 1
x_train, y_train, x_test, y_test = mnist_data
x_train = x_train[:1000]
y_train = y_train[:1000]
df = to_data_frame(spark_context, x_train, y_train, categorical=True)
test_df = to_data_frame(spark_context, x_test, y_test, categorical=True)
sgd = optimizers.SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
sgd_conf = optimizers.serialize(sgd)
# Initialize Spark ML Estimator
estimator = ElephasEstimator()
estimator.set_keras_model_config(classification_model.to_yaml())
estimator.set_optimizer_config(sgd_conf)
estimator.set_mode("synchronous")
estimator.set_loss("categorical_crossentropy")
estimator.set_metrics(['acc'])
estimator.set_predict_classes(False)
estimator.set_epochs(epochs)
estimator.set_batch_size(batch_size)
estimator.set_validation_split(0.1)
estimator.set_categorical_labels(True)
estimator.set_nb_classes(nb_classes)
# Fitting a model returns a Transformer
pipeline = Pipeline(stages=[estimator])
fitted_pipeline = pipeline.fit(df)
results = fitted_pipeline.transform(test_df)
# we should have an array of 10 elements in the prediction column, since we have 10 classes
# and therefore 10 probabilities
assert len(results.take(1)[0].prediction) == 10
model.add(Dense(512, input_shape=(input_dim, )))
model.add(Activation('relu'))
model.add(Dropout(0.5))
model.add(Dense(512))
model.add(Activation('relu'))
model.add(Dropout(0.5))
model.add(Dense(512))
model.add(Activation('relu'))
model.add(Dropout(0.5))
model.add(Dense(nb_classes))
model.add(Activation('softmax'))
model.compile(loss='categorical_crossentropy', optimizer='adam')
sgd = optimizers.SGD(lr=0.01)
sgd_conf = optimizers.serialize(sgd)
# Initialize Elephas Spark ML Estimator
estimator = ElephasEstimator()
estimator.set_keras_model_config(model.to_yaml())
estimator.set_optimizer_config(sgd_conf)
estimator.set_mode("synchronous")
estimator.set_loss("categorical_crossentropy")
estimator.set_metrics(['acc'])
estimator.setFeaturesCol("scaled_features")
estimator.setLabelCol("index_category")
estimator.set_epochs(10)
estimator.set_batch_size(128)
estimator.set_num_workers(1)
estimator.set_verbosity(0)
estimator.set_validation_split(0.15)
